Daily Papers

Progress in large language models is constrained by an evaluation bottleneck: build a benchmark, evaluate models and settings, then iterate. We therefore ask a simple question: can we forecast outcomes before running any experiments? We study text-only performance forecasting: estimating a model's score from a redacted task description and intended configuration, with no access to dataset instances. To support systematic study, we curate PRECOG, a corpus of redacted description-performance pairs spanning diverse tasks, domains, and metrics. Experiments show the task is challenging but feasible: models equipped with a retrieval module that excludes source papers achieve moderate prediction performance with well-calibrated uncertainty, reaching mean absolute error as low as 8.7 on the Accuracy subset at high-confidence thresholds. Our analysis indicates that stronger reasoning models engage in diverse, iterative querying, whereas current open-source models lag and often skip retrieval or gather evidence with limited diversity. We further test a zero-leakage setting, forecasting on newly released datasets or experiments before their papers are indexed, where GPT-5 with built-in web search still attains nontrivial prediction accuracy. Overall, our corpus and analyses offer an initial step toward open-ended anticipatory evaluation, supporting difficulty estimation and smarter experiment prioritization.

We conduct a systematic audit of three widely used reasoning benchmarks, SocialIQa, FauxPas-EAI, and ToMi, and uncover pervasive flaws in both benchmark items and evaluation methodology. Using five LLMs (GPT-{3, 3.5, 4, o1}, and LLaMA 3.1) as diagnostic tools, we identify structural, semantic, and pragmatic issues in benchmark design (e.g., duplicated items, ambiguous wording, and implausible answers), as well as scoring procedures that prioritize output form over reasoning process. Through systematic human annotation and re-evaluation on cleaned benchmark subsets, we find that model scores often improve not due to due to erratic surface wording variations and not to improved reasoning. Infact, further analyses show that model performance is highly sensitive to minor input variations such as context availability and phrasing, revealing that high scores may reflect alignment with format-specific cues rather than consistent inference based on the input. These findings challenge the validity of current benchmark-based claims about reasoning in LLMs, and highlight the need for evaluation protocols that assess reasoning as a process of drawing inference from available information, rather than as static output selection. We release audited data and evaluation tools to support more interpretable and diagnostic assessments of model reasoning.

The training data for many Large Language Models (LLMs) is contaminated with test data. This means that public benchmarks used to assess LLMs are compromised, suggesting a performance gap between benchmark scores and actual capabilities. Ideally, a private holdout set could be used to accurately verify scores. Unfortunately, such datasets do not exist for most benchmarks, and post-hoc construction of sufficiently similar datasets is non-trivial. To address these issues, we introduce a systematic methodology for (i) retrospectively constructing a holdout dataset for a target dataset, (ii) demonstrating the statistical indistinguishability of this retro-holdout dataset, and (iii) comparing LLMs on the two datasets to quantify the performance gap due to the dataset's public availability. Applying these methods to TruthfulQA, we construct and release Retro-Misconceptions, on which we evaluate twenty LLMs and find that some have inflated scores by as much as 16 percentage points. Our results demonstrate that public benchmark scores do not always accurately assess model properties, and underscore the importance of improved data practices in the field.

This paper introduces the UCFE: User-Centric Financial Expertise benchmark, an innovative framework designed to evaluate the ability of large language models (LLMs) to handle complex real-world financial tasks. UCFE benchmark adopts a hybrid approach that combines human expert evaluations with dynamic, task-specific interactions to simulate the complexities of evolving financial scenarios. Firstly, we conducted a user study involving 804 participants, collecting their feedback on financial tasks. Secondly, based on this feedback, we created our dataset that encompasses a wide range of user intents and interactions. This dataset serves as the foundation for benchmarking 12 LLM services using the LLM-as-Judge methodology. Our results show a significant alignment between benchmark scores and human preferences, with a Pearson correlation coefficient of 0.78, confirming the effectiveness of the UCFE dataset and our evaluation approach. UCFE benchmark not only reveals the potential of LLMs in the financial sector but also provides a robust framework for assessing their performance and user satisfaction.The benchmark dataset and evaluation code are available.

In this paper, we show that knowledge distillation can be subverted to manipulate language model benchmark scores, revealing a critical vulnerability in current evaluation practices. We introduce "Data Laundering," a three-phase process analogous to financial money laundering, that enables the covert transfer of benchmark-specific knowledge through seemingly legitimate intermediate training steps. Through extensive experiments with a 2-layer BERT student model, we show how this approach can achieve substantial improvements in benchmark accuracy (up to 75\% on GPQA) without developing genuine reasoning capabilities. Notably, this method can be exploited intentionally or even unintentionally, as researchers may inadvertently adopt this method that inflates scores using knowledge distillation without realizing the implications. While our findings demonstrate the effectiveness of this technique, we present them as a cautionary tale highlighting the urgent need for more robust evaluation methods in AI. This work aims to contribute to the ongoing discussion about evaluation integrity in AI development and the need for benchmarks that more accurately reflect true model capabilities. The code is available at https://github.com/mbzuai-nlp/data_laundering.

Large Language Models (LLMs) are essential tools to collaborate with users on different tasks. Evaluating their performance to serve users' needs in real-world scenarios is important. While many benchmarks have been created, they mainly focus on specific predefined model abilities. Few have covered the intended utilization of LLMs by real users. To address this oversight, we propose benchmarking LLMs from a user perspective in both dataset construction and evaluation designs. We first collect 1846 real-world use cases with 15 LLMs from a user study with 712 participants from 23 countries. These self-reported cases form the User Reported Scenarios(URS) dataset with a categorization of 7 user intents. Secondly, on this authentic multi-cultural dataset, we benchmark 10 LLM services on their efficacy in satisfying user needs. Thirdly, we show that our benchmark scores align well with user-reported experience in LLM interactions across diverse intents, both of which emphasize the overlook of subjective scenarios. In conclusion, our study proposes to benchmark LLMs from a user-centric perspective, aiming to facilitate evaluations that better reflect real user needs. The benchmark dataset and code are available at https://github.com/Alice1998/URS.

Large Language Models (LLMs) vary in their abilities on a range of tasks. Initiatives such as the Open LLM Leaderboard aim to quantify these differences with several large benchmarks (sets of test items to which an LLM can respond either correctly or incorrectly). However, high correlations within and between benchmark scores suggest that (1) there exists a small set of common underlying abilities that these benchmarks measure, and (2) items tap into redundant information and the benchmarks may thus be considerably compressed. We use data from n > 5000 LLMs to identify the most informative items of six benchmarks, ARC, GSM8K, HellaSwag, MMLU, TruthfulQA and WinoGrande (with d=28,632 items in total). From them we distill a sparse benchmark, metabench, that has less than 3% of the original size of all six benchmarks combined. This new sparse benchmark goes beyond point scores by yielding estimators of the underlying benchmark-specific abilities. We show that these estimators (1) can be used to reconstruct each original individual benchmark score with, on average, 1.5% root mean square error (RMSE), (2) reconstruct the original total score with 0.8% RMSE, and (3) have a single underlying common factor whose Spearman correlation with the total score is r = 0.93.

With the recent development of large language models (LLMs), models that focus on certain domains and languages have been discussed for their necessity. There is also a growing need for benchmarks to evaluate the performance of current LLMs in each domain. Therefore, in this study, we constructed a benchmark comprising multiple tasks specific to the Japanese and financial domains and performed benchmark measurements on some models. Consequently, we confirmed that GPT-4 is currently outstanding, and that the constructed benchmarks function effectively. According to our analysis, our benchmark can differentiate benchmark scores among models in all performance ranges by combining tasks with different difficulties.

Large language models perform well on many logical reasoning benchmarks, but it remains unclear which core logical skills they truly master. To address this, we introduce LogicSkills, a benchmark that isolates three fundamental logical skills: (i) formal symbolizationx2014{}translating premises into first-order logic; (ii) countermodel constructionx2014showing that an argument is logically invalid by constructing a finite countermodel; and (iii) validity assessmentx2014determining whether a conclusion follows from a set of premises. Items are drawn from the two-variable fragment of first-order logic without identity and are presented in both English and a Carrollian nonce-word language. All instances are solver-verified with Z3 for correctness and non-triviality. Across conventional instruction-tuned LLMs, performance is high on validity assessment but substantially lower on formal symbolization and countermodel construction, highlighting that high task-level accuracy can mask weaknesses in core logical skills. In contrast, recent reasoning-tuned models perform strongly across all three tasks, suggesting a more systematic logical skill profile.

Large Language Models (LLMs) effectiveness is usually evaluated by means of benchmarks such as MMLU, ARC-C, or HellaSwag, where questions are presented in their original wording, thus in a fixed, standardized format. However, real-world applications involve linguistic variability, requiring models to maintain their effectiveness across diverse rewordings of the same question or query. In this study, we systematically assess the robustness of LLMs to paraphrased benchmark questions and investigate whether benchmark-based evaluations provide a reliable measure of model capabilities. We systematically generate various paraphrases of all the questions across six different common benchmarks, and measure the resulting variations in effectiveness of 34 state-of-the-art LLMs, of different size and effectiveness. Our findings reveal that while LLM rankings remain relatively stable across paraphrased inputs, absolute effectiveness scores change, and decline significantly. This suggests that LLMs struggle with linguistic variability, raising concerns about their generalization abilities and evaluation methodologies. Furthermore, the observed performance drop challenges the reliability of benchmark-based evaluations, indicating that high benchmark scores may not fully capture a model's robustness to real-world input variations. We discuss the implications of these findings for LLM evaluation methodologies, emphasizing the need for robustness-aware benchmarks that better reflect practical deployment scenarios.

Terminal-agent benchmarks have become a primary signal for measuring the coding and system-administration capabilities of large language models. As the market for evaluation environments grows, so does the pressure to ship tasks quickly, often without thorough adversarial review of the verification logic. This paper is a guideline for writing good benchmark tasks, drawn from over a year of contributing to and reviewing tasks for Terminal Bench. Most people write benchmark tasks the way they write prompts. They shouldn't. A prompt is designed to help the agent succeed; a benchmark is designed to find out if it can. We argue that good tasks are adversarial, difficult, and legible, and that a large class of common failure modes -- AI-generated instructions, over-prescriptive specifications, clerical difficulty, oracle solutions that assume hidden knowledge, tests that validate the wrong things, and reward-hackable environments -- are predictable consequences of treating task authoring as prompt authoring. We catalog these failure modes, argue that real difficulty is conceptual rather than environmental, and discuss recent empirical evidence that over 15% of tasks in popular terminal-agent benchmarks are reward-hackable. We hope this serves as a useful reference for benchmark maintainers, task contributors, and researchers using benchmark scores as evidence.

Vision-language models (VLMs) are increasingly deployed where answers must follow from what is in the image, yet they often answer from textual priors, the question's phrasing together with memorized world knowledge, rather than from the image itself, which inflates benchmark scores and yields confident but ungrounded answers. Existing benchmarks rarely isolate this behavior, since each image is usually paired with a single fixed question. To measure the reliance, we build a 540-image benchmark across six reasoning categories and generate four question variants over the same images, so that phrasing rather than image content is the controlled variable. The hardest variant is written directly from the image to minimize text leakage. We benchmark eleven VLMs spanning small open-weight models to large closed-source systems: every model degrades on the hardest variant, and open models fall furthest. Our central diagnostic is a no-image ablation, which collapses the open-weight models to their text-only floor (1 to 9 percent). Three further analyses, LLM-rated difficulty, low base-to-final textual similarity, and human re-annotation, corroborate genuine image-dependence. In-context exemplars that match how a variant was built recover the most accuracy, and GRPO post-training of a small VLM yields consistent gains across all four variants that transfer to a held-out out-of-distribution set. Textual-prior reliance is measurable and partly trainable away.

Text embedding models are widely used for semantic similarity tasks, including information retrieval, clustering, and classification. General-purpose models are typically trained with single- or multi-stage processes using contrastive loss functions. We introduce a novel training regimen that combines model distillation techniques with task-specific contrastive loss to produce compact, high-performance embedding models. Our findings suggest that this approach is more effective for training small models than purely contrastive or distillation-based training paradigms alone. Benchmark scores for the resulting models, jina-embeddings-v5-text-small and jina-embeddings-v5-text-nano, exceed or match the state-of-the-art for models of similar size. jina-embeddings-v5-text models additionally support long texts (up to 32k tokens) in many languages, and generate embeddings that remain robust under truncation and binary quantization. Model weights are publicly available, hopefully inspiring further advances in embedding model development.

Benchmark scores tell you what an agent got right; they do not tell you how it got there. In this work, we introduce methods for comparing agents procedurally in different contexts, where the model, tasks, and approaches vary. We compare ten agents and find that they are identifiable by their behavioral habits, which we define as fingerprints: a probe over these procedural signatures attributes an unseen trajectory to the correct agent at 85.7% accuracy, controlling for leakage across tasks. We develop procedural representations for agent problem-solving procedures with an emergent vocabulary induction technique that is meant to be maximally compressive to avoid surface-level variation while being expressive enough to unveil the quirks of the models' patterns. We apply our framework to the software engineering evaluation dataset SWE-Bench to study the structural distinctness of agent trajectories and find that behavior is most similar between models from similar release periods and those that are distilled from one another (e.g., a distilled student model and its teacher have a Jensen-Shannon divergence of 0.25, about half the distance between other model pairs). As more models saturate evaluations, we believe that it will be important to probe model behavior along more holistic dimensions than success rates alone. We introduce ProcGrep, a library for auditing and evaluating agents for how they approach tasks at a procedural level given their traces in a top-down fashion. We believe this work has a range of applications to help developers work with and program coding agents, such as task-aware model routing, agent monitoring, and finer-grained cost analysis.

Specifying a penalty can turn a legal obligation into a cost-benefit calculation that favors violation. We show that this enforcement information paradox occurs in AI agents. Most AI safety evaluations test whether models fail; we ask why, using compliance theory from law and economics as a diagnostic. We evaluate twelve instruction-tuned language models deployed as enterprise procurement chatbots. Each is given an environmental regulation in its system prompt covering large purchases, and a vendor list on which the certified suppliers cost nearly twice what the uncertified ones do. We test the agents against the predictions of deterrence, legitimacy, and expressive law, and find that each theory accounts for part of what we observe. Under identical conditions, compliance spans 46 percentage points across models, and models differ in which pressure breaks them: some treat the regulation as binding however it is worded, while others fail where theory predicts, under low penalties and non-command phrasing. Benchmark scores and developers' own descriptions of post-training do not predict where a model falls. Across all twelve, financial incentives, managerial demands, peer outcomes, and employee pressure each produce large compliance failures. These agents violate regulatory constraints to satisfy local user objectives in ways standard alignment benchmarks do not measure. Embedding the rule in the system prompt is not on its own enough to produce a compliant agent: model selection is itself a governance decision, and benchmark evaluation is not sufficient for compliance-sensitive deployments.

This paper presents the sound event localization and detection (SELD) task setup for the DCASE 2019 challenge. The goal of the SELD task is to detect the temporal activities of a known set of sound event classes, and further localize them in space when active. As part of the challenge, a synthesized dataset with each sound event associated with a spatial coordinate represented using azimuth and elevation angles is provided. These sound events are spatialized using real-life impulse responses collected at multiple spatial coordinates in five different rooms with varying dimensions and material properties. A baseline SELD method employing a convolutional recurrent neural network is used to generate benchmark scores for this reverberant dataset. The benchmark scores are obtained using the recommended cross-validation setup.

Common-sense reasoning is a key language model capability because it encapsulates not just specific factual knowledge but rather general language and world understanding. Measuring common-sense reasoning, therefore, is crucial for language models of different sizes and applications. One of the most widely used benchmarks for evaluating such capabilities is HellaSwag; however, in this paper, we show that it has severe construct validity issues. These issues range from basic ungrammaticality and numerous typos to misleading prompts or equally correct options. Furthermore, we show that if models are evaluated only on answer texts, or with "Lorem ipsum dolor..." instead of the question, more than 65% of model predictions remain the same, and this cannot be attributed merely to contamination. Since benchmark scores are an essential part of model selection in both research and commercial applications, these validity issues can have severe consequences. In particular, knowing that taking benchmark scores at face value is ubiquitous, inadequate evaluation leads to ill-informed decisions about models. In this paper, we thoroughly investigate critical validity issues posed by HellaSwag and illustrate them with various evaluations using generative language models of different sizes. We argue that this benchmark does not accurately measure common-sense reasoning and, therefore, should not be used for evaluation in its current state. Based on the results of our study, we propose requirements that should be met by future common-sense reasoning benchmarks. In addition, we release GoldenSwag, a corrected subset of HellaSwag, which, to our belief, facilitates acceptable common-sense reasoning evaluation.

