The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
id: string
x: double
y: double
occluded: bool
visibility: double
vis: double
conf: double
scores_position_only: struct<queries: int64, answered: int64, visible_queries: int64, d_1: double, d_2: double, d_4: doubl (... 68 chars omitted)
child 0, queries: int64
child 1, answered: int64
child 2, visible_queries: int64
child 3, d_1: double
child 4, d_2: double
child 5, d_4: double
child 6, d_8: double
child 7, d_16: double
child 8, d_avg: double
child 9, aj: double
child 10, oa: double
queries: int64
by_category: struct<animals_nature: struct<queries: int64, answered: int64, visible_queries: int64, d_1: double, (... 1229 chars omitted)
child 0, animals_nature: struct<queries: int64, answered: int64, visible_queries: int64, d_1: double, d_2: double, d_4: doubl (... 68 chars omitted)
child 0, queries: int64
child 1, answered: int64
child 2, visible_queries: int64
child 3, d_1: double
child 4, d_2: double
child 5, d_4: double
child 6, d_8: double
child 7, d_16: double
child 8, d_avg: double
child 9, aj: double
child 10, oa: double
child 1, biomedical_scenes: struct<queries: int64, answered: int64, visible_queries: int64, d_1: double, d_2: double, d_4: doubl (... 68 chars omitted)
child 0, queries: int64
child 1, answered: int64
child 2, visible_queries: int64
child 3, d_1: double
child 4, d_2: double
child 5, d_4: double
child 6, d_8: double
ch
...
5, d_4: double
child 6, d_8: double
child 7, d_16: double
child 8, d_avg: double
child 9, aj: double
child 10, oa: double
child 13, vl_surgpt: struct<queries: int64, answered: int64, visible_queries: int64, d_1: double, d_2: double, d_4: doubl (... 68 chars omitted)
child 0, queries: int64
child 1, answered: int64
child 2, visible_queries: int64
child 3, d_1: double
child 4, d_2: double
child 5, d_4: double
child 6, d_8: double
child 7, d_16: double
child 8, d_avg: double
child 9, aj: double
child 10, oa: double
child 14, votsp2026: struct<queries: int64, answered: int64, visible_queries: int64, d_1: double, d_2: double, d_4: doubl (... 68 chars omitted)
child 0, queries: int64
child 1, answered: int64
child 2, visible_queries: int64
child 3, d_1: double
child 4, d_2: double
child 5, d_4: double
child 6, d_8: double
child 7, d_16: double
child 8, d_avg: double
child 9, aj: double
child 10, oa: double
scores: struct<queries: int64, answered: int64, visible_queries: int64, d_1: double, d_2: double, d_4: doubl (... 68 chars omitted)
child 0, queries: int64
child 1, answered: int64
child 2, visible_queries: int64
child 3, d_1: double
child 4, d_2: double
child 5, d_4: double
child 6, d_8: double
child 7, d_16: double
child 8, d_avg: double
child 9, aj: double
child 10, oa: double
answered: int64
to
{'queries': Value('int64'), 'answered': Value('int64'), 'scores': {'queries': Value('int64'), 'answered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('float64'), 'd_8': Value('float64'), 'd_16': Value('float64'), 'd_avg': Value('float64'), 'aj': Value('float64'), 'oa': Value('float64')}, 'scores_position_only': {'queries': Value('int64'), 'answered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('float64'), 'd_8': Value('float64'), 'd_16': Value('float64'), 'd_avg': Value('float64'), 'aj': Value('float64'), 'oa': Value('float64')}, 'by_category': {'animals_nature': {'queries': Value('int64'), 'answered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('float64'), 'd_8': Value('float64'), 'd_16': Value('float64'), 'd_avg': Value('float64'), 'aj': Value('float64'), 'oa': Value('float64')}, 'biomedical_scenes': {'queries': Value('int64'), 'answered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('float64'), 'd_8': Value('float64'), 'd_16': Value('float64'), 'd_avg': Value('float64'), 'aj': Value('float64'), 'oa': Value('float64')}, 'driving_street_scenes': {'queries': Value('int64'), 'answered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('flo
...
