Complete protocol-screened leaderboard

Educational scores retain the 20 August historical snapshot. High-Motion v2 compares CPU-rescored archived baselines with one new GRT run; no baseline inference was repeated. Download CSV · Versioned source hashes and comparison limits · Public reproduction code · Corrected-reference reproduction guide.

29 Educational results, 19 corrected High-Motion preview results, and 12 educational GRT comparison rows (60 CSV records). Educational: 634 QA / 317 videos. High-Motion v2: the fixed 1,000-record preview of 3,243 source records; metric-specific valid-reference coverage is shown separately. All numeric cells retain full precision in their title and downloadable data.

Educational High-FPS Videos

29 published methods. Rank by reported Open MOS, then Token F1; missing Open MOS sorts last and is never filled in. Open MOS judge: Qwen/Qwen3-VL-32B-Instruct. The verified GRT profiles use eight sampled frames, not a measured high-FPS operating point.

29 educational methods; 634 items each
Rank	Model	Method ID	Items	Open MOS ↑	Token F1 ↑	CER ↓	WER ↓	Patch recompute ↓	Reference patch compute ↓	Sampling density (fps)	Mean throughput (fps) ↑	Source
1	Gemini 3.1 Pro Preview (gemini-3.1-pro-preview)	gemini-3.1-pro-preview	634	1.6041	0.159633	1.26881	1.24166	—	—	—	—	gemini
2	GRT (Qwen2.5-VL 7B, route floor 0.80 / 0.55, cap 48)	grt_qwen2_5_vl_7b_dual_floor_s080_o055_cap48	634	1.59148	0.0489087	1.06401	1.05377	0.847719	0.847719	0.00747917	2.48043	open
3	GRT (Qwen2.5-VL 3B, threshold 0.3)	grt_qwen2_5_vl_3b_t03	634	1.58991	0.0995837	1.13915	1.08351	0.885659	0.885659	0.00747917	1.67264	open
4	Qwen3-VL 8B Instruct	qwen3_vl_8b	634	1.29811	0.0360398	1.05889	1.05198	—	—	—	—	open
5	Qwen2.5-VL 7B Instruct	qwen2_5_vl_7b	634	1.28391	0.0347846	1.04624	1.03975	—	—	—	—	open
6	Gemini 3.6 Flash (gemini-3.6-flash)	gemini-3.6-flash	634	1.24921	0.0814611	1.10168	1.06796	—	—	—	—	gemini
7	Qwen3-VL 2B Instruct	qwen3_vl_2b	634	1.14511	0.0331448	1.0455	1.0365	—	—	—	—	open
8	Qwen3-VL 4B Instruct	qwen3_vl_4b	634	1.1388	0.0348606	1.05593	1.05119	—	—	—	—	open
9	Qwen2.5-VL 3B Instruct	qwen2_5_vl_3b	634	1.13565	0.0343388	1.04393	1.03399	—	—	—	—	open
10	Qwen2.5-VL 72B Instruct	qwen2_5_vl_72b	634	0.944795	0.0268628	1.05934	1.04789	—	—	—	—	open
11	Qwen2-VL 2B Instruct	qwen2_vl_2b	634	0.790221	0.0200933	1.03692	1.0334	—	—	—	—	open
12	Qwen2.5-VL 32B Instruct	qwen2_5_vl_32b	634	0.5	0.0250276	1.05448	1.04856	—	—	—	—	open
13	Qwen3-VL 32B Instruct	qwen3_vl_32b	634	0.175079	0.0184676	1.05239	1.05085	—	—	—	—	open
14	GRT (LLaVA-OneVision Qwen2 0.5B, verified Route31)	grt_llava_onevision_0_5b_hf_route31_t0001	634	0.119874	0.0141964	1.11093	1.10273	0.867154	0.867154	0.00747917	4.86514	open
15	LLaVA-OneVision Qwen2 0.5B	llava_onevision_0_5b	634	0.116719	0.0140879	1.03765	1.03756	—	—	—	—	open
16	LLaVA-OneVision 1.5 8B Instruct	llava_onevision_1_5_8b	634	—	0.0349748	1.07243	1.05686	—	—	—	—	open
17	Qwen2-VL 7B Instruct	qwen2_vl_7b	634	—	0.0305096	1.05309	1.04788	—	—	—	—	open
18	Phi-4 Multimodal Instruct	phi4_multimodal	634	—	0.0275287	1.03008	1.02762	—	—	—	—	open
19	InternVL3 1B	internvl3_1b	634	—	0.0271622	1.062	1.0606	—	—	—	—	open
20	InternVL3 8B	internvl3_8b	634	—	0.0250303	1.07046	1.06601	—	—	—	—	open
21	VideoLLaMA3 7B	videollama3_7b	634	—	0.0249369	1.05106	1.04975	—	—	—	—	open
22	LongVA 7B	longva_7b	634	—	0.0246616	1.04678	1.04241	—	—	—	—	open
23	InternVL3 2B	internvl3_2b	634	—	0.0223091	1.03408	1.03574	—	—	—	—	open
24	InternVL2.5 8B	internvl2_5_8b	634	—	0.0188533	1.05738	1.05719	—	—	—	—	open
25	InternVL2.5 4B	internvl2_5_4b	634	—	0.01792	1.02051	1.02362	—	—	—	—	open
26	InternVL2.5 1B	internvl2_5_1b	634	—	0.0178427	1.03147	1.0336	—	—	—	—	open
27	LLaVA-OneVision Qwen2 7B	llava_onevision_original	634	—	0.0175202	1.04056	1.03982	—	—	—	—	open
28	VideoLLaMA3 2B	videollama3_2b	634	—	0.0158476	1.0245	1.02683	—	—	—	—	open
29	InternVL2.5 2B	internvl2_5_2b	634	—	0.0119298	1.01729	1.02037	—	—	—	—	open