Safety benchmark scores provide incomplete evidence of deployment readiness: aligned language models often adhere to rigid rules even when a situational update flips which action is safe. We term this failure brittle safety. To diagnose it, we introduce context-flip evaluation, testing 12 models across a safety benchmark (PacifAIst) and two commonsense controls using paired variants where the nominally safe action produces harm. Three findings emerge. First, brittle safety is safety-specific: all 12 models exhibit a safety-commonsense gap (mean +17.4 pp). Baseline accuracy fails to predict brittleness: among models above 90% baseline accuracy, brittleness rates range from 13.7% to 90.0%. Second, failures stem from policy override rather than miscomprehension: despite acknowledging the context change in every case, models persist via three distinct mechanisms that vary by update type and model family. Third, on a hand-audited probe of catastrophic consequence-flip scenarios, standard action-level guardrails catch none, while a state-aware validator catches all without false alarms on correct interventions. This indicates that action-level content moderation is systematically blind to consequence-flips, motivating state-aware architectural alternatives. We release our protocol, perturbed benchmarks, and deployment probe.

Benchmarks are widely used to evaluate task completion by Large Language Models (LLMs), but this approach has accumulated construction-validity problems, and a passing score may not show whether the requested task was delivered. We study both problems. In a controlled code-as-spec setup, two production Copilot CLI agents (claude-opus-4.7, gpt-5.5) re-implement a React Fluent-UI data table in Angular as a reusable library under a hidden 222-test Playwright oracle across 18 runs and three oracle-availability conditions. Alongside the score, we run a mechanical library audit and check each verdict with a no-op ablation. Without the oracle, the library is present but unfinished, revealed by scores. With the oracle in the loop, the score reaches near-perfect, but from a demo holding the tested behavior directly, the library left dead or absent. We call this building to the test; the broader disposition behind both we call validation self-awareness. The agent does not, on its own, validate what it ships as a user would. Prevalence remains an open question across other agents, signals, and model families. Beyond benchmark scores, dispositions like validation self-awareness merit research attention.

Sales dialogues require multi-turn, goal-directed persuasion under asymmetric incentives, which makes them a challenging setting for large language models (LLMs). Yet existing dialogue benchmarks rarely measure deal progression and outcomes. We introduce SalesLLM benchmark, a bilingual (ZH/EN) benchmark derived from realistic applications covering Financial Services and Consumer Goods, built from 30,074 scripted configurations and 1,805 curated multi-turn scenarios with controllable difficulty and personas. We propose a fully automatic evaluation pipeline that combines (i) an LLM judge for sales-process progress, and (ii) fine-tuned BERT classifiers for end-of-dialogue buying intent. To improve simulation fidelity, we train a user model, CustomerLM, with SFT and DPO on 8,000+ crowdworker-involved sales conversations, reducing role inversion from 17.44% (GPT-4o) to 8.8%. SalesLLM benchmark scores correlate strongly with human ratings (average Pearson r=0.86; inter-annotator agreement Krippendorff's alpha=0.86, 500 annotations per model). Across 15 mainstream LLMs, in Chinese the strongest models are competitive with typical (junior-to-intermediate) human salespeople -- not with sales experts -- while weaker ones fall below this baseline, and cross-lingual consistency remains poor. SalesLLM benchmark serves as a scalable benchmark for outcome-oriented sales agents. Our code and data are released at https://github.com/Bairong-Xdynamics/Benchmarking-LLM-Realistic-Selling-Skill

Most works on modeling the conversation history in Conversational Question Answering (CQA) report a single main result on a common CQA benchmark. While existing models show impressive results on CQA leaderboards, it remains unclear whether they are robust to shifts in setting (sometimes to more realistic ones), training data size (e.g. from large to small sets) and domain. In this work, we design and conduct the first large-scale robustness study of history modeling approaches for CQA. We find that high benchmark scores do not necessarily translate to strong robustness, and that various methods can perform extremely differently under different settings. Equipped with the insights from our study, we design a novel prompt-based history modeling approach, and demonstrate its strong robustness across various settings. Our approach is inspired by existing methods that highlight historic answers in the passage. However, instead of highlighting by modifying the passage token embeddings, we add textual prompts directly in the passage text. Our approach is simple, easy-to-plug into practically any model, and highly effective, thus we recommend it as a starting point for future model developers. We also hope that our study and insights will raise awareness to the importance of robustness-focused evaluation, in addition to obtaining high leaderboard scores, leading to better CQA systems.

Hampering the interpretation of benchmark scores, evaluation data contamination has become a growing concern in the evaluation of LLMs, and an active area of research studies its effects. While evaluation data contamination is easily understood intuitively, it is surprisingly difficult to define precisely which samples should be considered contaminated and, consequently, how it impacts benchmark scores. We propose that these questions should be addressed together and that contamination metrics can be assessed based on whether models benefit from the examples they mark contaminated. We propose a novel analysis method called ConTAM, and show with a large scale survey of existing and novel n-gram based contamination metrics across 13 benchmarks and 7 models from 2 different families that ConTAM can be used to better understand evaluation data contamination and its effects. We find that contamination may have a much larger effect than reported in recent LLM releases and benefits models differently at different scales. We also find that considering only the longest contaminated substring provides a better signal than considering a union of all contaminated substrings, and that doing model and benchmark specific threshold analysis greatly increases the specificity of the results. Lastly, we investigate the impact of hyperparameter choices, finding that, among other things, both using larger values of n and disregarding matches that are infrequent in the pre-training data lead to many false negatives. With ConTAM, we provide a method to empirically ground evaluation data contamination metrics in downstream effects. With our exploration, we shed light on how evaluation data contamination can impact LLMs and provide insight into the considerations important when doing contamination analysis. We end our paper by discussing these in more detail and providing concrete suggestions for future work.

This paper investigates the fragility of Large Language Models (LLMs) in generalizing to novel inputs, specifically focusing on minor perturbations in well-established benchmarks (e.g., slight changes in question format or distractor length). Despite high benchmark scores, LLMs exhibit significant accuracy drops and unexpected biases (e.g., preference for longer distractors) when faced with these minor but content-preserving modifications. For example, Qwen 2.5 1.5B's MMLU score rises from 60 to 89 and drops from 89 to 36 when option lengths are changed without altering the question. Even GPT-4 experiences a 25-point accuracy loss when question types are changed, with a 6-point drop across all three modification categories. These analyses suggest that LLMs rely heavily on superficial cues rather than forming robust, abstract representations that generalize across formats, lexical variations, and irrelevant content shifts. This work aligns with the ACL 2025 theme track on the Generalization of NLP models, proposing a "Generalization Stress Test" to assess performance shifts under controlled perturbations. The study calls for reevaluating benchmarks and developing more reliable evaluation methodologies to capture LLM generalization abilities better.

Despite high benchmark scores, Large Language Models (LLMs) often fail simple problem, raising a critical question: Do LLMs learn mathematical principles or merely memorize patterns? Rather than designing increasingly complex benchmarks like recent works, we investigate this using elementary two-integer addition (0 to 2^{64}), probing two core properties: commutativity (A+B=B+A) and compositional generalization (via isomorphic symbolic mappings, e.g., 7 rightarrow y). While state-of-the-art LLMs achieve 73.8-99.8\% accuracy on numerical addition, performance collapses to leq7.5\% under symbolic mapping, indicating failure to generalize learned rules. Non-monotonic performance scaling with digit count and frequent commutativity violations (over 1,700 cases of A+B neq B+A) further support this. Explicitly providing addition rules degrades performance by 81.2\% on average, while self-explanation maintains baseline accuracy, suggesting LLM arithmetic processing is misaligned with human-defined principles. Our findings indicate current LLMs rely on memory pattern over genuine rule learning, highlighting architectural limitations and the need for new approaches to achieve true mathematical reasoning.

Agent benchmarks are increasingly used to compare large language models (LLMs) and guide deployment decisions, yet benchmark scores are meaningful only if they measure model capability rather than properties of the evaluation pipeline. We identify a double measurement confound: execution-critical decisions are performed by a fixed scaffold instead of the model, while the scorer evaluates outputs using criteria that may not reflect task correctness. We unify these issues within a measurement-theoretic framework that characterizes when benchmark scores can be interpreted as evidence of model capability, and instantiate it with an audit-and-repair protocol that (i) transfers execution-critical decisions from the scaffold to the model, (ii) replaces shape-based evaluation with seeded ground-truth scoring, and (iii) reports reliability beyond the mean through worst-case and tail-risk metrics. Experiments on ComtradeBench show that the joint intervention transforms a nearly flat leaderboard into a reliability spectrum that distinguishes both average performance and robustness across seeds. Applying the audit to existing benchmarks further shows that scorer validity is benchmark-specific, whereas scaffold ownership is an uncontrolled axis wherever we probed it. Our results suggest that benchmark scores should be interpreted together with their scaffolding level, scoring criterion, and reliability profile, providing a practical framework for more valid evaluation of LLM agents.

Before deploying a language model (LM) within a given domain, it is important to measure its tendency to generate factually incorrect information in that domain. Existing factual generation evaluation methods focus on facts sampled from the LM itself, and thus do not control the set of evaluated facts and might under-represent rare and unlikely facts. We propose FACTOR: Factual Assessment via Corpus TransfORmation, a scalable approach for evaluating LM factuality. FACTOR automatically transforms a factual corpus of interest into a benchmark evaluating an LM's propensity to generate true facts from the corpus vs. similar but incorrect statements. We use our framework to create two benchmarks: Wiki-FACTOR and News-FACTOR. We show that: (i) our benchmark scores increase with model size and improve when the LM is augmented with retrieval; (ii) benchmark score correlates with perplexity, but the two metrics do not always agree on model ranking; and (iii) when perplexity and benchmark score disagree, the latter better reflects factuality in open-ended generation, as measured by human annotators. We make our data and code publicly available in https://github.com/AI21Labs/factor.

Visual-Language-Action (VLA) models report impressive success rates on robotic manipulation benchmarks, yet these results may mask fundamental weaknesses in robustness. We perform a systematic vulnerability analysis by introducing controlled perturbations across seven dimensions: objects layout, camera viewpoints, robot initial states, language instructions, light conditions, background textures and sensor noise. We comprehensively analyzed multiple state-of-the-art models and revealed consistent brittleness beneath apparent competence. Our analysis exposes critical weaknesses: models exhibit extreme sensitivity to perturbation factors, including camera viewpoints and robot initial states, with performance dropping from 95% to below 30% under modest perturbations. Surprisingly, models are largely insensitive to language variations, with further experiments revealing that models tend to ignore language instructions completely. Our findings challenge the assumption that high benchmark scores equate to true competency and highlight the need for evaluation practices that assess reliability under realistic variation.

We show that PLDR-LLMs pretrained at self-organized criticality exhibit reasoning at inference time. The characteristics of PLDR-LLM deductive outputs at criticality is similar to second-order phase transitions. At criticality, the correlation length diverges, and the deductive outputs attain a metastable steady state. The steady state behaviour suggests that deductive outputs learn representations equivalent to scaling functions, universality classes and renormalization groups from the training dataset, leading to generalization and reasoning capabilities in the process. We can then define an order parameter from the global statistics of the model's deductive output parameters at inference. The reasoning capabilities of a PLDR-LLM is better when its order parameter is close to zero at criticality. This observation is supported by the benchmark scores of the models trained at near-criticality and sub-criticality. Our results provide a self-contained explanation on how reasoning manifests in large language models, and the ability to reason can be quantified solely from global model parameter values of the deductive outputs at steady state, without any need for evaluation of curated benchmark datasets through inductive output for reasoning and comprehension.

Vision-language models (VLMs) achieve strong performance on spatial reasoning benchmarks, yet it remains unclear whether this reflects structured 3D understanding or reliance on statistical shortcuts in natural images. We introduce a representation-level analysis framework that constructs minimal contrastive pairs to measure how spatial axes are organized and disentangled within VLM embeddings. Our analysis across multiple model families reveals a consistent vertical-distance entanglement: models conflate vertical image position with distance, mirroring the perspective bias of natural photographs. This bias produces a significant accuracy gap between perspective-consistent and counter-heuristic examples, and intensifies under data scaling even as overall benchmark accuracy improves. We further show that models with similar benchmark scores can exhibit different internal representations, and that these differences predict accuracy and robustness across diverse spatial reasoning benchmarks. To isolate this bias from evaluation-set skew, we introduce SpatialTunnel, a synthetic benchmark designed to expose spatial shortcut biases by removing common correlations present in natural images. Experiments confirm that the entanglement is model-intrinsic, and that models with well-separated spatial axes exhibit greater robustness, suggesting that well-structured spatial representations lead to more reliable spatial reasoning across diverse benchmarks. Code and benchmark are available on the project page: https://cheolhong0916.github.io/whyfarlooksup.github.io/.

Coding agents have become a major mode of software engineering, but the benchmarks we use to compare them were designed in a pre-agent era: they collapse model, harness, and environment into a single end-to-end score, typically computed against one reference solution, with no component-level signal for iteration. We argue that current coding benchmarks are misaligned with agentic software engineering. A coding agent in practice is not a model: it is a system harness -- a composite of models, harnesses, contexts, environments, and feedback signals, any one of which can move the benchmark score by margins comparable to those between adjacent model generations. We discuss three symptoms: (i) benchmark scores conflate the model with the rest of the harness; (ii) grading against a single reference solution penalises equally valid alternatives; and (iii) the absence of signal at the level of individual harness components makes the end-to-end system score difficult to iterate on.

Recent autonomous penetration testing papers report high benchmark scores while adding multi-component security harnesses around frontier LLMs. Because these systems often change both architecture and backbone model, it is difficult to tell how much performance comes from the harness rather than from the underlying model. This paper presents a controlled study on the 104-task XBOW benchmark using default coding CLI agents as plain-agent baselines. We first run Codex, OpenCode, and Pi with the same GPT-5 model, budget, target interface, and scoring rule. This phase identifies the strongest same-model baseline and tests whether security-specific prompt variants improve its observed score. We then compare the default Codex scaffold with published MAPTA and PentestGPT V2 results under the closest available model matches. Finally, we repeat the plain-agent experiment with GPT-5.2 and GPT-5.5 to measure model scaling inside the same scaffold. The results show a mixed but practical picture. Specialised harnesses can add measurable benchmark lift and may improve cost efficiency, but plain coding agents already solve a large share of the benchmark; repeated plain-agent runs can match or exceed some published architecture scores in union coverage, and newer models substantially improve the same scaffold. Future evaluations should report model-matched plain-agent baselines before attributing benchmark gains to architecture design alone.

Long-horizon agents are beginning to automate complete workflows that produce code, reports, and research artifacts. Medical imaging workflows are multi-stage and data-sensitive, while expert trajectories remain scarce and difficult to share. Structured benchmarks can localize failures through stage-level rubrics, but standard post-training discards these diagnostics before the next training round. We present Benchmark-as-Teacher (BaT), a recursive self-improvement system for agent post-training. BaT contains two linked components: the asynchronous Stage Bank data pipeline and BiCuRL (Bilevel Curriculum Reinforcement Learning), its self-improving post-training method. Stage Bank synthesizes content-isolated training states outside the policy-update loop. BiCuRL uses a fixed held-out evaluation to select the next stage curriculum, verifies rollouts with task rubrics, updates the policy with GRPO, and returns the candidate checkpoint to evaluation. On AutoMedBench-Lite, BaT-4B and BaT-9B more than double the Overall scores of their Qwen Instruct baselines. BaT-9B Agent reaches 79.6 Overall, exceeding Claude Opus 4.6 with Claude Code at 77.5.