swered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('float64'), 'd_8': Value('float64'), 'd_16': Value('float64'), 'd_avg': Value('float64'), 'aj': Value('float64'), 'oa': Value('float64')}, 'tapvid_kubric': {'queries': Value('int64'), 'answered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('float64'), 'd_8': Value('float64'), 'd_16': Value('float64'), 'd_avg': Value('float64'), 'aj': Value('float64'), 'oa': Value('float64')}, 'tapvid_rgb_stacking': {'queries': Value('int64'), 'answered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('float64'), 'd_8': Value('float64'), 'd_16': Value('float64'), 'd_avg': Value('float64'), 'aj': Value('float64'), 'oa': Value('float64')}, 'vl_surgpt': {'queries': Value('int64'), 'answered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('float64'), 'd_8': Value('float64'), 'd_16': Value('float64'), 'd_avg': Value('float64'), 'aj': Value('float64'), 'oa': Value('float64')}, 'votsp2026': {'queries': Value('int64'), 'answered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('float64'), 'd_8': Value('float64'), 'd_16': Value('float64'), 'd_avg': Value('float64'), 'aj': Value('float64'), 'oa': Value('float64')}}}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 274, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 544, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3074, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2578, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2603, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 577, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 459, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 373, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2394, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2322, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
id: string
x: double
y: double
occluded: bool
visibility: double
vis: double
conf: double
scores_position_only: struct<queries: int64, answered: int64, visible_queries: int64, d_1: double, d_2: double, d_4: doubl (... 68 chars omitted)
child 0, queries: int64
child 1, answered: int64
child 2, visible_queries: int64
child 3, d_1: double
child 4, d_2: double
child 5, d_4: double
child 6, d_8: double
child 7, d_16: double
child 8, d_avg: double
child 9, aj: double
child 10, oa: double
queries: int64
by_category: struct<animals_nature: struct<queries: int64, answered: int64, visible_queries: int64, d_1: double, (... 1229 chars omitted)
child 0, animals_nature: struct<queries: int64, answered: int64, visible_queries: int64, d_1: double, d_2: double, d_4: doubl (... 68 chars omitted)
child 0, queries: int64
child 1, answered: int64
child 2, visible_queries: int64
child 3, d_1: double
child 4, d_2: double
child 5, d_4: double
child 6, d_8: double
child 7, d_16: double
child 8, d_avg: double
child 9, aj: double
child 10, oa: double
child 1, biomedical_scenes: struct<queries: int64, answered: int64, visible_queries: int64, d_1: double, d_2: double, d_4: doubl (... 68 chars omitted)
child 0, queries: int64
child 1, answered: int64
child 2, visible_queries: int64
child 3, d_1: double
child 4, d_2: double
child 5, d_4: double
child 6, d_8: double
ch
...
5, d_4: double
child 6, d_8: double
child 7, d_16: double
child 8, d_avg: double
child 9, aj: double
child 10, oa: double
child 13, vl_surgpt: struct<queries: int64, answered: int64, visible_queries: int64, d_1: double, d_2: double, d_4: doubl (... 68 chars omitted)
child 0, queries: int64
child 1, answered: int64
child 2, visible_queries: int64
child 3, d_1: double
child 4, d_2: double
child 5, d_4: double
child 6, d_8: double
child 7, d_16: double
child 8, d_avg: double
child 9, aj: double
child 10, oa: double
child 14, votsp2026: struct<queries: int64, answered: int64, visible_queries: int64, d_1: double, d_2: double, d_4: doubl (... 68 chars omitted)
child 0, queries: int64
child 1, answered: int64
child 2, visible_queries: int64
child 3, d_1: double
child 4, d_2: double
child 5, d_4: double
child 6, d_8: double
child 7, d_16: double
child 8, d_avg: double
child 9, aj: double
child 10, oa: double
scores: struct<queries: int64, answered: int64, visible_queries: int64, d_1: double, d_2: double, d_4: doubl (... 68 chars omitted)
child 0, queries: int64
child 1, answered: int64
child 2, visible_queries: int64
child 3, d_1: double
child 4, d_2: double
child 5, d_4: double
child 6, d_8: double
child 7, d_16: double
child 8, d_avg: double
child 9, aj: double
child 10, oa: double
answered: int64
to
{'queries': Value('int64'), 'answered': Value('int64'), 'scores': {'queries': Value('int64'), 'answered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('float64'), 'd_8': Value('float64'), 'd_16': Value('float64'), 'd_avg': Value('float64'), 'aj': Value('float64'), 'oa': Value('float64')}, 'scores_position_only': {'queries': Value('int64'), 'answered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('float64'), 'd_8': Value('float64'), 'd_16': Value('float64'), 'd_avg': Value('float64'), 'aj': Value('float64'), 'oa': Value('float64')}, 'by_category': {'animals_nature': {'queries': Value('int64'), 'answered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('float64'), 'd_8': Value('float64'), 'd_16': Value('float64'), 'd_avg': Value('float64'), 'aj': Value('float64'), 'oa': Value('float64')}, 'biomedical_scenes': {'queries': Value('int64'), 'answered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('float64'), 'd_8': Value('float64'), 'd_16': Value('float64'), 'd_avg': Value('float64'), 'aj': Value('float64'), 'oa': Value('float64')}, 'driving_street_scenes': {'queries': Value('int64'), 'answered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('flo
...