High-Motion v2: right-ring reference correction, preview-1000

19 methods: 18 archived baselines rescored on CPU and one new HF 0.5B GRT run. All use the same fixed first 1,000 source records, not a full 3,243-record evaluation. The reference target is rightRingFingerMetacarpal, the named right-palm/ring-finger-base proxy; questions and original sampled positions are unchanged. This versioned correction does not claim to recover the original annotation constructor.

Rank by Grid Accuracy descending; exact ties share a competition rank and are ordered by method ID. Null scores are unranked. Every metric is a macro mean over its defined per-record values, with visible metric-specific row and slot/edge coverage. A dash means undefined under the reference mask, not zero. Invalid slots never shift later predictions; FDE uses the original final slot, and transitions require adjacent valid original slots. Token F1 uses canonical label bags on valid positions, with surplus outputs penalized. All 1,000 records remain counted even when no positions are scoreable.

New GRT predictions versus rescored, configuration-checked archived predictions; archived baseline weight revisions and consumed-tensor identity are unproven. This is not a freshly rerun byte-identical paired experiment or a statistical-significance claim.

19 corrected-reference preview methods; eight slots per record
Rank	Model	Method ID	Cached/new records	Grid Accuracy ↑	Grid Accuracy ↑ coverage	Grid ADE ↓	Grid ADE ↓ coverage	Grid FDE ↓	Grid FDE ↓ coverage	Transition Accuracy ↑	Transition Accuracy ↑ coverage	Token F1 ↑	Token F1 ↑ coverage	Prediction provenance
1	LLaVA-OneVision-2-8B-Instruct	llava_onevision_2_8b	1000	0.512279	861 rows / 6015 slots	0.279313	861 rows / 6015 slots	0.206212	669 rows / 669 slots	0.731701	852 rows / 5101 edges	0.513393	861 rows / 6015 slots	archived_baseline
2	Qwen3-VL-8B-Instruct	qwen3_vl_8b	1000	0.426745	861 rows / 6015 slots	0.357624	861 rows / 6015 slots	0.310443	669 rows / 669 slots	0.734879	852 rows / 5101 edges	0.426938	861 rows / 6015 slots	archived_baseline
3	LLaVA-OneVision-1.5-8B-Instruct	llava_onevision_1_5_8b	1000	0.41958	861 rows / 6015 slots	0.374625	861 rows / 6015 slots	0.35337	669 rows / 669 slots	0.729748	852 rows / 5101 edges	0.41958	861 rows / 6015 slots	archived_baseline
4	VideoLLaMA3-7B	videollama3_7b	1000	0.416147	861 rows / 6015 slots	0.479898	861 rows / 6015 slots	0.591313	669 rows / 669 slots	0.440789	852 rows / 5101 edges	0.462768	861 rows / 6015 slots	archived_baseline
5	Qwen3-VL-32B-Instruct	qwen3_vl_32b	1000	0.360901	861 rows / 6015 slots	0.384434	861 rows / 6015 slots	0.392697	669 rows / 669 slots	0.72526	852 rows / 5101 edges	0.363749	861 rows / 6015 slots	archived_baseline
6	Qwen3-VL-4B-Instruct	qwen3_vl_4b	1000	0.359986	861 rows / 6015 slots	0.359877	861 rows / 6015 slots	0.368279	669 rows / 669 slots	0.735113	852 rows / 5101 edges	0.359986	861 rows / 6015 slots	archived_baseline
7	Qwen3-VL-2B-Instruct	qwen3_vl_2b	1000	0.275784	861 rows / 6015 slots	0.423396	861 rows / 6015 slots	0.471795	669 rows / 669 slots	0.313081	852 rows / 5101 edges	0.35428	861 rows / 6015 slots	archived_baseline