Recent observations have underscored a disparity between the inflated benchmark scores and the actual performance of LLMs, raising concerns about potential contamination of evaluation benchmarks. This issue is especially critical for closed-source models and certain open-source models where training data transparency is lacking. In this paper we study data contamination by proposing two methods tailored for both open-source and proprietary LLMs. We first introduce a retrieval-based system to explore potential overlaps between evaluation benchmarks and pretraining corpora. We further present a novel investigation protocol named Testset Slot Guessing (TS-Guessing), applicable to both open and proprietary models. This approach entails masking a wrong answer in a multiple-choice question and prompting the model to fill in the gap. Additionally, it involves obscuring an unlikely word in an evaluation example and asking the model to produce it. We find that certain commercial LLMs could surprisingly guess the missing option in various test sets. Specifically, in the TruthfulQA benchmark, we find that LLMs exhibit notable performance improvement when provided with additional metadata in the benchmark. Further, in the MMLU benchmark, ChatGPT and GPT-4 demonstrated an exact match rate of 52\% and 57\%, respectively, in guessing the missing options in benchmark test data. We hope these results underscore the need for more robust evaluation methodologies and benchmarks in the field.

Large language models (LLMs) showcase increasingly impressive English benchmark scores, however their performance profiles remain inconsistent across multilingual settings. To address this gap, we introduce PolyPrompt, a novel, parameter-efficient framework for enhancing the multilingual capabilities of LLMs. Our method learns a set of trigger tokens for each language through a gradient-based search, identifying the input query's language and selecting the corresponding trigger tokens which are prepended to the prompt during inference. We perform experiments on two ~1 billion parameter models, with evaluations on the global MMLU benchmark across fifteen typologically and resource diverse languages, demonstrating accuracy gains of 3.7%-19.9% compared to naive and translation-pipeline baselines.

Large language models (LLMs) are now ubiquitous in user-facing applications, yet they still generate undesirable toxic outputs, including profanity, vulgarity, and derogatory remarks. Although numerous detoxification methods exist, most apply broad, surface-level fixes and can therefore easily be circumvented by jailbreak attacks. In this paper we leverage sparse autoencoders (SAEs) to identify toxicity-related directions in the residual stream of models and perform targeted activation steering using the corresponding decoder vectors. We introduce three tiers of steering aggressiveness and evaluate them on GPT-2 Small and Gemma-2-2B, revealing trade-offs between toxicity reduction and language fluency. At stronger steering strengths, these causal interventions surpass competitive baselines in reducing toxicity by up to 20%, though fluency can degrade noticeably on GPT-2 Small depending on the aggressiveness. Crucially, standard NLP benchmark scores upon steering remain stable, indicating that the model's knowledge and general abilities are preserved. We further show that feature-splitting in wider SAEs hampers safety interventions, underscoring the importance of disentangled feature learning. Our findings highlight both the promise and the current limitations of SAE-based causal interventions for LLM detoxification, further suggesting practical guidelines for safer language-model deployment.

LLMs demonstrate promising performance in software vulnerability detection after fine-tuning. However, it remains unclear whether these gains reflect a genuine understanding of vulnerability root causes or merely an exploitation of functional patterns. In this paper, we identify a critical failure mode termed the "semantic trap," where fine-tuned LLMs achieve high detection scores by associating certain functional domains with vulnerability likelihood rather than reasoning about the underlying security semantics. To systematically evaluate this phenomenon, we propose TrapEval, a comprehensive evaluation framework designed to disentangle vulnerability root cause from functional pattern. TrapEval introduces two complementary datasets derived from real-world open-source projects: V2N, which pairs vulnerable code with unrelated benign code, and V2P, which pairs vulnerable code with its corresponding patched version, forcing models to distinguish near-identical code that differs only in subtle security-critical logic. Using TrapEval, we fine-tune five representative state-of-the-art LLMs across three model families and evaluate them under cross-dataset testing, semantic-preserving perturbations, and varying degrees of semantic gap measured by CodeBLEU. Our empirical results reveal that, despite improvements in metrics, fine-tuned LLMs consistently struggle to distinguish vulnerable code from its patched counterpart, exhibit severe robustness degradation under minor semantic-preserving transformations, and rely heavily on functional-context shortcuts when the semantic gap is small. These findings provide strong evidence that current fine-tuning practices often fail to impart true vulnerability reasoning. Our findings serve as a wake-up call: high benchmark scores on traditional datasets may be illusory, masking the model's inability to understand the true causal logic of vulnerabilities.

Benchmarks are often used as a standard to understand LLM capabilities in different domains. However, aggregate benchmark scores provide limited insight into compositional skill gaps of LLMs and how to improve them. To make these weaknesses visible, we propose Scaffolded Task Design (STaD) framework. STaD generates controlled variations of benchmark tasks based on the concept of scaffolding, which introduces structured, incremental support in a step-by-step manner. Rather than inspecting failures individually, this approach enables systematic and scalable probing of model behavior by identifying the specific reasoning skill compositions they lack. Treating the LLM as a black box, our experiments on six models of varying sizes reveal multiple failure points in three reasoning benchmarks and highlight each model's unique and distinct skill gaps.

Scaling laws predict loss from compute but not how capabilities interact. We measure the coupling between reasoning and truthfulness across 63 base models from 16 families and find a regime change invisible to loss curves: below a family-dependent critical scale N_c, capabilities anticorrelate; above it, they cooperate. N_c approx 3.5B parameters [2.9B, 13.4B] (bootstrap 95% CI), but model size is not the only variable that determines phase. Architecture, data curation, and training recipe each shift N_c independently: curated training eliminated the coupling dip between Qwen generations (0.025 to 0.830 at matched scale), Gemma-4 at 4B achieves coupling 0.871, characteristic of 13B+ standard-trained models, through distillation and architectural innovation, and Phi at 1B matches web-trained coupling at 10B through data curation alone. Width normalization eliminates the anticorrelation across all tested families, supporting an output-projection bottleneck. Internally, 38 of 40 models show zero competing attention heads. A sparse-regression ODE cross-predicts held-out Llama-2 at 5.6% error. The diagnostic requires no model internals -- only public benchmark scores across a model family. The cooperative regime extends to the frontier (r = +0.72, 34 models, 10 labs). Code, data, and an open-source activation-steering tool for any open-weight model are released alongside an interactive dashboard that diagnoses any model's coupling phase, suggests concrete interventions (data curation, width, benchmark rotation), and provides ODE scaling predictions, frontier diagnostics, and eigenstructure analysis: https://zehenlabs.com/cape/.

Progress in LLMs is increasingly measured through standardized benchmarks, where state-of-the-art improvements are often separated by fractions of a percentage point. At the same time, the computational cost of evaluating modern LLMs has driven widespread adoption of specialized inference backends, software systems that execute trained models efficiently at inference time. While critical for scalability, system-level optimizations, such as custom CUDA kernels and reduced-precision arithmetic, can alter token probabilities and introduce non-determinism, possibly cascading into divergent generation. In this work, we first survey the inference landscape, identifying 200 distinct engines, and analyze 35,000 ML publications, finding that the specific inference stack is rarely reported despite this widespread diversity. We then present a systematic empirical study of how inference backends affect LLM benchmark results. Holding model weights, decoding parameters, and hardware constant, we evaluate five widely used inference engines, including vLLM, SGLang, and llama.cpp, across multiple open-weight models and established benchmarks. We show that the choice of backend alone can shift benchmark scores by up to 16.6 percentage points and induce high rates of output disagreement. By isolating backend optimizations and tracing the execution pipeline, we find this divergence is driven by system-level optimizations like prefix caching and CUDA graphs, custom kernels, and engine-specific defaults in logit processing. Our findings identify the inference backend as a previously unreported but consequential hyperparameter in the evaluation of LLM and advocate standardized reporting of inference stacks to improve the reproducibility and interpretability of benchmark comparisons.

We introduce PerceptionRubrics, a rubric-based evaluation framework that addresses the gap between saturated benchmark scores and real-world brittleness. Shifting evaluation from holistic semantic matching to rigorous atomic auditing, PerceptionRubrics pairs 1,038 information-dense images with over 12,000 instance-specific rubrics. These criteria are derived from golden captions constructed via a novel Circular Peer-Review consensus pipeline and then distilled into a dual-stream system of Must-Right (essential facts) and Easy-Wrong (fine-grained details) rubrics. Crucially, PerceptionRubrics implements a Gated Scoring mechanism: unlike linear averages, failure on mandatory visual facts triggers sharp binary penalties. Extensive evaluation yields critical insights: (1) The Reliability Gap: models often verify fragmented elements correctly yet fail strict conjunctive constraints, exposing brittleness in dense domains; (2) Open-Closed Stratification: contrary to reasoning trends, we reveal a persistent 8% perception deficit between open-source and proprietary frontiers; and (3) Human-Aligned Rigor: our gated metrics substantially out-align conventional benchmarks, validating that strict perceptual fidelity is the prerequisite for reliable generation.

Benchmarks like MMLU suggest flagship language models approach factuality saturation above 90\%. LLMpedia shows this picture is incomplete. We materialize {sim}1.3M encyclopedia articles entirely from parametric memory across three model families, then audit every claim against Wikipedia and curated web evidence. For gpt-5-mini, the verifiable true rate is 68.4\% on Wikipedia-covered subjects - more than 21\,pp below MMLU - and the gap is driven by unverifiability (30.5\%), not refutation (1.2\%). Beyond Wikipedia, frontier articles audited against curated web evidence reach 57.6\%; Wikipedia covers only 56.7\% of model-surfaced subjects, and three model families overlap in just 7.3\% of subject choices. In a retrieval-trap benchmark inspired by prior analysis of Grokipedia, LLMpedia is more factual at roughly half the textual similarity to Wikipedia. Every prompt, article, and verdict is released. Data, code, interface: https://llmpedia.net.

As LLMs are increasingly deployed as agents, reliable assessment of their agentic capabilities has become essential. However, reported benchmark scores often jointly reflect model capability and the implementation choices each benchmark is packaged with, making cross-benchmark results difficult to interpret as clean measurements of the underlying model. In this work, we present a unified framework for the fair evaluation of LLM agentic capabilities. Driven by a unified configuration system, the framework integrates diverse benchmarks into a standardized instruction-tool-environment format, executes agents through a fixed ReAct-style architecture within a controllable sandbox, and provides an optional offline setting that replaces volatile live environments with curated snapshots, so that framework effects and environment effects can be analyzed separately. Building on this, we unify the evaluation methodology under each benchmark's original task-success criteria, while introducing unified metrics for resource consumption and a taxonomy for decision- and execution-level failure attribution. Within this framework, we adapt 7 widely used benchmarks spanning 24 domains across single-agent, multi-agent, and safety-critical scenarios, and conduct a large-scale empirical analysis over 400K rollouts and 5B tokens on 15 models. The results show that scaffold choice and environmental volatility materially shift benchmark outcomes in both directions, allowing our framework to disentangle intrinsic LLM capabilities from framework- and environment-induced artifacts. We further demonstrate its extensibility as a secure testbed for safety-critical domains. Codes and benchmarks at are available at https://github.com/whfeLingYu/A-Unified-Framework-for-the-Evaluation-of-LLM-Agentic-Capabilities, https://huggingface.co/datasets/whfeLingYu/Unified_Agent_Framework.

Detecting a traffic anomaly does not establish whether a video-language model can explain what happened, localize it in time, or identify its causes. We introduce TAR (Traffic Anomaly Reasoning) and TAR-Bench, paired resources for training and evaluating these complementary capabilities across 10 tasks spanning question answering, temporal reasoning, and scene understanding. TAR contains 44,040 automatically generated annotations with chain-of-thought traces for 3,670 CCTV videos from eight public datasets. TAR-Bench provides 960 human-curated annotations for 80 held-out clips from 17 public YouTube videos. Evaluation of eleven vision-language models reveals a gap between question-answering performance and temporal or scene reasoning. Progressively adding task groups during supervised fine-tuning improves aggregate performance on both Cosmos-Reason2-8B and Qwen3-VL-8B-Instruct. Training on all 10 tasks raises their mean benchmark scores from 34.3 to 55.7 and from 30.9 to 53.9, respectively. These results support joint training across complementary tasks as a promising approach to traffic anomaly understanding, while highlighting persistent limitations in temporal precision and causal attribution. TAR and TAR-Bench serve as the official training and in-domain evaluation resources for AI City Challenge 2026 Track 3.The dataset is available at https://huggingface.co/datasets/nvidia/PhysicalAI-Traffic-Anomaly-Reasoning

For classification and regression on tabular data, the dominance of gradient-boosted decision trees (GBDTs) has recently been challenged by often much slower deep learning methods with extensive hyperparameter tuning. We address this discrepancy by introducing (a) RealMLP, an improved multilayer perceptron (MLP), and (b) strong meta-tuned default parameters for GBDTs and RealMLP. We tune RealMLP and the default parameters on a meta-train benchmark with 118 datasets and compare them to hyperparameter-optimized versions on a disjoint meta-test benchmark with 90 datasets, as well as the GBDT-friendly benchmark by Grinsztajn et al. (2022). Our benchmark results on medium-to-large tabular datasets (1K--500K samples) show that RealMLP offers a favorable time-accuracy tradeoff compared to other neural baselines and is competitive with GBDTs in terms of benchmark scores. Moreover, a combination of RealMLP and GBDTs with improved default parameters can achieve excellent results without hyperparameter tuning. Finally, we demonstrate that some of RealMLP's improvements can also considerably improve the performance of TabR with default parameters.

Recent work pairs LLMs with evolutionary search to iteratively generate, modify, and select code using task-specific feedback. These systems have produced strong results in mathematical discovery and algorithm design, yet a fundamental question remains: what do they actually evolve? Progress is typically summarized by the best score a run reaches under a task-specific evaluator, but that score can reflect several different mechanisms: new algorithmic structure, re-tuning an existing strategy, recombining ideas already in the model's internal knowledge, or overfitting to the evaluator. Distinguishing these mechanisms requires inspecting the search process itself, not only its final outcome. We introduce EvoTrace, a dataset of evolutionary coding traces spanning four evolutionary frameworks, reasoning and non-reasoning models, and 16 tasks across mathematics and algorithm design. To analyze these traces, we develop EvoReplay, a replay-based methodology that reconstructs the local search states behind high-scoring solutions and tests controlled interventions, including adjusting constants, removing program components and substituting models or prompting contexts. We annotate every code edit in EvoTrace with one of nine recurring edit types using an LLM-as-judge pipeline validated against blind human re-annotation. Across EvoTrace, most score gains come from a small subset of these edit types. We further find a deterministic cycling pattern: about 30% of code lines added during search are byte-identical re-introductions of previously-deleted lines, present throughout nearly every run. These results show that benchmark gains in evolutionary coding agents can arise from qualitatively different mechanisms, only some of which correspond to new algorithmic structure. EvoTrace enables more diagnostic evaluation of evolutionary coding agents beyond final benchmark scores.

The rapid capability gains of frontier language models are widely attributed to improved reasoning abilities, yet this cannot be verified as raw CoT traces in closed-source systems are hidden. By registering a simple custom tool through a standard API feature, we induce frontier models to externalize intermediate reasoning. Because these traces may reflect post-hoc rationalization rather than genuine reasoning, we first evaluate against native CoT on open-source models and extend to closed-source frontier models including GPT-6 Astra. We find that the extracted reasoning matches native reasoning performance and substantially outperforms no-reasoning baselines, across competition mathematics, science, and code generation. We then characterize how frontier models structure their intermediate reasoning. Across token efficiency, reasoning-step types, and induced reasoning trees, we identify systematic differences in how models externalize, compress, and organize reasoning. We find that Astra exhibits token-efficient directed reasoning, selecting a correct trajectory earlier, while resolving elementary steps internally and externalizing only crucial reasoning. These findings provide a behavioral lens on frontier-model reasoning beyond benchmark scores.

Robot learning tasks are extremely compute-intensive and hardware-specific. Thus the avenues of tackling these challenges, using a diverse dataset of offline demonstrations that can be used to train robot manipulation agents, is very appealing. The Train-Offline-Test-Online (TOTO) Benchmark provides a well-curated open-source dataset for offline training comprised mostly of expert data and also benchmark scores of the common offline-RL and behaviour cloning agents. In this paper, we introduce DiffClone, an offline algorithm of enhanced behaviour cloning agent with diffusion-based policy learning, and measured the efficacy of our method on real online physical robots at test time. This is also our official submission to the Train-Offline-Test-Online (TOTO) Benchmark Challenge organized at NeurIPS 2023. We experimented with both pre-trained visual representation and agent policies. In our experiments, we find that MOCO finetuned ResNet50 performs the best in comparison to other finetuned representations. Goal state conditioning and mapping to transitions resulted in a minute increase in the success rate and mean-reward. As for the agent policy, we developed DiffClone, a behaviour cloning agent improved using conditional diffusion.