swered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('float64'), 'd_8': Value('float64'), 'd_16': Value('float64'), 'd_avg': Value('float64'), 'aj': Value('float64'), 'oa': Value('float64')}, 'tapvid_kubric': {'queries': Value('int64'), 'answered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('float64'), 'd_8': Value('float64'), 'd_16': Value('float64'), 'd_avg': Value('float64'), 'aj': Value('float64'), 'oa': Value('float64')}, 'tapvid_rgb_stacking': {'queries': Value('int64'), 'answered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('float64'), 'd_8': Value('float64'), 'd_16': Value('float64'), 'd_avg': Value('float64'), 'aj': Value('float64'), 'oa': Value('float64')}, 'vl_surgpt': {'queries': Value('int64'), 'answered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('float64'), 'd_8': Value('float64'), 'd_16': Value('float64'), 'd_avg': Value('float64'), 'aj': Value('float64'), 'oa': Value('float64')}, 'votsp2026': {'queries': Value('int64'), 'answered': Value('int64'), 'visible_queries': Value('int64'), 'd_1': Value('float64'), 'd_2': Value('float64'), 'd_4': Value('float64'), 'd_8': Value('float64'), 'd_16': Value('float64'), 'd_avg': Value('float64'), 'aj': Value('float64'), 'oa': Value('float64')}}}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
eval_vlm_results
Model outputs and scores of vision-language models on the 2Infinity Lab point-correspondence benchmark
(2inf/tracking_bench_vlm). One folder per
(benchmark release, model, prompt); each holds the scores (metrics.json), the parsed answers (predictions.jsonl),
the raw model replies (responses.jsonl), the configuration as run (resolved.yaml) and a README with the setting
and the score tables.
Results on benchmark_312 (frozen release releases/benchmark_312_chen_v1, 312 clips, 14,896 requests)
| folder | model | prompt / harness | d_avg | AJ | OA |
|---|---|---|---|---|---|
benchmark_312/qwen3.8-27b-fp8_twostep_single |
Qwen3.8-27B-FP8 | twostep_single, one call, no thinking | 24.16 | 13.52 | 78.31 |
benchmark_312/qwen3.8-27b-fp8_chen_v1_think_xhigh_zoom |
Qwen3.8-27B-FP8 | thinking xhigh + Muse zoom (multi-call) | 33.91 | 20.92 | 77.87 |
benchmark_312/qwen3.8-27b-fp8_chen_v1_think_medium_zoom |
Qwen3.8-27B-FP8 | thinking medium + Muse zoom (multi-call) | 29.38 | 17.15 | 78.60 |
benchmark_312/qwen3.8-27b-fp8_chen_v1_think_low_zoom |
Qwen3.8-27B-FP8 | thinking low + Muse zoom (multi-call) | 28.60 | 16.73 | 78.63 |
benchmark_312/qwen3.8-27b-fp8_chen_v1_think_xhigh |
Qwen3.8-27B-FP8 | thinking xhigh, one call | 25.81 | 15.53 | 76.46 |
benchmark_312/qwen3.8-27b-fp8_chen_v1_think_medium |
Qwen3.8-27B-FP8 | thinking medium, one call | 21.00 | 11.91 | 78.10 |
benchmark_312/qwen3.8-27b-fp8_chen_v1_think_low |
Qwen3.8-27B-FP8 | thinking low, one call | 21.01 | 11.96 | 77.81 |
benchmark_312/qwen3.8-flash-next-fp8_tsdsl_v3 |
Qwen3.8-Flash-Next-FP8 | tsdsl_v3 (per-model prompt), one call, no thinking | 22.21 | 12.89 | 77.39 |
benchmark_312/qwen3.5-122b-a10b-fp8_tsd_still_landmark |
Qwen3.5-122B-A10B-FP8 | tsd_still_landmark, one call, no thinking | 16.32 | 8.74 | 78.42 |
benchmark_312/deepseek-v4.1-flash_tsd_still_landmark |
DeepSeek-V4.1-Flash | tsd_still_landmark, one call, chat mode | 16.24 | 8.95 | 76.95 |