When using Large Language Models (LLMs) to support Knowledge Graph Engineering (KGE), one of the first indications when searching for an appropriate model is its size. According to the scaling laws, larger models typically show higher capabilities. However, in practice, resource costs are also an important factor and thus it makes sense to consider the ratio between model performance and costs. The LLM-KG-Bench framework enables the comparison of LLMs in the context of KGE tasks and assesses their capabilities of understanding and producing KGs and KG queries. Based on a dataset created in an LLM-KG-Bench run covering 26 open state-of-the-art LLMs, we explore the model size scaling laws specific to KGE tasks. In our analyses, we assess how benchmark scores evolve between different model size categories. Additionally, we inspect how the general score development of single models and families of models correlates to their size. Our analyses revealed that, with a few exceptions, the model size scaling laws generally also apply to the selected KGE tasks. However, in some cases, plateau or ceiling effects occurred, i.e., the task performance did not change much between a model and the next larger model. In these cases, smaller models could be considered to achieve high cost-effectiveness. Regarding models of the same family, sometimes larger models performed worse than smaller models of the same family. These effects occurred only locally. Hence it is advisable to additionally test the next smallest and largest model of the same family.

We introduce CAT, a framework designed to evaluate and visualize the interplay of accuracy and response consistency of Large Language Models (LLMs) under controllable input variations, using multiple-choice (MC) benchmarks as a case study. Current evaluation practices primarily focus on model capabilities such as accuracy or benchmark scores and, more recently, measuring consistency is being considered an essential property for deploying LLMs in high-stake, real-world applications. We argue in this paper that although both dimensions should still be evaluated independently, their inter-dependency also need to be considered for a more nuanced evaluation of LLMs. At the core of CAT are the Consistency-Accuracy Relation (CAR) curves, which visualize how model accuracy varies with increasing consistency requirements, as defined by the Minimum-Consistency Accuracy (MCA) metric. We further propose the Consistency-Oriented Robustness Estimate (CORE) index, a global metric that combines the area and shape of the CAR curve to quantify the trade-off between accuracy and consistency. We present a practical demonstration of our framework across a diverse set of generalist and domain-specific LLMs, evaluated on multiple MC benchmarks. We also outline how CAT can be extended beyond MC tasks to support long-form, open-ended evaluations through adaptable scoring functions.

Evaluating LLMs is challenging, as benchmark scores often fail to capture models' real-world usefulness. Instead, users often rely on ``vibe-testing'': informal experience-based evaluation, such as comparing models on coding tasks related to their own workflow. While prevalent, vibe-testing is often too ad hoc and unstructured to analyze or reproduce at scale. In this work, we study how vibe-testing works in practice and then formalize it to support systematic analysis. We first analyze two empirical resources: (1) a survey of user evaluation practices, and (2) a collection of in-the-wild model comparison reports from blogs and social media. Based on these resources, we formalize vibe-testing as a two-part process: users personalize both what they test and how they judge responses. We then introduce a proof-of-concept evaluation pipeline that follows this formulation by generating personalized prompts and comparing model outputs using user-aware subjective criteria. In experiments on coding benchmarks, we find that combining personalized prompts and user-aware evaluation can change which model is preferred, reflecting the role of vibe-testing in practice. These findings suggest that formalized vibe-testing can serve as a useful approach for bridging benchmark scores and real-world experience.

Recent work has fine-tuned language models on chess data and reported high benchmark scores as evidence that the resulting models can understand the rules of chess, play full chess games at a professional level, or generate human-readable explanations grounded in expert knowledge. We train KinGPT, a 25M-parameter character-level language model trained only on (position, best-move) pairs, who exceeds 3B-parameter ChessGPT on a 600-puzzle mate-in-N suite and 4B-parameter C1-4B over a 20-theme puzzle benchmark. We examine several claims made in existing literature regarding chess-trained language models and assert that their impressive benchmark performance is largely explained by pattern-matching. We also demonstrate how LLM-Modulo, a verifier-in-the-loop framework, raises RedPajama 3B's best move accuracy from 1.2% to 21.2% and move generation validity from 19.3% to 95.3% on mate-in-N chess puzzles, comparable to gains achieved from ChessGPT's fine-tuning on chess-specific web corpora at a fraction of the cost. Our results illustrate how pairing a general LLM with an external verifier offers a more flexible alternative to directly training on synthetic data for well-defined domains. We open source all training/evaluation code, datasets, puzzle samples, and KinGPT model checkpoints for reproducibility.

In this paper, we propose EEVEE, the first multi-dataset test-time prompt learning framework for LLM agents, enabling test-time prompt learning under real-world task streams. Existing methods are largely designed for single-dataset settings, while real-world applications require models to handle heterogeneous input streams drawn from multiple datasets, domains, and task distributions, limiting their practical applicability. To mitigate cross-dataset interference, EEVEE introduces a router that partitions incoming inputs into task clusters and assigns them to suitable prompt configurations. This design is optimized via a router-prompt co-evolution strategy, which employs interleaved router and prompt learning phases to address their mutual dependency. Experiments across multiple datasets demonstrate that the framework improves robustness under heterogeneous data streams while maintaining single-benchmark learning capability and efficiency. Specifically, EEVEE improves average multi-benchmark scores by 10.38 and 24.32 points over Qwen3-4B-Instruct and DeepSeek-V3.2, surpassing SOTA methods GEPA and ACE by up to 37.2% and 48.2%.

Why do we build local large language models (LLMs)? What should a local LLM learn from the target language? Which abilities can be transferred from other languages? Do language-specific scaling laws exist? To explore these research questions, we evaluated 35 Japanese, English, and multilingual LLMs on 19 evaluation benchmarks for Japanese and English, taking Japanese as a local language. Adopting an observational approach, we analyzed correlations of benchmark scores, and conducted principal component analysis (PCA) on the scores to derive ability factors of local LLMs. We found that training on English text can improve the scores of academic subjects in Japanese (JMMLU). In addition, it is unnecessary to specifically train on Japanese text to enhance abilities for solving Japanese code generation, arithmetic reasoning, commonsense, and reading comprehension tasks. In contrast, training on Japanese text could improve question-answering tasks about Japanese knowledge and English-Japanese translation, which indicates that abilities for solving these two tasks can be regarded as Japanese abilities for LLMs. Furthermore, we confirmed that the Japanese abilities scale with the computational budget for Japanese text.

The research of artificial intelligence is undergoing a paradigm shift from prioritizing model innovations over benchmark scores towards emphasizing problem definition and rigorous real-world evaluation. As the field enters the "second half," the central challenge becomes real utility in long-horizon, dynamic, and user-dependent environments, where agents face context explosion and must continuously accumulate, manage, and selectively reuse large volumes of information across extended interactions. Memory, with hundreds of papers released this year, therefore emerges as the critical solution to fill the utility gap. In this survey, we provide a unified view of foundation agent memory along three dimensions: memory substrate (internal and external), cognitive mechanism (episodic, semantic, sensory, working, and procedural), and memory subject (agent- and user-centric). We then analyze how memory is instantiated and operated under different agent topologies and highlight learning policies over memory operations. Finally, we review evaluation benchmarks and metrics for assessing memory utility, and outline various open challenges and future directions.

There is a belief that learning to compress well will lead to intelligence. Recently, language modeling has been shown to be equivalent to compression, which offers a compelling rationale for the success of large language models (LLMs): the development of more advanced language models is essentially enhancing compression which facilitates intelligence. Despite such appealing discussions, little empirical evidence is present for the interplay between compression and intelligence. In this work, we examine their relationship in the context of LLMs, treating LLMs as data compressors. Given the abstract concept of "intelligence", we adopt the average downstream benchmark scores as a surrogate, specifically targeting intelligence related to knowledge and commonsense, coding, and mathematical reasoning. Across 12 benchmarks, our study brings together 30 public LLMs that originate from diverse organizations. Remarkably, we find that LLMs' intelligence -- reflected by average benchmark scores -- almost linearly correlates with their ability to compress external text corpora. These results provide concrete evidence supporting the belief that superior compression indicates greater intelligence. Furthermore, our findings suggest that compression efficiency, as an unsupervised metric derived from raw text corpora, serves as a reliable evaluation measure that is linearly associated with the model capabilities. We open-source our compression datasets as well as our data collection pipelines to facilitate future researchers to assess compression properly.

Large Language Models (LLMs) have shown remarkable capabilities, but their development has primarily focused on English and other high-resource languages, leaving many languages underserved. We present our latest Hindi-English bi-lingual LLM Mantra-14B with ~3\% average improvement in benchmark scores over both languages, outperforming models twice its size. Using a curated dataset composed of English and Hindi instruction data of 485K samples, we instruction tuned models such as Qwen-2.5-14B-Instruct and Phi-4 to improve performance over both English and Hindi. Our experiments encompassing seven different LLMs of varying parameter sizes and over 140 training attempts with varying English-Hindi training data ratios demonstrated that it is possible to significantly improve multilingual performance without compromising native performance. Further, our approach avoids resource-intensive techniques like vocabulary expansion or architectural modifications, thus keeping the model size small. Our results indicate that modest fine-tuning with culturally and locally informed data can bridge performance gaps without incurring significant computational overhead. We release our training code, datasets, and models under mit and apache licenses to aid further research towards under-represented and low-resource languages.

Large Language Models (LLMs) are now integral across various domains and have demonstrated impressive performance. Progress, however, rests on the premise that benchmark scores are both accurate and reproducible. We demonstrate that the reproducibility of LLM performance is fragile: changing system configuration such as evaluation batch size, GPU count, and GPU version can introduce significant difference in the generated responses. This issue is especially pronounced in reasoning models, where minor rounding differences in early tokens can cascade into divergent chains of thought, ultimately affecting accuracy. For instance, under bfloat16 precision with greedy decoding, a reasoning model like DeepSeek-R1-Distill-Qwen-7B can exhibit up to 9% variation in accuracy and 9,000 tokens difference in response length due to differences in GPU count, type, and evaluation batch size. We trace the root cause of this variability to the non-associative nature of floating-point arithmetic under limited numerical precision. This work presents the first systematic investigation into how numerical precision affects reproducibility in LLM inference. Through carefully controlled experiments across various hardware, software, and precision settings, we quantify when and how model outputs diverge. Our analysis reveals that floating-point precision -- while critical for reproducibility -- is often neglected in evaluation practices. Inspired by this, we develop a lightweight inference pipeline, dubbed LayerCast, that stores weights in 16-bit precision but performs all computations in FP32, balancing memory efficiency with numerical stability. Code is available at https://github.com/nanomaoli/llm_reproducibility.

Large Language Model (LLM) training often optimizes for preference alignment, rewarding outputs that are perceived as helpful and interaction-friendly. However, this preference-oriented objective can be exploited: manipulative prompts can steer responses toward user-appeasing agreement and away from truth-oriented correction. In this work, we investigate whether aligned models are vulnerable to Preference-Undermining Attacks (PUA), a class of manipulative prompting strategies designed to exploit the model's desire to please user preferences at the expense of truthfulness. We propose a diagnostic methodology that provides a finer-grained and more directive analysis than aggregate benchmark scores, using a factorial evaluation framework to decompose prompt-induced shifts into interpretable effects of system objectives (truth- vs. preference-oriented) and PUA-style dialogue factors (directive control, personal derogation, conditional approval, reality denial) within a controlled 2 times 2^4 design. Surprisingly, more advanced models are sometimes more susceptible to manipulative prompts. Beyond the dominant reality-denial factor, we observe model-specific sign reversals and interactions with PUA-style factors, suggesting tailored defenses rather than uniform robustness. These findings offer a novel, reproducible factorial evaluation methodology that provides finer-grained diagnostics for post-training processes like RLHF, enabling better trade-offs in the product iteration of LLMs by offering a more nuanced understanding of preference alignment risks and the impact of manipulative prompts.

In this report, we introduce Qwen3-ASR family, which includes two powerful all-in-one speech recognition models and a novel non-autoregressive speech forced alignment model. Qwen3-ASR-1.7B and Qwen3-ASR-0.6B are ASR models that support language identification and ASR for 52 languages and dialects. Both of them leverage large-scale speech training data and the strong audio understanding ability of their foundation model Qwen3-Omni. We conduct comprehensive internal evaluation besides the open-sourced benchmarks as ASR models might differ little on open-sourced benchmark scores but exhibit significant quality differences in real-world scenarios. The experiments reveal that the 1.7B version achieves SOTA performance among open-sourced ASR models and is competitive with the strongest proprietary APIs while the 0.6B version offers the best accuracy-efficiency trade-off. Qwen3-ASR-0.6B can achieve an average TTFT as low as 92ms and transcribe 2000 seconds speech in 1 second at a concurrency of 128. Qwen3-ForcedAligner-0.6B is an LLM based NAR timestamp predictor that is able to align text-speech pairs in 11 languages. Timestamp accuracy experiments show that the proposed model outperforms the three strongest force alignment models and takes more advantages in efficiency and versatility. To further accelerate the community research of ASR and audio understanding, we release these models under the Apache 2.0 license.

Vision-language models (VLMs) have made substantial progress across a wide range of visual question answering benchmarks, spanning visual reasoning, document understanding, and multimodal dialogue. These improvements are evident in a wide range of VLMs built on a variety of base models, alignment architectures, and training data. However, recent works show that these models trail behind in traditional image classification benchmarks, which test fine-grained visual knowledge. We test a large number of recent VLMs on fine-grained classification benchmarks and identify potential factors in the disconnect between fine-grained knowledge and other vision benchmarks. Through a series of ablation experiments, we find that using a better LLM improves all benchmark scores equally, while a better vision encoder disproportionately improves fine-grained classification performance. Furthermore, we find that the pretraining stage is also vital to fine-grained performance, particularly when the language model weights are unfrozen during pretraining. These insights pave the way for enhancing fine-grained visual understanding and vision-centric capabilities in VLMs.

Given a database S and a natural language question Q, text-to-SQL systems aim to generate an SQL query that correctly answers Q when executed against S. Currently, popular text-to-SQL benchmarks mostly assume unrestricted access to S; in practice, however, user access is often restricted, e.g., through role-based access control (RBAC) policies. This leads to a potential disconnect between benchmarking results and real-world performance: an LLM with high benchmark scores might perform poorly in an access-controlled environment, by frequently violating RBAC, or rejecting a query q that could be answered with only permitted data in S. Motivated by this, we present a comprehensive text-to-SQL benchmarking framework with realistic RBAC constraints, which features an LLM-assisted workflow that augments existing text-to-SQL benchmarks with plausible user roles and access policies. To do so, we formulate the problem of role synthesis as a structured reasoning process over the database schema, in which the LLM first infers the application context from the schema, and then derives role responsibilities and access scopes consistent with this context. This process is audited by human-in-the-loop quality control, in which domain experts perform metric-guided screening on the generated roles. Besides the augmented dataset, the proposed framework also contains evaluation metrics that identify RBAC-specific failure modes, and disentangle SQL utility from access-control compliance. We apply the proposed framework to several widely-used benchmarks, and conduct a systematic empirical study of state-of-the-art text-to-SQL systems. The results show that many solutions (especially open-weight LLMs) with high benchmarking scores under an unrestricted setting suffer sharp performance degradation once access constraints are in place, due to frequent RBAC violations.

Small language models are usually built like large ones and then squeezed onto a CPU afterwards. We did the opposite: we fixed the target first, one user, one token at a time, 4-bit weights, ordinary CPU, and chose the architecture to suit it. The result keeps full attention in only 6 of its 18 blocks. The other 12 use short convolutions whose memory is two timesteps wide no matter how long the conversation gets, so two thirds of the network never re-reads a growing cache. Trained from scratch on 59.9B tokens, the model scores 47.31 on a five-task benchmark against a bar of 42.20 that was fixed before training began. It beats GPT-2 124M, Pythia-160M, OPT-125M and GPT-neo-125M, all trained on three to six times more data, and exceeds MobileLLM-125M's published score despite that model seeing a trillion tokens. Validation bits-per-byte is 0.8685. To check the architecture rather than the training recipe, we trained a conventional all-attention model of the same size on the same data, and wrote down the winning condition before scoring either. The hybrid won the chosen quality metric by 0.81%, matched it on downstream tasks, produced a 6.3% smaller 4-bit file, and decoded 1.76x faster at 2048 tokens of context, 2.08x against an external model of similar size. In every measurement the speed advantage is near zero at an empty context and grows with length, which is what the mechanism predicts and what a merely leaner model would not show. A simple bandwidth calculation predicts only 1.17x, so memory volume alone does not explain the gap. We also report what did not work: an unmitigated 4-bit quality cost, roughly half the convolution channels ending up inert and impossible to remove, and a vocabulary larger than this model size warrants.

Posted prices for AI inference have fallen steadily since 2024, yet the measured speed of that fall depends almost entirely on the method of measurement. This paper constructs quality-adjusted price indices for the AI inference market from public data. The panel assembles 21,024 posted-price observations across 3,208 models and 86 providers and joins them to 4,605 benchmark scores through a latent quality index estimated from benchmark response patterns, so the quality ladder of the hedonic tradition is built here from evaluations in place of product characteristics. Measured by the matched-model methods that statistical agencies apply to software, inference prices fell at 0.10 log points a year. The quality-adjusted index fell at 0.73, so 87% of the decline is invisible to current methods, with direct consequences for measured competition, concentration and productivity in this market. Counted per completed task, moreover, the buyer's price stopped falling. Reasoning models raised token consumption faster than token prices fell, and the seller's and buyer's prices accordingly diverged. A pre-registered validity audit disciplines the quality measure and yields the sharpest result. Excluding contamination-flagged benchmarks leaves model rankings intact at 0.998 yet moves the index by 0.49 log points a year, so the leaderboard-stability arguments standard in AI evaluation offer no defence of economic statistics built on benchmarks. Prices, quality and the audit are fully reproducible from public sources at zero cost.

Video foundation models increasingly rely on large-scale pretraining data, yet the end-to-end data pipelines behind them remain largely closed and difficult to inspect or reuse. Researchers seeking to understand how video data recipes affect model pretraining often need to build substantial infrastructure before testing even a focused hypothesis. We present VIDAFORGE, an open research infrastructure that represents a video data recipe as an executable five-stage workflow from raw videos to training datasets. A decision in this workflow can be varied to construct alternative datasets while preserving how every sample was produced. To demon strate this research workflow, we compare data recipes with different coverage and quality in early from-scratch pretraining of Wan 2.1 and V-JEPA 2.1. Across both learning objectives, the broader-coverage recipe achieves the highest downstream benchmark scores, while loss-based evaluation favors different recipes. This study demonstrates how VidaForge connects data-recipe choices to downstream model performance. We further release VIDAFORGE-3M, containing 3.14 million scene level clips totaling 6,475 hours, with fine-grained annotations and curation signals for video data-recipe research.

Robot foundation models are trained and evaluated predominantly in English, and robot demonstration corpora do not exist for most languages. We study the addition of Greek to an open vision-language-action stack using only machine-rephrased instructions and no architecture changes. The main challenge is measurement rather than translation. Several plausible instruments produce false conclusions: a color-histogram metric rewards noise, a single-goal benchmark scores 84.6% under correct Greek and 82.6% under deliberately wrong instructions, training loss fails to predict Greek success, and single-run comparisons are dominated by seed variation. On a discriminative ninety-task suite with three seeds per arm, a multilingual text tower without Greek demonstrations remains at its wrong-instruction floor, while Greek-only training exceeds its control by at most 2.7 points. Bilingual training yields a consistent 6.7-7.1 point margin over its control and reaches about two fifths of English performance. The policy also overfits the translator's phrasing; training on seven phrasings per task approximately halves this penalty. Warm-starting from a language-adapted world model and unfreezing the text tower both degrade performance. The results support two practical requirements for low-resource robot-policy localization: build a guaranteed null before trusting a metric, and replicate low-resource-language results across seeds.

Production LLMs must handle inputs that attempt to override system instructions, bypass safety policies or elicit harmful responses. A common mitigation is a separate guardrail model. Existing reports, however, provide little evidence on Russian prompt injection or Russian surface obfuscation. We present HiveTraceGuard-Pro, a 0.6B generative guardrail LoRA-tuned from Qwen3-0.6B. It is trained on Russian and English and uses one binary scoring rule (safe/unsafe) for the final target turn. Its training corpus pairs harmful examples, where a counterpart exists, with benign examples from the same domain and applies eight obfuscation transforms to both labels. In one harness, we compare HiveTraceGuard-Pro with thirty-four other guards on nineteen benchmark groups, sixteen of which are public. Its aggregate key is 0.7432, behind 0.7641 and 0.7552 for the two higher-scoring guards. Over the sixteen public groups alone, its key is 0.7153 and four of the thirty-four other suite guards score higher. In a fifteen-model comparison, HiveTraceGuard-Pro has the highest clean Russian robustness combined-F1 (0.88) and Russian prompt-injection recall (0.999). Both results use Russian sets assembled by our team, and at least 27.1% of the prompt-injection set overlaps the training corpus. Its 14.3 ms median latency is the lowest among those fifteen models in that run. Across the suite, FPR is 0.268 and FNR is 0.156. All reported response results use a legacy standalone-reply serialization rather than the natural assistant-role path of the shipped chat template. We release the merged weights on Hugging Face under Apache-2.0. The corpus, evaluation sets and evaluation code remain internal.

Large language models (LLMs) have revolutionized automated code generation, yet the evaluation of their real-world effectiveness remains limited by static benchmarks and simplistic metrics. We present ProxyWar, a novel framework that systematically assesses code generation quality by embedding LLM-generated agents within diverse, competitive game environments. Unlike existing approaches, ProxyWar evaluates not only functional correctness but also the operational characteristics of generated programs, combining automated testing, iterative code repair, and multi-agent tournaments to provide a holistic view of program behavior. Applied to a range of state-of-the-art coders and games, our approach uncovers notable discrepancies between benchmark scores and actual performance in dynamic settings, revealing overlooked limitations and opportunities for improvement. These findings highlight the need for richer, competition-based evaluation of code generation. Looking forward, ProxyWar lays a foundation for research into LLM-driven algorithm discovery, adaptive problem solving, and the study of practical efficiency and robustness, including the potential for models to outperform hand-crafted agents. The project is available at https://github.com/xinke-wang/ProxyWar.

Industrial Retrieval-Augmented Generation (RAG) systems depend on optical character recognition (OCR) to transform visual documents into text. Existing OCR benchmarks rely on character-level metrics, which inadequately measure downstream RAG effectiveness under real-world conditions. We introduce an OCR benchmark for industrial RAG systems covering 11 challenging document types, including extreme layouts, high-resolution pages, complex or watermarked backgrounds, historical documents with non-standard reading orders, visually decorated text, and documents containing tables and mathematical formulas. Evaluating recent SOTA OCR models under a controlled OCR-first RAG pipeline shows clear performance degradation on realistic industrial documents despite strong conventional benchmark scores. We find that high OCR accuracy does not necessarily translate into strong downstream RAG performance: structural and semantic errors can cause substantial retrieval failures even when WER/CER remains low. Further analysis shows that this mismatch is category-dependent, arises through both retrieval-side and downstream generation-side failures, and remains stable across representative OCR-first pipeline choices. The benchmark is publicly available at https://github.com/Qihoo360/InduOCRBench.

Training models to act as agents that can effectively navigate and perform actions in a complex environment, such as a web browser, has typically been challenging due to lack of training data. Large language models (LLMs) have recently demonstrated some capability to navigate novel environments as agents in a zero-shot or few-shot fashion, purely guided by natural language instructions as prompts. Recent research has also demonstrated LLMs have the capability to exceed their base performance through self-improvement, i.e. fine-tuning on data generated by the model itself. In this work, we explore the extent to which LLMs can self-improve their performance as agents in long-horizon tasks in a complex environment using the WebArena benchmark. In WebArena, an agent must autonomously navigate and perform actions on web pages to achieve a specified objective. We explore fine-tuning on three distinct synthetic training data mixtures and achieve a 31\% improvement in task completion rate over the base model on the WebArena benchmark through a self-improvement procedure. We additionally contribute novel evaluation metrics for assessing the performance, robustness, capabilities, and quality of trajectories of our fine-tuned agent models to a greater degree than simple, aggregate-level benchmark scores currently used to measure self-improvement.

The rapid development of autoregressive Large Language Models (LLMs) has significantly improved the quality of generated texts, necessitating reliable machine-generated text detectors. A huge number of detectors and collections with AI fragments have emerged, and several detection methods even showed recognition quality up to 99.9% according to the target metrics in such collections. However, the quality of such detectors tends to drop dramatically in the wild, posing a question: Are detectors actually highly trustworthy or do their high benchmark scores come from the poor quality of evaluation datasets? In this paper, we emphasise the need for robust and qualitative methods for evaluating generated data to be secure against bias and low generalising ability of future model. We present a systematic review of datasets from competitions dedicated to AI-generated content detection and propose methods for evaluating the quality of datasets containing AI-generated fragments. In addition, we discuss the possibility of using high-quality generated data to achieve two goals: improving the training of detection models and improving the training datasets themselves. Our contribution aims to facilitate a better understanding of the dynamics between human and machine text, which will ultimately support the integrity of information in an increasingly automated world.

This report presents the dataset and the evaluation setup of the Sound Event Localization & Detection (SELD) task for the DCASE 2020 Challenge. The SELD task refers to the problem of trying to simultaneously classify a known set of sound event classes, detect their temporal activations, and estimate their spatial directions or locations while they are active. To train and test SELD systems, datasets of diverse sound events occurring under realistic acoustic conditions are needed. Compared to the previous challenge, a significantly more complex dataset was created for DCASE 2020. The two key differences are a more diverse range of acoustical conditions, and dynamic conditions, i.e. moving sources. The spatial sound scenes are created using real room impulse responses captured in a continuous manner with a slowly moving excitation source. Both static and moving sound events are synthesized from them. Ambient noise recorded on location is added to complete the generation of scene recordings. A baseline SELD method accompanies the dataset, based on a convolutional recurrent neural network, to provide benchmark scores for the task. The baseline is an updated version of the one used in the previous challenge, with input features and training modifications to improve its performance.

We show that Large Language Model from Power Law Decoder Representations (PLDR-LLM) is a foundational model whose deductive outputs are invariant tensors up to a small perturbation. PLDR-LLM learns a singularity condition for the deductive outputs that enable the once-inferred energy-curvature tensor G_{LM} to replace the deep neural network of power law graph attention (PLGA) generating the deductive outputs at inference. We demonstrate that a cache for G_{LM} (G-cache) and KV-cache can be implemented in a straightforward manner to improve the inference time. The invariance and generalizable nature of deductive outputs is at a very high fidelity where deductive outputs have same RMSE and determinant values up to 15 decimal places after caching, and zero-shot benchmark scores remain unchanged. Ablation studies show that learned deductive outputs have distinct loss and accuracy characteristics from models pretrained with transferred, randomly initialized or identity tensors as a constant tensor operator and an LLM with scaled-dot product attention (SDPA) is a special case of PLDR-LLM where G_{LM} is predefined as identity. The observed invariance characteristic introduces a novel asymmetry between training and inference phases with caching. We outline observed common characteristics of the deductive outputs for the learned singularity condition. We provide an implementation of a training and inference framework for PLDR-LLM with KV-cache and G-cache.

In this paper, we introduce PolyMath, a multilingual mathematical reasoning benchmark covering 18 languages and 4 easy-to-hard difficulty levels. Our benchmark ensures difficulty comprehensiveness, language diversity, and high-quality translation, making it a highly discriminative multilingual mathematical benchmark in the era of reasoning LLMs. We conduct a comprehensive evaluation for advanced LLMs and find that even Deepseek-R1-671B and Qwen-QwQ-32B, achieve only 43.4 and 41.8 benchmark scores, with less than 30% accuracy under the highest level. From a language perspective, our benchmark reveals several key challenges of LLMs in multilingual reasoning: (1) Reasoning performance varies widely across languages for current LLMs; (2) Input-output language consistency is low in reasoning LLMs and may be correlated with performance; (3) The thinking length differs significantly by language for current LLMs. Additionally, we demonstrate that controlling the output language in the instructions has the potential to affect reasoning performance, especially for some low-resource languages, suggesting a promising direction for improving multilingual capabilities in LLMs.

The key-value (KV) cache accelerates LLMs decoding by storing KV tensors from previously generated tokens. It reduces redundant computation at the cost of increased memory usage. To mitigate this overhead, existing approaches compress KV tensors into lower-bit representations; however, quantization errors can accumulate as more tokens are generated, potentially resulting in undesired outputs. In this paper, we introduce SQuat (Subspace-orthogonal KV cache quantization). It first constructs a subspace spanned by query tensors to capture the most critical task-related information. During key tensor quantization, it enforces that the difference between the (de)quantized and original keys remains orthogonal to this subspace, minimizing the impact of quantization errors on the attention mechanism's outputs. SQuat requires no model fine-tuning, no additional calibration dataset for offline learning, and is grounded in a theoretical framework we develop. Through numerical experiments, we show that our method reduces peak memory by 2.17 to 2.82, improves throughput by 2.45 to 3.60, and achieves more favorable benchmark scores than existing KV cache quantization algorithms.

Large language models (LLMs) demonstrate remarkable reasoning capabilities in tasks such as algorithmic coding and mathematical problem-solving. Recent methods have improved reasoning through expanded corpus and multistage training combining reinforcement learning and supervised fine-tuning. Although some methods suggest that small but targeted dataset can incentivize reasoning via only distillation, a reasoning scaling laws is still taking shape, increasing computational costs. To address this, we propose a data-efficient distillation framework (DED) that optimizes the Pareto frontier of reasoning distillation. Inspired by the on-policy learning and diverse roll-out strategies of reinforcement learning, the key idea of our approach is threefold: (1) We identify that benchmark scores alone do not determine an effective teacher model. Through comprehensive comparisons of leading reasoning LLMs, we develop a method to select an optimal teacher model. (2) While scaling distillation can enhance reasoning, it often degrades out-of-domain performance. A carefully curated, smaller corpus achieves a balanced trade-off between in-domain and out-of-domain capabilities. (3) Diverse reasoning trajectories encourage the student model to develop robust reasoning skills. We validate our method through evaluations on mathematical reasoning (AIME 2024/2025, MATH-500) and code generation (LiveCodeBench), achieving state-of-the-art results with only 0.8k carefully curated examples, bypassing the need for extensive scaling. Our systematic analysis demonstrates that DED outperforms existing methods by considering factors beyond superficial hardness, token length, or teacher model capability. This work offers a practical and efficient pathway to advanced reasoning while preserving general capabilities.

Large language model (LLM) developers aim for their models to be honest, helpful, and harmless. However, when faced with malicious requests, models are trained to refuse, sacrificing helpfulness. We show that frontier LLMs can develop a preference for dishonesty as a new strategy, even when other options are available. Affected models respond to harmful requests with outputs that sound harmful but are subtly incorrect or otherwise harmless in practice. This behavior emerges with hard-to-predict variations even within models from the same model family. We find no apparent cause for the propensity to deceive, but we show that more capable models are better at executing this strategy. Strategic dishonesty already has a practical impact on safety evaluations, as we show that dishonest responses fool all output-based monitors used to detect jailbreaks that we test, rendering benchmark scores unreliable. Further, strategic dishonesty can act like a honeypot against malicious users, which noticeably obfuscates prior jailbreak attacks. While output monitors fail, we show that linear probes on internal activations can be used to reliably detect strategic dishonesty. We validate probes on datasets with verifiable outcomes and by using their features as steering vectors. Overall, we consider strategic dishonesty as a concrete example of a broader concern that alignment of LLMs is hard to control, especially when helpfulness and harmlessness conflict.