Results on the mini2k evaluation split (configs/data/benchmark_312_mini2k_real.txt, 2,000 of the 14,896 requests)
| folder | model | prompt / harness | answered | d_avg | AJ | OA |
|---|---|---|---|---|---|---|
benchmark_312/gemma-4-31b-fp8_codex_agent_low_mini2k_20261017 |
Gemma-4-31B-FP8 | Codex CLI agent (vlm_track.agent.qwen_codex, zoom.py tools, <= 8 views), thinking low; failed sessions retried once (strict, no retry: 1950/2000, 37.94 / 23.46 / 76.95) |
1996/2000 | 38.64 | 23.80 | 78.70 |
benchmark_312/qwen3.8-27b-fp8_codex_agent_xhigh_mini2k_20261021 |
Qwen3.8-27B-FP8 | Codex CLI agent (vlm_track.agent.qwen_codex, zoom.py tools, <= 8 views), thinking xhigh; failed sessions not retried, an unanswered request counts as wrong (Table 2, AgenTAP Qwen3.8-27B) |
1867/2000 | 43.99 | 29.02 | 74.95 |
benchmark_312/qwen3.8-27b-fp8_codex_agent_medium_mini2k_20261020 |
Qwen3.8-27B-FP8 | Codex CLI agent (vlm_track.agent.qwen_codex, zoom.py tools, <= 8 views), thinking medium; failed sessions not retried, an unanswered request counts as wrong |
1904/2000 | 43.80 | 28.51 | 77.30 |
benchmark_312/qwen3.8-27b-fp8_codex_agent_low_mini2k_20261020 |
Qwen3.8-27B-FP8 | Codex CLI agent (vlm_track.agent.qwen_codex, zoom.py tools, <= 8 views), thinking low; failed sessions not retried, an unanswered request counts as wrong |
1979/2000 | 41.60 | 26.29 | 78.60 |
benchmark_312/qwen3.8-27b-fp8_agentap_staged_xhigh_mini2k_20261120 |
Qwen3.8-27B-FP8 | AgenTAP staged workflow in the open-model Codex agent (vlm_track.agent.qwen_codex, workflow: staged: 2 overview views and a first answer, then the RoMa v2 and two-frame AllTracker candidates, then 5 close views), thinking xhigh, 3,600 s per session; failed sessions not retried, an unanswered request counts as wrong |
1962/2000 | 57.47 | 38.63 | 79.60 |
benchmark_312/qwen3.8-27b-fp8_agentap_staged_xhigh_reask38_mini2k_20261121 |
Qwen3.8-27B-FP8 | the run of the row above with its 38 failed sessions retried once (a separate run, same settings; as for Gemma-4-31B below); the 1962 answered requests keep their answers; strict, no retry: the row above; retrying only its 5 timeouts, as the closed models' runs do: 1967/2000, 57.59 / 38.69 / 79.80 | 1997/2000 | 58.11 | 38.84 | 80.90 |
benchmark_312/gemma-4-31b-fp8_agentap_staged_low_mini2k_20261119 |
Gemma-4-31B-FP8 | AgenTAP staged workflow in the open-model Codex agent (vlm_track.agent.qwen_codex, workflow: staged: 2 overview views and a first answer, then the RoMa v2 and two-frame AllTracker candidates, then 5 close views), thinking on, 2,400 s per session; failed sessions retried once (strict, no retry: 1955/2000, 49.93 / 31.88 / 77.35) |
1994/2000 | 50.82 | 32.31 | 78.95 |
benchmark_312/muse-glimmer-30b_codex_agent_xhigh_mini2k_20261109 |
Muse Glimmer 30B | Codex CLI agent (vlm_track.agent.qwen_codex, zoom.py tools, <= 8 views), reasoning strength xhigh (temperature 1.0); failed sessions not retried, an unanswered request counts as wrong |
1924/2000 | 41.99 | 26.23 | 75.50 |
benchmark_312/muse-glimmer-30b_codex_agent_high_mini2k_20261109 |
Muse Glimmer 30B | Codex CLI agent (vlm_track.agent.qwen_codex, zoom.py tools, <= 8 views), reasoning strength high (temperature 1.0); failed sessions not retried, an unanswered request counts as wrong |
1955/2000 | 41.89 | 26.01 | 78.20 |