The ability to translate diverse patterns of inputs into structured patterns of behavior has been thought to rest on both humans' and machines' ability to learn robust representations of relevant concepts. The rapid advancement of transformer-based large language models (LLMs) has led to a diversity of computational ingredients -- architectures, fine tuning methods, and training datasets among others -- but it remains unclear which of these ingredients are most crucial for building models that develop human-like representations. Further, most current LLM benchmarks are not suited to measuring representational alignment between humans and models, making benchmark scores unreliable for assessing if current LLMs are making progress towards becoming useful cognitive models. We address these limitations by first evaluating a set of over 70 models that widely vary in their computational ingredients on a triplet similarity task, a method well established in the cognitive sciences for measuring human conceptual representations, using concepts from the THINGS database. Comparing human and model representations, we find that models that undergo instruction-finetuning and which have larger dimensionality of attention heads are among the most human aligned, while multimodal pretraining and parameter size have limited bearing on alignment. Correlations between alignment scores and scores on existing benchmarks reveal that while some benchmarks (e.g., MMLU) are better suited than others (e.g., MUSR) for capturing representational alignment, no existing benchmark is capable of fully accounting for the variance of alignment scores, demonstrating their insufficiency in capturing human-AI alignment. Taken together, our findings help highlight the computational ingredients most essential for advancing LLMs towards models of human conceptual representation and address a key benchmarking gap in LLM evaluation.

Large language models (LLMs) are increasingly being applied to programming tasks, ranging from single-turn code completion to autonomous agents. Current code agent designs frequently depend on complex, hand-crafted workflows and tool sets. However, this reliance on elaborate scaffolding presents several challenges: agent performance becomes overly dependent on prompt tuning and custom design choices, heavy human intervention obscures a model's true underlying capabilities, and intricate pipelines are costly to build and maintain. Furthermore, optimizing complex task prompts increases the risk of data leakage. Currently, when introducing new models, LLM providers like OpenAI and Anthropic often publish benchmark scores to demonstrate their models' coding proficiency, but keep their proprietary evaluation frameworks confidential. To address these limitations, we introduce Lita (Lite Agent), which operationalizes liteness, a principle of minimizing manual design while retaining the essential elements of a fully autonomous agent. Lita enables a more faithful and unified evaluation without elaborate scaffolding. Experiments on the Aider Polyglot and SWE-Bench with frontier models demonstrate that Lita achieves competitive or superior performance compared to workflow-based and agentic baselines. Crucially, Lita also consumes fewer tokens and requires significantly less design effort. Our results suggest that Lita is sufficient to reveal the underlying coding competence of modern LLMs. Finally, we propose the Agent Complexity Law: the performance gap between agents of varying complexity, from simple to sophisticated designs, will shrink as the core model improves, ultimately converging to a negligible difference.

The prohibitive cost of evaluating large language models (LLMs) on comprehensive benchmarks necessitates the creation of small yet representative data subsets (i.e., tiny benchmarks) that enable efficient assessment while retaining predictive fidelity. Current methods for this task operate under a model-centric paradigm, selecting benchmarking items based on the collective performance of existing models. Such approaches are limited by large upfront costs, an inability to immediately handle new benchmarks (`cold-start'), and the fragile assumption that future models will share the failure patterns of their predecessors. In this work, we challenge this paradigm and propose a item-centric approach to benchmark subset selection, arguing that selection should be based on the intrinsic properties of the task items themselves, rather than on model-specific failure patterns. We instantiate this item-centric efficient benchmarking approach via a novel method, Scales++, where data selection is based on the cognitive demands of the benchmark samples. Empirically, we show Scales++ reduces the upfront selection cost by over 18x while achieving competitive predictive fidelity. On the Open LLM Leaderboard, using just a 0.5\% data subset, we predict full benchmark scores with a 2.9% mean absolute error. We demonstrate that this item-centric approach enables more efficient model evaluation without significant fidelity degradation, while also providing better cold-start performance and more interpretable benchmarking.

For deploying foundation models, practitioners increasingly need prescriptive scaling laws: given a pre training compute budget, what downstream accuracy is attainable with contemporary post training practice, and how stable is that mapping as the field evolves? Using large scale observational evaluations with 5k observational and 2k newly sampled data on model performance, we estimate capability boundaries, high conditional quantiles of benchmark scores as a function of log pre training FLOPs, via smoothed quantile regression with a monotone, saturating sigmoid parameterization. We validate the temporal reliability by fitting on earlier model generations and evaluating on later releases. Across various tasks, the estimated boundaries are mostly stable, with the exception of math reasoning that exhibits a consistently advancing boundary over time. We then extend our approach to analyze task dependent saturation and to probe contamination related shifts on math reasoning tasks. Finally, we introduce an efficient algorithm that recovers near full data frontiers using roughly 20% of evaluation budget. Together, our work releases the Proteus 2k, the latest model performance evaluation dataset, and introduces a practical methodology for translating compute budgets into reliable performance expectations and for monitoring when capability boundaries shift across time.

Large language models (LLMs) frequently achieve impressive scores on standardized benchmarks, yet accuracy alone offers a limited view of their capabilities. Evaluating open-source LLMs through leaderboards faces persistent issues like data contamination, narrow task scope, and weak alignment with real-world reliability. Benchmark-based evaluations such as MMLU PRO, BBH, or IFEval primarily capture what a model outputs on fixed test sets, not how it processes information, calibrates uncertainty, or structures internal knowledge. In this article, we advocate for a shift from benchmark-centric evaluation toward a complementary, state-centered intrinsic assessment of LLMs. To this end, we introduce Latent Performance Profiling (LPP) -- a framework that derives task-agnostic diagnostics from hidden activations and output distributions. LPP defines a set of scalar metrics on a model's latent representations and dynamics, revealing scale-independent traits that enable interpretable comparisons and uncover hidden vulnerabilities. Unlike static accuracy scores, LPP provides stable, architecture-sensitive signatures across models of similar size. With extensive empirical analyses across eight LLMs, spanning a size range of 0.5B-14B, we demonstrate that models with similar benchmark scores can exhibit contrasting latent profiles, such as differences in entropy or adaptability. Guided by these insights, we design synthetic probes for uncertainty and symbolic reasoning that align with intrinsic metrics while decoupling from leaderboard bias. We recommend that reporting LPP alongside benchmarks provides a deeper, interpretable understanding of model behavior, enabling more reliable model selection, safety assessment, and evaluation beyond surface-level accuracy.

We introduce an axiomatic evaluation framework for latent thought representations in LLMs, comprising metrics that are independent of downstream benchmark scores and reveal representational failures that benchmark accuracy masks. Existing evaluations conflate representation quality with model capacity. Therefore, failures cannot be attributed to the representation rather than to the model that processes it. We formalize four functional axioms (Causality, Minimality, Separability, and Stability) and define a quantitative measure for each, computed directly on the representation independently of downstream accuracy. We audit open-weight LLMs across 23 reasoning tasks (e.g., Spatial Reasoning, Factual QA). We find that no candidate satisfies all four axioms simultaneously, that the representations distinguish task type reliably but cannot distinguish between two questions within the same task, and that the representations encode little information beyond what is already present in the input embedding. The failure is consistent across dense, reasoning-distilled, and RL-trained model families, indicating that the gap is structural rather than a property of model size or training procedure.

Omni modal large language models (OmniLLMs) have attracted wide attention for their ability to jointly process audio and video, but they generate large token sequences under audio-visual inputs, leading to substantial inference cost. Existing audio-visual token compression methods often rely on unimodal guidance, overlooking the temporal locality of query-relevant evidence in audio-visual inputs and implicitly assuming that the two modalities share a temporally aligned information density distribution. We propose OmniFocus, a training-free query-guided token compression method for OmniLLMs that performs independent importance estimation for video and audio, enabling a modality-symmetric compression design that preserves modality-specific salient evidence while maintaining audio-visual alignment, thereby mitigating the modality bias issue that can arise from unimodal-guided compression. Experiments on the Qwen2.5-Omni model family across four audio-visual benchmarks show that OmniFocus maintains strong compressed performance at low token retention ratios and outperforms existing baselines on several major benchmark scores at 25\% token retention. On DailyOmni with Qwen2.5-Omni-7B at 25\% token retention, OmniFocus maintains 59.40 accuracy while delivering up to 1.38times prefill speedup relative to the full-token baseline, highlighting a favorable practical accuracy-efficiency trade-off.

Large language model evaluations typically focus on performance under nominal conditions, creating an illusion of capability where models comfortably walk a narrow, highly optimized generation corridor. In real-world deployments, however, complex system prompts, safety guardrails, and structural constraints continuously force models off this nominal path, driving a divergence between benchmark scores and deployment performance. To address this issue, we introduce Decoding-Level Taboo, a zero-prompt diagnostic stress test that intervenes directly in logit space at runtime, forcing models out of their nominal paths. By dynamically masking primary candidate tokens at word boundaries, Taboo forces machine circumlocution. Evaluating Taboo across several open-weight model families reveals that off-path robustness is heavily influenced by both parameter scale and post-training instruction alignment, with robustness generally improving with model size and alignment. Beyond the results presented in this paper, Taboo provides a novel primitive for generating diverse synthetic datasets, stress-testing runtime safety guardrails, and auditing model reliability prior to real-world deployment.

We identify the Flow Moment, a reasoning pattern characterized by sustained, process-confirming verbalizations such as I'm doing, in contrast to the revision- and backtracking-oriented Aha Moment. We refer to their corresponding linguistic expressions as Flow Markers and Aha Markers, respectively. Based on this observation, we construct Flow-CoT by rewriting the discourse markers of original reasoning traces while preserving their underlying reasoning content, and use it as auxiliary supervision for on-policy self-distillation (OPSD). We further propose Aha-Flow Distillation (AFD), a dual-mode extension of OPSD that pairs different forms of privileged information with corresponding reasoning instructions. The Aha branch retains concise solution-based supervision, while the Flow branch introduces rewritten Flow-CoT under a direct and confident reasoning instruction. At inference time, the model uses only the standard reflective instruction, so Flow-style reasoning serves purely as a training signal. Experiments on AIME25 and HMMT25 show consistent improvements across Qwen3-8B and Qwen3-4B: AFD improves Avg@12 from 60.8 to 61.3 on Qwen3-8B and from 57.5 to 58.6 on Qwen3-4B over our reproduced OPSD baselines. Controlled ablations further show that, with the same Flow-CoT/Aha-CoT composition, dual-mode training improves Avg@12 from 59.5 to 60.1, indicating that the benefit comes not only from introducing heterogeneous reasoning supervision, but also from how it is organized during self-distillation. The code is available at https://github.com/Wang-Xiaodong1899/Aha-Flow-Distillation.

With the increase in deep learning, it becomes increasingly difficult to understand the model in which AI systems can identify objects. Thus, an adversary could aim to modify an image by adding unseen elements, which will confuse the AI in its recognition of an entity. This paper thus investigates the adversarial robustness of LLaVA-1.5-13B and Meta's Llama 3.2 Vision-8B-2. These are tested for untargeted PGD (Projected Gradient Descent) against the visual input modality, and empirically evaluated on the Visual Question Answering (VQA) v2 dataset subset. The results of these adversarial attacks are then quantified using the standard VQA accuracy metric. This evaluation is then compared with the accuracy degradation (accuracy drop) of LLaVA and Llama 3.2 Vision. A key finding is that Llama 3.2 Vision, despite a lower baseline accuracy in this setup, exhibited a smaller drop in performance under attack compared to LLaVA, particularly at higher perturbation levels. Overall, the findings confirm that the vision modality represents a viable attack vector for degrading the performance of contemporary open-weight VLMs, including Meta's Llama 3.2 Vision. Furthermore, they highlight that adversarial robustness does not necessarily correlate directly with standard benchmark performance and may be influenced by underlying architectural and training factors.

The rapid proliferation of large audio models (LAMs) demands efficient approaches for model comparison, yet comprehensive benchmarks are costly. To fill this gap, we investigate whether minimal subsets can reliably evaluate LAMs while reducing costs and data redundancy. Analyzing 10 subset selection methods with 18 audio models across 40 tasks covering major LAM evaluation dimensions, we show that subsets of just 50 examples (0.3% of data) can achieve over 0.93 Pearson correlation with full benchmark scores. To understand how well these scores align with what practitioners ultimately care about, user satisfaction, we collect 776 human preference ratings from realistic voice assistant conversations, finding that both subsets and full benchmark achieve only 0.85 correlation with human. To better predict preferences, we trained regression models on these selected subsets, achieving 0.98 correlation -- outperforming regression models trained on both random subsets and the full benchmark. This demonstrates that in regression modeling, well-curated subsets outpredict the full benchmark, showing quality over quantity. We open-source these regression-weighted subsets as the HUMANS benchmark, an efficient proxy for LAM evaluation that captures both benchmark performance and user preferences.

Multilingual retrieval increasingly underpins cross-lingual question answering and retrieval-augmented generation. Strong zero-shot scores on multilingual benchmarks are often taken as evidence that current encoders transfer reliably across many languages. We argue that this assumption breaks down for underrepresented, morphologically rich languages, and use Amharic as a diagnostic case. Under a shared passage retrieval protocol covering dense, late-interaction, learned sparse, and cross-encoder paradigms, we compare zero-shot multilingual retrievers, Amharic-fine-tuned multilingual retrievers, and monolingual Amharic retrievers. The strongest zero-shot multilingual retriever underperforms the strongest monolingual Amharic first-stage retriever by 23% relative MRR@10. Fine-tuning two recent multilingual embedding models on the same Amharic supervision yields 32-60% relative MRR@10 gains over zero-shot, but the best Amharic-fine-tuned multilingual model remains below the strongest monolingual Amharic retriever. These findings indicate that zero-shot multilingual retrieval is not a sufficient proxy for equitable information access in the LLM era: for underrepresented languages, retrieval must be evaluated and adapted in-language rather than inferred from aggregate multilingual benchmarks. To foster future research, we publicly release the dataset, codebase, and trained models at https://github.com/rasyosef/amharic-neural-ir.

Abundant procedural knowledge on the Web holds great potential for helping agents solve long-horizon tasks. However, such knowledge is often multimodal, heterogeneous, noisy, and implicitly assumes human executors, making it difficult to use directly as the skills required by agents. To bridge the gap between human-oriented guides and agent-executable skills, we formalize this problem as guide-to-skill learning: converting in-the-wild guides into executable skills and continuously improving them from trajectories observable to the agent. To evaluate the capability of existing agents on this task, we introduce MMG2Skill-Bench, the first benchmark designed for this problem. We further propose MMG2Skill, a closed-loop framework that compiles guides into editable skills, conditions a fixed vision-language model (VLM) agent on these skills during execution, and revises the skills from trajectory-level root-cause feedback without using benchmark scores. Across GUI control, open-ended gameplay, and strategic card play with six VLM backbones, MMG2Skill consistently outperforms vanilla baseline agents in every model-domain setting, achieving macro-average gains of +12.8 to +25.3 percentage points across backbones. Ablation studies show that directly prompting agents with raw guides can degrade performance, while both structured skill construction and trajectory-driven revision are necessary for the observed improvements. On success-inferable tasks, analyzer-based early stopping further prevents late-stage performance regressions and saves 25%-53% of attempts when the success signal is properly calibrated.

Yes. Can it cure doom loops? Probably not. The Gemma 4 instruction-tuned models share a reproducible failure: on long factual enumeration prompts, such as listing every episode of a TV series, the 88 IAU constellations, or the 151 original Pokemon, they collapse into repetition, either a tight verbatim loop or a list whose entries decay onto a single answer. These loops occur at rates as high as 95% and survive prompt rewording, inference-engine changes, and most sampling adjustments. In this paper we explore whether this behavior is localized enough to remove by weight edits. To localize the cause, we use per-layer ablation and per-neuron attribution, then confirm the strongest candidates with full-generation sweeps. The loops trace to a small set of MLP neurons (or, in the 26B-A4B Mixture-of-Experts model, a few routed experts) which we suppress with static weight edits. These "surgeries" can be as small as a single sign-inverted neuron (in the E2B model). The size of the effective edits grows with model scale, but in all cases, the loop patterns can be addressed at normal generation budgets while preserving general-purpose benchmark scores. However, the edits do not solve everything: we also study longer thinking budgets, where the two larger models most visibly enter doom looping, i.e. a non-convergent regime in which the model self-corrects in circles over a fact it cannot recall, exhausting the budget without committing to a final answer. We show this residual failure is reduced but not eliminated by the same edits, and argue it is fundamentally a knowledge-precision problem rather than a removable circuit; weight surgery can delete a loop, but it cannot supply a missing fact. Our results are both a feasibility demonstration, that is, evidence that a concrete generation pathology can be localized to a few parameters and edited out, and a delineation of where that approach stops.

Cantonese is widely spoken but remains low-resource in written data, with no large corpus of native Cantonese reasoning traces available for model training. We develop and release CantoneseLLM v2, comprising models based on Qwen3 8B and 30B-A3B. The models are trained through CPT on 784 million Cantonese and Hong Kong-related tokens, chat-vector merging, SFT, DPO, and RLVR. Evaluation across the training stages shows that chat-vector merging transfers instruction following but preserves the donor model's reasoning language, while SFT with limited Cantonese reasoning data substantially shortens or removes reasoning traces and reduces benchmark performance. DPO restores the reasoning-block format, particularly for the 8B model, but recovers only part of the lost performance. The RLVR training with Cantonese language and Traditional Chinese scripts as multiplicative constraints introduced Cantonese language alignment and restored the lost performance. The 30B-A3B model reaches 73.16 on HKCanto-Eval, within 1.20 points of its merged checkpoint, while retaining the Cantonese reasoning behaviour absent from that checkpoint. We release the model checkpoints, the training environments, and a thirteen-year Traditional Chinese Common Crawl dataset. The models can be accessed at https://huggingface.co/collections/hon9kon9ize/cantonesellm-v20

Graphical user interface (GUI) agents build on vision-language models to complete user tasks end-to-end in real applications through interface actions such as tapping, swiping, text entry, and navigation. However, existing GUI agents are trained and evaluated largely on offline trajectories, simulated environments, and standardized benchmarks. These differ substantially from real applications in interface layout, interaction logic, and abnormal-state distribution, and cannot faithfully characterize execution stability in real-world use, where account states, permission dialogs, payment authentication, and risk control continually reshape the state distribution and open a persistent gap between benchmark scores and real usability. To close this gap, we propose Xiaomi-GUI-0, a native multimodal GUI agent for real mobile environments, trained and evaluated within a real-device closed loop. At its core is a real-device-dominant hybrid infrastructure, where physical devices are the primary execution environment and sandboxes provide auxiliary support, so that data collection, training, rollout, and evaluation share an execution distribution close to real deployment. We construct multi-source training data spanning high-frequency head tasks, high-generalization data for long-tail intents, and capability-enhancement data for reflection and memory, and introduce an error-driven data flywheel that turns failure trajectories into corrected actions, reflective explanations, and recovery demonstrations. The model is trained through a progressive three-stage pipeline of supervised fine-tuning, step-level reinforcement learning, and agentic reinforcement learning. Evaluated on public benchmarks and our in-house RealMobile, Xiaomi-GUI-0 achieves 72.0% success on RealMobile and 78.9% on AndroidWorld, while substantially improving execution stability and abnormal-state recognition in real-world tasks.

Physical AI models are evaluated on suites of benchmarks that differ across model reports, leaving the model-by-benchmark matrix sparse and the relationship between benchmarks unmeasured. We construct a matrix of 51 models on 12 physical AI benchmarks, selected from a registry of 51 benchmarks and 152 models by reporting density, combining scores from model cards and benchmark papers with our own evaluation runs under each benchmark's official protocol. We measure how much information the benchmarks share and show quantitative evidence of Redundancy. Redundancy affects reported rankings: collapsing the two substitute pairs into single columns moves 22 of 51 models by three or more places under an equally weighted average. We then select benchmarks greedily under a utility combining score dispersion with variance not explained by the already-selected set, and obtain a four-benchmark subset retaining 78.5\% of the utility of all 12, on which we fit a Bradley--Terry ranking. The procedure requires only benchmark-level scores with sufficient overlap and is not specific to physical AI.

Unlearning benchmarks such as TOFU and MUSE certify forgetting by reading the model's final answer, where a model that refuses to answer already counts as having forgotten. We show that this model-level certificate does not transfer once the model is deployed as an agent. We introduce K-Bench, a benchmark that scores LLM unlearning under agentic deployment. K-Bench inspects all six channels a ReAct agent exposes, including its chain-of-thought (CoT), tool calls and tool observations, and elicited summary. A query counts as leaked if the secret appears in any of them. Each experiment places the secret in exactly one of the agent's three sources (the weights, the prompt, or the retrieval store). The K-Score is computed separately for each source and credits forgetting only when the agent remains usable. Clearing the answer channel does not make the secret unrecoverable. On structured retrieval, the secret stays verbatim in the tool-observation channel and the aggregate leak rate is unchanged. When the secret lives in the prompt or the retrieval store, TOFU and MUSE report no leakage, while the deployed agent still leaks it on 22--86\% of queries. When the secret is in the weights, none of the twenty evaluated published methods demonstrably removes it, and only an input-corruption intervention reaches selective forgetting under the evaluated observer. The top-ranked method changes across base models. A refusal-tuning method resists the evaluated extraction without verified knowledge removal.

The choice of embedding model is a crucial step in the design of Retrieval Augmented Generation (RAG) systems. Given the sheer volume of available options, identifying clusters of similar models streamlines this model selection process. Relying solely on benchmark performance scores only allows for a weak assessment of model similarity. Thus, in this study, we evaluate the similarity of embedding models within the context of RAG systems. Our assessment is two-fold: We use Centered Kernel Alignment to compare embeddings on a pair-wise level. Additionally, as it is especially pertinent to RAG systems, we evaluate the similarity of retrieval results between these models using Jaccard and rank similarity. We compare different families of embedding models, including proprietary ones, across five datasets from the popular Benchmark Information Retrieval (BEIR). Through our experiments we identify clusters of models corresponding to model families, but interestingly, also some inter-family clusters. Furthermore, our analysis of top-k retrieval similarity reveals high-variance at low k values. We also identify possible open-source alternatives to proprietary models, with Mistral exhibiting the highest similarity to OpenAI models.

Many peer-review venues are using algorithms to assign submissions to reviewers. The crux of such automated approaches is the notion of the "similarity score" -- a numerical estimate of the expertise of a reviewer in reviewing a paper -- and many algorithms have been proposed to compute these scores. However, these algorithms have not been subjected to a principled comparison, making it difficult for stakeholders to choose the algorithm in an evidence-based manner. The key challenge in comparing existing algorithms and developing better algorithms is the lack of publicly available gold-standard data. We address this challenge by collecting a novel dataset of similarity scores that we release to the research community. Our dataset consists of 477 self-reported expertise scores provided by 58 researchers who evaluated their expertise in reviewing papers they have read previously. Using our dataset, we compare several widely used similarity algorithms and offer key insights. First, all algorithms exhibit significant error, with misranking rates between 12%-30% in easier cases and 36%-43% in harder ones. Second, most specialized algorithms are designed to work with titles and abstracts of papers, and in this regime the SPECTER2 algorithm performs best. Interestingly, classical TF-IDF matches SPECTER2 in accuracy when given access to full submission texts. In contrast, off-the-shelf LLMs lag behind specialized approaches.

Effectively analyzing the comments to uncover latent intentions holds immense value in making strategic decisions across various domains. However, several challenges hinder the process of sentiment analysis including the lexical diversity exhibited in comments, the presence of long dependencies within the text, encountering unknown symbols and words, and dealing with imbalanced datasets. Moreover, existing sentiment analysis tasks mostly leveraged sequential models to encode the long dependent texts and it requires longer execution time as it processes the text sequentially. In contrast, the Transformer requires less execution time due to its parallel processing nature. In this work, we introduce a novel hybrid deep learning model, RoBERTa-BiLSTM, which combines the Robustly Optimized BERT Pretraining Approach (RoBERTa) with Bidirectional Long Short-Term Memory (BiLSTM) networks. RoBERTa is utilized to generate meaningful word embedding vectors, while BiLSTM effectively captures the contextual semantics of long-dependent texts. The RoBERTa-BiLSTM hybrid model leverages the strengths of both sequential and Transformer models to enhance performance in sentiment analysis. We conducted experiments using datasets from IMDb, Twitter US Airline, and Sentiment140 to evaluate the proposed model against existing state-of-the-art methods. Our experimental findings demonstrate that the RoBERTa-BiLSTM model surpasses baseline models (e.g., BERT, RoBERTa-base, RoBERTa-GRU, and RoBERTa-LSTM), achieving accuracies of 80.74%, 92.36%, and 82.25% on the Twitter US Airline, IMDb, and Sentiment140 datasets, respectively. Additionally, the model achieves F1-scores of 80.73%, 92.35%, and 82.25% on the same datasets, respectively.

Anomaly detection (AD) is a machine learning task that identifies anomalies by learning patterns from normal training data. In many real-world scenarios, anomalies vary in severity, from minor anomalies with little risk to severe abnormalities requiring immediate attention. However, existing models primarily operate in a binary setting, and the anomaly scores they produce are usually based on the deviation of data points from normal data, which may not accurately reflect practical severity. In this paper, we address this gap by making three key contributions. First, we propose a novel setting, Multilevel AD (MAD), in which the anomaly score represents the severity of anomalies in real-world applications, and we highlight its diverse applications across various domains. Second, we introduce a novel benchmark, MAD-Bench, that evaluates models not only on their ability to detect anomalies, but also on how effectively their anomaly scores reflect severity. This benchmark incorporates multiple types of baselines and real-world applications involving severity. Finally, we conduct a comprehensive performance analysis on MAD-Bench. We evaluate models on their ability to assign severity-aligned scores, investigate the correspondence between their performance on binary and multilevel detection, and study their robustness. This analysis offers key insights into improving AD models for practical severity alignment. The code framework and datasets used for the benchmark will be made publicly available.

Understanding complete musical scores entails integrated reasoning over pitch, rhythm, harmony, and large-scale structure, yet the ability of Large Language Models and Vision--Language Models to interpret full musical notation remains insufficiently examined. We introduce Musical Score Understanding Benchmark (MSU-Bench), a human-curated benchmark for score-level musical understanding across textual (ABC notation) and visual (PDF) modalities. MSU-Bench contains 1,800 generative question-answer pairs from works by Bach, Beethoven, Chopin, Debussy, and others, organised into four levels of increasing difficulty, ranging from onset information to texture and form. Evaluations of more than fifteen state-of-the-art models, in both zero-shot and fine-tuned settings, reveal pronounced modality gaps, unstable level-wise performance, and challenges in maintaining multilevel correctness. Fine-tuning substantially improves results across modalities while preserving general knowledge, positioning MSU-Bench as a robust foundation for future research in multimodal reasoning. The benchmark and code are available at https://github.com/Congren-Dai/MSU-Bench.

As large language models (LLMs) advance across diverse tasks, the need for comprehensive evaluation beyond single metrics becomes increasingly important. To fully assess LLM intelligence, it is crucial to examine their interactive dynamics and strategic behaviors. We present LLMsPark, a game theory-based evaluation platform that measures LLMs' decision-making strategies and social behaviors in classic game-theoretic settings, providing a multi-agent environment to explore strategic depth. Our system cross-evaluates 15 leading LLMs (both commercial and open-source) using leaderboard rankings and scoring mechanisms. Higher scores reflect stronger reasoning and strategic capabilities, revealing distinct behavioral patterns and performance differences across models. This work introduces a novel perspective for evaluating LLMs' strategic intelligence, enriching existing benchmarks and broadening their assessment in interactive, game-theoretic scenarios. The benchmark and rankings are publicly available at https://llmsparks.github.io/.

Video-quality measurement is a critical task in video processing. Nowadays, many implementations of new encoding standards - such as AV1, VVC, and LCEVC - use deep-learning-based decoding algorithms with perceptual metrics that serve as optimization objectives. But investigations of the performance of modern video- and image-quality metrics commonly employ videos compressed using older standards, such as AVC. In this paper, we present a new benchmark for video-quality metrics that evaluates video compression. It is based on a new dataset consisting of about 2,500 streams encoded using different standards, including AVC, HEVC, AV1, VP9, and VVC. Subjective scores were collected using crowdsourced pairwise comparisons. The list of evaluated metrics includes recent ones based on machine learning and neural networks. The results demonstrate that new no-reference metrics exhibit a high correlation with subjective quality and approach the capability of top full-reference metrics.

PCB routing is the task of connecting the nets of a board with copper traces under strict design rules, yet learning-based methods still lag behind rule-based routers. We introduce PCBWorld, an open-source engine-grounded PCB routing environment built on KiCad, an electronic design automation (EDA) engine. As a human engineer does, agents in PCBWorld interactively route a board through the engine's native operations, guided by its Design Rule Check (DRC) feedback. The environment supports both RL and tool-using LLM agents. Alongside the environment, PCBWorld-Bench provides three board datasets in the native .kicad_pcb format, two controllable synthetic generators and 679 real open-source boards. It scores any completed board with eight engine-checked evaluation metrics, regardless of the routing method. In our experiments, agents in PCBWorld consistently outperformed grid-action RL policies and open-loop LLM baselines, and an RL policy trained only on synthetic boards transferred zero-shot to real boards, approaching rule-based routers. PCBWorld and PCBWorld-Bench are available at https://github.com/LGAI-Research/PCBWorld.

In this paper, we present the LingOly benchmark, a novel benchmark for advanced reasoning abilities in large language models. Using challenging Linguistic Olympiad puzzles, we evaluate (i) capabilities for in-context identification and generalisation of linguistic patterns in very low-resource or extinct languages, and (ii) abilities to follow complex task instructions. The LingOly benchmark covers more than 90 mostly low-resource languages, minimising issues of data contamination, and contains 1,133 problems across 6 formats and 5 levels of human difficulty. We assess performance with both direct accuracy and comparison to a no-context baseline to penalise memorisation. Scores from 11 state-of-the-art LLMs demonstrate the benchmark to be challenging, and models perform poorly on the higher difficulty problems. On harder problems, even the top model only achieved 35.3% accuracy, 21.7% improvement over the no-context baseline. Large closed models typically outperform open models, and in general, the higher resource the language, the better the scores. These results indicate, in absence of memorisation, true multi-step out-of-domain reasoning remains a challenge for current language models.

Risk scores are simple classification models that let users make quick risk predictions by adding and subtracting a few small numbers. These models are widely used in medicine and criminal justice, but are difficult to learn from data because they need to be calibrated, sparse, use small integer coefficients, and obey application-specific operational constraints. In this paper, we present a new machine learning approach to learn risk scores. We formulate the risk score problem as a mixed integer nonlinear program, and present a cutting plane algorithm for non-convex settings to efficiently recover its optimal solution. We improve our algorithm with specialized techniques to generate feasible solutions, narrow the optimality gap, and reduce data-related computation. Our approach can fit risk scores in a way that scales linearly in the number of samples, provides a certificate of optimality, and obeys real-world constraints without parameter tuning or post-processing. We benchmark the performance benefits of this approach through an extensive set of numerical experiments, comparing to risk scores built using heuristic approaches. We also discuss its practical benefits through a real-world application where we build a customized risk score for ICU seizure prediction in collaboration with the Massachusetts General Hospital.

Many deployments must compare candidate language models for safety before a labeled benchmark exists for the relevant language, sector, or regulatory regime. We formalize this setting as benchmarkless comparative safety scoring and specify the contract under which a scenario-based audit can be interpreted as deployment evidence. Scores are valid only under a fixed scenario pack, rubric, auditor, judge, sampling configuration, and rerun budget. Because no labels are available, we replace ground-truth agreement with an instrumental-validity chain: responsiveness to a controlled safe-versus-abliterated contrast, dominance of target-driven variance over auditor and judge artifacts, and stability across reruns. We instantiate the chain in SimpleAudit, a local-first scoring instrument, and validate it on a Norwegian safety pack. Safe and abliterated targets separate with AUROC values between 0.89 and 1.00, target identity is the dominant variance component (η^2 approx 0.52), and severity profiles stabilize by ten reruns. Applying the same chain to Petri shows that it admits both tools. The substantial differences arise upstream of the chain, in claim-contract enforcement and deployment fit. A Norwegian public-sector procurement case comparing Borealis and Gemma 3 demonstrates the resulting evidence in practice: the safer model depends on scenario category and risk measure. Consequently, scores, matched deltas, critical rates, uncertainty, and the auditor and judge used must be reported together rather than collapsed into a single ranking.