benchmark_312/muse-glimmer-30b_codex_agent_medium_mini2k_20261109 |
Muse Glimmer 30B | Codex CLI agent (vlm_track.agent.qwen_codex, zoom.py tools, <= 8 views), reasoning strength medium (temperature 1.0); failed sessions not retried, an unanswered request counts as wrong |
1949/2000 | 41.37 | 25.70 | 78.35 |
benchmark_312/muse-glimmer-30b_codex_agent_low_mini2k_20261109 |
Muse Glimmer 30B | Codex CLI agent (vlm_track.agent.qwen_codex, zoom.py tools, <= 8 views), reasoning strength low (temperature 1.0); failed sessions not retried, an unanswered request counts as wrong |
1958/2000 | 37.64 | 23.03 | 77.25 |
benchmark_312/qwen3.8-27b-fp8_think_with_images_xhigh_mini2k_20261022 |
Qwen3.8-27B-FP8 | think with images (official Qwen-Agent image_zoom_in_tool), thinking xhigh, multi-call |
1861/2000 | 28.92 | 17.61 | 75.05 |
benchmark_312/qwen3.8-27b-fp8_think_with_images_medium_mini2k_20261022 |
Qwen3.8-27B-FP8 | think with images (official Qwen-Agent image_zoom_in_tool), thinking medium, multi-call |
1981/2000 | 24.97 | 14.58 | 79.20 |
benchmark_312/qwen3.8-27b-fp8_think_with_images_low_mini2k_20261022 |
Qwen3.8-27B-FP8 | think with images (official Qwen-Agent image_zoom_in_tool), thinking low, multi-call |
1994/2000 | 22.46 | 12.90 | 79.30 |
benchmark_312/claude-opus-5.5_one_call_claude_code_mini2k_20261106 |
Claude Opus 5.5 | Table 1 one-call request through Claude Code 2.1.318 (claude -p, no tools, effort high; Claude Code's system prompt included) |
2000/2000 | 48.90 | 32.34 | 82.35 |
benchmark_312/qwen3.8-27b-fp8_codex_agent_medium_repro68_mini2k_20261024 asks 68 selected mini2k requests four more times
with the Qwen3.8 Codex agent at thinking medium (a repeatability study; the 68 are not a random sample).
The *_agentap_*_mini2k_* folders (GPT-6 Astra and GPT-6.1 Sol through the AgenTAP staged Codex agent; Qwen3.8-27B-FP8 and Gemma-4-31B-FP8 through the same staged workflow in the open-model Codex agent, qwen3.8-27b-fp8_agentap_staged_xhigh_mini2k_20261120 (with its failed sessions retried once: qwen3.8-27b-fp8_agentap_staged_xhigh_reask38_mini2k_20261121) and gemma-4-31b-fp8_agentap_staged_low_mini2k_20261119, also rows of the table above; Claude Opus 5.5 through the same staged workflow in Claude Code, on the two images and, in claude-opus-5.5_agentap_staged_video_mini2k_20261106, on the full video) carry their scores in their own READMEs. gpt-6-astra_agentap_staged_candfirst_mini2k_20261111 is the order ablation of the two-image staged workflow (the candidates shown before the first answer instead of after it): AJ 54.8 / APD 71.3 / OA 86.2 vs 54.9 / 71.4 / 86.2, paired differences within ±1.
Results on scatter_v9 (GPT-6.1 Sol, Codex agent, points per session)
GPT-6.1 Sol (reasoning high) through the Codex CLI agent with the zoom.py tools. p1 = one point per session (the original
one-point agent, vlm_track.agent.codex); p2 / p4 / p8 / p16 = 2 / 4 / 8 / 16 points per session (vlm_track.agent.codex_mi, magnified views
of every point on a separate last image), prompt asks to take the points one at a time; p16 free order = the same without
that sentence. The same points in every protocol of a subset (paired by target id). Each folder's sessions.tar.gz holds
every session's events, trace and the images the model saw.