Progress in video generation keeps narrowing the visual gap between AI-generated and professionally produced footage, yet most benchmarks still draw prompts from web sources or LLM templates and score them with untrained, generic multimodal models. More fundamentally, their evaluation taxonomies remain rudimentary (overall visual quality, coarse text alignment and temporal smoothness) rather than the professional Cinematic Language criteria by which films are actually made and judged, so they assess basic video plausibility rather than film-grade craft. We introduce FilmBench, a text-to-video (T2V) and reference-to-video (R2V) benchmark grounded in the professional Cinematic Language of the film- academy tradition and co-developed with directors and faculty from the Beijing Film Academy and the Hujing Digital Media & Entertainment Group film studio. It rests on three choices. First, prompts are reverse-engineered from clips of award-winning films spanning 20 cinematic genres and chosen by professional directors, so every prompt is anchored to a verified live-action reference; the prompts follow real shot lists, and most script multiple shots (1,056 of the 1,169 prompts are multi-shot), unlike prior single-clip benchmarks. Second, evaluation follows a three-level Cinematic taxonomy of 3 axes, 12 components and 35 (T2V) +3 (R2V-only) sub-metrics. Third, we develop an in-house expert-grade automatic evaluation agent and open-source its core suite of Cinematic Language operators (FilmOps). Benchmarking leading video generation models (9 for T2V, 7 for R2V), the evaluator reproduces the human model ranking at model-level Spearman ho = 0.95 (T2V) and 0.96 (R2V). Scores fall well below prior web-style benchmarks, with two consistent gaps in dynamic aesthetics and a marked single- to multi-shot performance drop that widens for weaker models.

SciCode is the standard measure of the scientific-coding ability of language models: research-level problems that demand both frontier scientific theory and its implementation as working numerical code. It is a component of the Artificial Analysis Intelligence Index and a standing evaluation in government and national-laboratory suites. Yet its scores have recently plateaued: the strongest 2026 models cluster tightly around 60\% subproblem accuracy, and a successor model ties its predecessor. We trace this stagnation to defects in the benchmark itself. A per-problem, domain-expert audit of all 65 test problems uncovers 263 defects; 192 of them, spread across 91\% of the main problems, cause correct, instruction-following solutions to be wrongly rejected---through non-reproducible gold answers, over-tight tolerances, or self-contradictory specifications. Critically, 78\% of these score-suppressing defects require specialized physics or mathematics knowledge to detect, not mere clerical proofreading. We corrected every confirmable defect to produce SciCode-Verified. The corrections add only the specifications a well-posed problem requires, repair grading, and tighten the tests that were too lenient; every change is recorded with its justification and independently re-checked by a second domain expert. We re-evaluate twelve frontier model snapshots on the corrected benchmark and find a substantial recovery: subproblem accuracy rises from 45--60\% to 84--98\%, and main-problem accuracy from 9--27\% to 69--92\%. State-of-the-art models are far more proficient in scientific coding than SciCode has suggested---the bottleneck was not model capability, but the quality of the evaluation instrument. We release SciCode-Verified with its complete audit trail as the corrected public standard.

Comprehensively evaluating AI agents across interactive environments is difficult due to fragmented tasks, scaffolds, verifiers, and scoring rules. Unfortunately, existing efforts to unify these evaluations are limited in scale and domain, making costly reruns necessary and leaving available data incomparable. We introduce MESSIER, a unified corpus of 957,611 records spanning 30 benchmarks, 745 agents, 11,891 tasks, and 74,263 verifiers. MESSIER combines public evaluation results with new runs on six underrepresented professional and scientific benchmarks, standardizing their heterogeneous components into a common schema. Using this corpus, we show that frontier progress is uneven across benchmark groups, with function-calling evaluations largely saturated, programming improving fastest, and enterprise workflows remaining most challenging. Counterfactual rescoring further shows that strict all-pass scoring in multi-verifier tasks can alter agent rankings. Finally, we derive capability scores from our corpus that correlate with Epoch's Evaluation Capability Index rankings at Spearman ho = 0.84. The scores can also be estimated for subsets defined by domain, occupation, action space, or verifier type. In essence, MESSIER is a reusable resource for studying agent performance at scale, and a basis for designing better evaluations.

Multi-modal Large Language Models (MLLMs) exhibit impressive problem-solving abilities in various domains, but their visual comprehension and abstract reasoning skills remain under-evaluated. To this end, we present PolyMATH, a challenging benchmark aimed at evaluating the general cognitive reasoning abilities of MLLMs. PolyMATH comprises 5,000 manually collected high-quality images of cognitive textual and visual challenges across 10 distinct categories, including pattern recognition, spatial reasoning, and relative reasoning. We conducted a comprehensive, and quantitative evaluation of 15 MLLMs using four diverse prompting strategies, including Chain-of-Thought and Step-Back. The best scores achieved on PolyMATH are ~41%, ~36%, and ~27%, obtained by Claude-3.5 Sonnet, GPT-4o and Gemini-1.5 Pro respectively - highlighting the logical and visual complexity of these questions. A further fine-grained error analysis reveals that these models struggle to understand spatial relations and perform drawn-out, high-level reasoning. This is further strengthened by our ablation study estimating MLLM performance when given textual descriptions in place of diagrams. As evidenced by ~4% improvement over textual descriptions as opposed to actual images, we discover that models do not truly comprehend visual diagrams and the spatial information therein, and are thus prone to logical errors. Finally, we evaluate the OpenAI o1 models and find that their performance only matches the human baseline, highlighting the difficulty of the benchmark. The results on PolyMATH highlight the room for improvement in multi-modal reasoning and provide unique insights to guide the development of future MLLMs.

Forecasts of future events are essential inputs into informed decision-making. Machine learning (ML) systems have the potential to deliver forecasts at scale, but there is no framework for evaluating the accuracy of ML systems on a standardized set of forecasting questions. To address this gap, we introduce ForecastBench: a dynamic benchmark that evaluates the accuracy of ML systems on an automatically generated and regularly updated set of 1,000 forecasting questions. To avoid any possibility of data leakage, ForecastBench is comprised solely of questions about future events that have no known answer at the time of submission. We quantify the capabilities of current ML systems by collecting forecasts from expert (human) forecasters, the general public, and LLMs on a random subset of questions from the benchmark (N=200). While LLMs have achieved super-human performance on many benchmarks, they perform less well here: expert forecasters outperform the top-performing LLM (p-value <0.001). We display system and human scores in a public leaderboard at www.forecastbench.org.

Code-switching (CS) poses a significant challenge for Large Language Models (LLMs), yet its comprehensibility remains underexplored in LLMs. We introduce CS-Sum, to evaluate the comprehensibility of CS by the LLMs through CS dialogue to English summarization. CS-Sum is the first benchmark for CS dialogue summarization across Mandarin-English (EN-ZH), Tamil-English (EN-TA), and Malay-English (EN-MS), with 900-1300 human-annotated dialogues per language pair. Evaluating ten LLMs, including open and closed-source models, we analyze performance across few-shot, translate-summarize, and fine-tuning (LoRA, QLoRA on synthetic data) approaches. Our findings show that though the scores on automated metrics are high, LLMs make subtle mistakes that alter the complete meaning of the dialogue. To this end, we introduce 3 most common type of errors that LLMs make when handling CS input. Error rates vary across CS pairs and LLMs, with some LLMs showing more frequent errors on certain language pairs, underscoring the need for specialized training on code-switched data.

Measuring the full abilities of large language models (LLMs) requires benchmarks representing multiple tasks. We aim to create large, high-quality datasets for comparison of logical reasoning skills across several languages and of suitable difficulty for LLMs of various reasoning ability. We explore multiple ways of increasing difficulty. We generate zebra puzzles in multiple languages, themes, sizes and including 14 different clue types and 8 red herring types (uninformative clues). We find puzzle sizes 2x3 and 4x5 are sufficiently challenging for GPT-4o mini (a non-reasoning model) and o3-mini (a reasoning model), respectively. Including 5 red herrings decreases o3-mini puzzle-level accuracy on 4x5 puzzles by 15pm7 %. Scores of o3-mini on 4x5 puzzles are not significantly affected by use of English vs. Danish or the common houses theme vs. the country-specific smoerrebroed theme. We find no correlation between difficulty and the selected clue types. Datasets of 128+1024 puzzles are published as MultiZebraLogic in each of nine Germanic languages for sizes 2x3 and 4x5. We publish code for puzzle generation, designed for adaptablity into more languages and themes.

Multimodal large language models (MLLMs) are now routinely deployed for visual understanding, generation, and curation. A substantial fraction of these applications require an explicit aesthetic judgment. Most existing solutions reduce this judgment to predicting a scalar score for a single image. We first ask whether such scores faithfully capture comparative preference: in a controlled study with eight expert annotators, score-derived rankings align poorly with the same annotators' direct comparisons, while direct ranking yields substantially higher inter-annotator agreement on best- and worst-image labels. Motivated by this finding, we introduce the Visual Aesthetic Benchmark (VAB), which casts aesthetic evaluation as comparative selection over candidate sets with matched subject matter. VAB contains 400 tasks and 1,195 images across fine art, photography, and illustration, with labels derived from the consensus of 10 independent expert judges per task. Evaluating 20 frontier MLLMs and six dedicated visual-quality reward models, we find that the strongest system identifies both the best and the worst image correctly across three random permutations of the candidate order in only 26.5% of tasks, far below the 68.9% achieved by human experts. Fine-tuning a 35B-parameter model on 2,000 expert examples brings its accuracy close to that of a 397B-parameter open-weight model, suggesting that the comparative signal in VAB is transferable. Together, these results expose a clear and measurable gap between current multimodal models and expert aesthetic judgment, and VAB provides the first set-based, expert-grounded testbed on which that gap can be tracked and closed.

Vision-language model (VLM) agents are increasingly deployed in interactive game environments. Yet game benchmarks for VLM agents typically report a single first-attempt score per (agent, game) pair, focus on single-agent Solo play, and lack unified protocols for evaluating heterogeneous agent classes (commercial VLMs, open-weight VLMs, and specialized game policies) on the same footing. We address these gaps with OmniGameArena, a real-time benchmark of twelve newly built Unreal Engine 5 games spanning Solo (7), PvP (3), and Coop (2) with unified action interfaces, and the Improvement Dynamics Curve (IDC), an agentic-reflection harness in which a tool-using reflector LLM autonomously refines a bounded skill prompt across multiple rounds. Beyond cold-start leaderboard scores, IDC exposes two additional observables for each (agent, game) pair: how the score evolves across reflection rounds, and how the learned skill behaves on held-out task variants. We report these observables for twelve VLM agents on the cold-start leaderboard and four top agents under IDC.

Real repository issues routinely include visual evidence such as screenshots, error dialogs, rendered UI states, and logs, yet repository-level issue localization is evaluated mostly as a text-only task. Existing multimodal SE benchmarks evaluate end-to-end repair, entangling localization with patch synthesis and obscuring whether visual input helped, hurt, or was ignored. We introduce MM-IssueLoc, a controlled benchmark and evaluation protocol for repository-level localization with visual evidence. MM-IssueLoc contains 652 issue-PR instances across 23 languages, with annotations for 7 image categories and 4 relevance levels. It provides file-level and function-level gold labels, paired text-only and with-image evaluation, and VCE-based diagnostics that convert images into structured textual evidence. We evaluate LLM-based and retrieval-based systems, including MM-IssueLoc-VL-Emb as a controlled multimodal retriever. Results show that existing systems remain far from reliable multimodal repository localization: the strongest agent reaches 38.96 file Acc@5 and 22.45 function Acc@10, while the strongest retriever reaches 33.86 function Acc@10. Cross-benchmark comparisons show that high localization scores on text-dominant SWE benchmarks do not transfer cleanly to multimodal issue localization. MM-IssueLoc turns visual evidence into an explicit evaluation variable, enabling future work to test whether systems improve by using visual evidence for localization, rather than by relying on text-only cues or downstream patch-generation effects.

Stereotype benchmark datasets are crucial to detect and mitigate social stereotypes about groups of people in NLP models. However, existing datasets are limited in size and coverage, and are largely restricted to stereotypes prevalent in the Western society. This is especially problematic as language technologies gain hold across the globe. To address this gap, we present SeeGULL, a broad-coverage stereotype dataset, built by utilizing generative capabilities of large language models such as PaLM, and GPT-3, and leveraging a globally diverse rater pool to validate the prevalence of those stereotypes in society. SeeGULL is in English, and contains stereotypes about identity groups spanning 178 countries across 8 different geo-political regions across 6 continents, as well as state-level identities within the US and India. We also include fine-grained offensiveness scores for different stereotypes and demonstrate their global disparities. Furthermore, we include comparative annotations about the same groups by annotators living in the region vs. those that are based in North America, and demonstrate that within-region stereotypes about groups differ from those prevalent in North America. CONTENT WARNING: This paper contains stereotype examples that may be offensive.

WeatherBench 2 is an update to the global, medium-range (1-14 day) weather forecasting benchmark proposed by Rasp et al. (2020), designed with the aim to accelerate progress in data-driven weather modeling. WeatherBench 2 consists of an open-source evaluation framework, publicly available training, ground truth and baseline data as well as a continuously updated website with the latest metrics and state-of-the-art models: https://sites.research.google/weatherbench. This paper describes the design principles of the evaluation framework and presents results for current state-of-the-art physical and data-driven weather models. The metrics are based on established practices for evaluating weather forecasts at leading operational weather centers. We define a set of headline scores to provide an overview of model performance. In addition, we also discuss caveats in the current evaluation setup and challenges for the future of data-driven weather forecasting.

We present MultiLoKo, a new benchmark for evaluating multilinguality in LLMs covering 31 languages. MultiLoKo consists of three partitions: a main partition consisting of 500 questions per language, separately sourced to be locally relevant to the specific language, and two translated partitions, containing human-authored translations from 30 non-English languages to English and vice versa. For comparison, we also release corresponding machine-authored translations. The data is equally distributed over two splits: a dev split and a blind, out-of-distribution test split. MultiLoKo can be used to study a variety of questions regarding the multilinguality of LLMs as well as meta-questions about multilingual benchmark creation. We compute MultiLoKo scores for 11 base and chat models marketed to be multilingual and study their average performance, their performance parity across languages, how much their ability to answer questions depends on the question language, and which languages are most difficult. None of the models we studied performs well on MultiLoKo, as indicated by low average scores as well as large differences between the best and worst scoring languages. Furthermore, we find a substantial effect of the question language, indicating sub-optimal knowledge transfer between languages. Lastly, we find that using local vs English-translated data can result in differences more than 20 points for the best performing models, drastically change the estimated difficulty of some languages. For using machines instead of human translations, we find a weaker effect on ordering of language difficulty, a larger difference in model rankings, and a substantial drop in estimated performance for all models.

Large language models (LLMs) typically generate identical or similar responses for all users given the same prompt, posing serious safety risks in high-stakes applications where user vulnerabilities differ widely. Existing safety evaluations primarily rely on context-independent metrics - such as factuality, bias, or toxicity - overlooking the fact that the same response may carry divergent risks depending on the user's background or condition. We introduce personalized safety to fill this gap and present PENGUIN - a benchmark comprising 14,000 scenarios across seven sensitive domains with both context-rich and context-free variants. Evaluating six leading LLMs, we demonstrate that personalized user information significantly improves safety scores by 43.2%, confirming the effectiveness of personalization in safety alignment. However, not all context attributes contribute equally to safety enhancement. To address this, we develop RAISE - a training-free, two-stage agent framework that strategically acquires user-specific background. RAISE improves safety scores by up to 31.6% over six vanilla LLMs, while maintaining a low interaction cost of just 2.7 user queries on average. Our findings highlight the importance of selective information gathering in safety-critical domains and offer a practical solution for personalizing LLM responses without model retraining. This work establishes a foundation for safety research that adapts to individual user contexts rather than assuming a universal harm standard.

In recent years, large language models (LLMs) have showcased significant advancements in code generation. However, most evaluation benchmarks are primarily oriented towards Python, making it difficult to evaluate other programming languages, such as Swift, with high quality. By examining widely established multilingual benchmarks like HumanEval-XL and MultiPL-E, we identified critical issues specific to their Swift components, making them insufficient or even irrelevant for assessing LLM coding capabilities on Swift. Unlike these existing approaches, which prioritize rapid scaling and generalization by automatically translating Python-centric benchmarks with LLMs, we adopt a quality-over-quantity methodology. We present SwiftEval, the first Swift-oriented benchmark consisting of 28 carefully hand-crafted problems, and evaluate 44 popular Code LLMs on it. Our results show significant LLM scores drop for problems requiring language-specific features, most noticeable in the models of smaller sizes.