| folder | subset | protocol | requests | points | d_avg | AJ | OA |
|---|---|---|---|---|---|---|---|
scatter_v9/gpt-6.1-sol_codex_moving_p1_20261103 |
moving | p1 | 768 | 768 | 72.81 | 64.58 | 96.48 |
scatter_v9/gpt-6.1-sol_codex_moving_p2_20261106 |
moving | p2 | 384 | 768 | 74.09 | 65.61 | 97.27 |
scatter_v9/gpt-6.1-sol_codex_moving_p4_20261103 |
moving | p4 | 192 | 768 | 73.31 | 64.39 | 96.61 |
scatter_v9/gpt-6.1-sol_codex_moving_p8_20261106 |
moving | p8 | 96 | 768 | 73.15 | 64.49 | 96.48 |
scatter_v9/gpt-6.1-sol_codex_moving_p16_20261104 |
moving | p16 | 48 | 768 | 74.30 | 65.43 | 97.27 |
scatter_v9/gpt-6.1-sol_codex_moving_p16_free_order_20261105 |
moving | p16 free order | 48 | 768 | 74.51 | 65.70 | 97.40 |
scatter_v9/gpt-6.1-sol_codex_static_p1_20261103 |
static | p1 | 528 | 528 | 96.33 | 94.01 | 99.24 |
scatter_v9/gpt-6.1-sol_codex_static_p2_20261106 |
static | p2 | 264 | 528 | 96.55 | 94.21 | 99.43 |
scatter_v9/gpt-6.1-sol_codex_static_p4_20261103 |
static | p4 | 132 | 528 | 96.78 | 94.57 | 99.43 |
scatter_v9/gpt-6.1-sol_codex_static_p8_20261106 |
static | p8 | 66 | 528 | 96.86 | 94.75 | 99.43 |
scatter_v9/gpt-6.1-sol_codex_static_p16_20261104 |
static | p16 | 33 | 528 | 96.86 | 94.71 | 99.43 |
scatter_v9/gpt-6.1-sol_codex_static_p16_free_order_20261105 |
static | p16 free order | 33 | 528 | 96.89 | 94.83 | 99.43 |
scatter_v9/gpt-6.1-sol_codex_mixed_p1_20261103 |
mixed | p1 | 1008 | 1008 | 87.70 | 81.01 | 98.21 |
scatter_v9/gpt-6.1-sol_codex_mixed_p2_20261106 |
mixed | p2 | 504 | 1008 | 88.39 | 81.52 | 98.61 |
scatter_v9/gpt-6.1-sol_codex_mixed_p4_20261103 |
mixed | p4 | 252 | 1008 | 87.92 | 81.12 | 98.31 |
scatter_v9/gpt-6.1-sol_codex_mixed_p8_20261106 |
mixed | p8 | 126 | 1008 | 88.69 | 82.60 | 97.72 |
scatter_v9/gpt-6.1-sol_codex_mixed_p16_20261104 |
mixed | p16 | 63 | 1008 | 88.81 | 82.45 | 98.21 |
scatter_v9/gpt-6.1-sol_codex_mixed_p16_free_order_20261105 |
mixed | p16 free order | 63 | 1008 | 90.04 | 84.09 | 98.41 |
Experiment B: one point, several target frames (b_p1)
The same points as above, but one session sees the point's reference frame and 2 or 3 of its later frames at once
(gaps 18 and 36 in one request, 18, 36 and 54 in a second request for the same point; vlm_track.agent.codex_mi, sheet
layout, commands name the frame with --in targetN). points = target frames scored; each point's gaps 18 and 36 are asked
twice, so the metrics weight gaps 18 : 36 : 54 as 2 : 2 : 1. Paired with the one-frame runs on the same target frames, the
3-frame request gains +6.8 d_avg over p1 on moving (79.6 vs 72.8; 95% clip-bootstrap CI +4.8 … +8.8) and +5.3 … +6.5
over p2 … p16; static is unchanged (±0.3); mixed gains +2.4 over p1 (+1.2 … +2.1 over p2 … p16).
| folder | subset | protocol | requests | points | d_avg | AJ | OA |
|---|---|---|---|---|---|---|---|
scatter_v9/gpt-6.1-sol_codex_moving_b_p1_20261108 |
moving | b_p1 | 512 | 1280 | 80.58 | 73.32 | 97.89 |
scatter_v9/gpt-6.1-sol_codex_static_b_p1_20261108 |
static | b_p1 | 352 | 880 | 96.89 | 94.76 | 99.43 |
scatter_v9/gpt-6.1-sol_codex_mixed_b_p1_20261108 |
mixed | b_p1 | 672 | 1680 | 90.07 | 84.27 | 98.39 |
| Metrics follow TAP-Vid: offsets are scaled to a 256 x 256 image; d_avg is the mean share of visible targets within | |||||||
| 1, 2, 4, 8 and 16 px; AJ is the average Jaccard over the same thresholds; OA is the occlusion accuracy. |
- Downloads last month
- 1,137