8	Qwen2.5-VL-72B-Instruct	qwen2_5_vl_72b	1000	0.235602	861 rows / 6015 slots	0.513407	861 rows / 6015 slots	0.569024	669 rows / 669 slots	0.648879	852 rows / 5101 edges	0.258947	861 rows / 6015 slots	archived_baseline
9	Qwen2.5-VL-32B-Instruct	qwen2_5_vl_32b	1000	0.176484	861 rows / 6015 slots	0.505219	861 rows / 6015 slots	0.504983	669 rows / 669 slots	0.56006	852 rows / 5101 edges	0.189601	861 rows / 6015 slots	archived_baseline
10	LongVA-7B	longva_7b	1000	0.0760605	861 rows / 6015 slots	0.770476	861 rows / 6015 slots	0.632581	669 rows / 669 slots	0.120532	852 rows / 5101 edges	0.155908	861 rows / 6015 slots	archived_baseline
11	Phi-4-multimodal-instruct	phi4_multimodal	1000	0.0722015	861 rows / 6015 slots	1.1359	861 rows / 6015 slots	1.21451	669 rows / 669 slots	0.0180751	852 rows / 5101 edges	0.11382	861 rows / 6015 slots	archived_baseline
12	Qwen2-VL-2B-Instruct	qwen2_vl_2b	1000	0.0656546	861 rows / 6015 slots	0.771788	861 rows / 6015 slots	0.625801	669 rows / 669 slots	0.225872	852 rows / 5101 edges	0.105367	861 rows / 6015 slots	archived_baseline
13	LLaVA-OneVision HF 7B	llava_onevision_original	1000	0.0624772	861 rows / 6015 slots	0.881176	861 rows / 6015 slots	0.846691	669 rows / 669 slots	0.185616	852 rows / 5101 edges	0.116204	861 rows / 6015 slots	archived_baseline
14	VideoLLaMA3-2B	videollama3_2b	1000	0.0549375	861 rows / 6015 slots	1.12491	861 rows / 6015 slots	1.15021	669 rows / 669 slots	0.000503018	852 rows / 5101 edges	0.115104	861 rows / 6015 slots	archived_baseline
15	Qwen2-VL-7B-Instruct	qwen2_vl_7b	1000	0.0524114	861 rows / 6015 slots	0.855721	861 rows / 6015 slots	0.905482	669 rows / 669 slots	0.30432	852 rows / 5101 edges	0.069898	861 rows / 6015 slots	archived_baseline
16	GRT · LLaVA-OneVision HF 0.5B (motion SSIM 0.001)	grt_llava_hf_0_5b_motion_ssim_t0001	1000	0.0495036	861 rows / 6015 slots	1.0862	861 rows / 6015 slots	1.06721	669 rows / 669 slots	0.00534038	852 rows / 5101 edges	0.0450525	861 rows / 6015 slots	new_grt
17	LLaVA-OneVision HF 0.5B	llava_onevision_0_5b	1000	0.0468683	861 rows / 6015 slots	1.09641	861 rows / 6015 slots	1.07772	669 rows / 669 slots	0.0056841	852 rows / 5101 edges	0.0418983	861 rows / 6015 slots	archived_baseline
18	Qwen2.5-VL-3B-Instruct	qwen2_5_vl_3b	1000	0.0150863	861 rows / 6015 slots	0.932863	861 rows / 6015 slots	0.998825	669 rows / 669 slots	0.576009	852 rows / 5101 edges	0.0201869	861 rows / 6015 slots	archived_baseline
19	Qwen2.5-VL-7B-Instruct	qwen2_5_vl_7b	1000	0.0149204	861 rows / 6015 slots	1.18876	861 rows / 6015 slots	1.30077	669 rows / 669 slots	0.579396	852 rows / 5101 edges	0.0135712	861 rows / 6015 slots	archived_baseline

GRT versus its corresponding HF 0.5B baseline

GRT exceeds the corresponding HF 0.5B baseline on observed Grid Accuracy. Point tolerance: 1e-12. Every observed difference is retained, including regressions. This is a point-estimate comparison, not a statistical-significance claim.

All five corrected-reference GRT versus baseline differences
Metric	Rescored HF 0.5B baseline	GRT	GRT minus baseline	Oriented improvement (positive is better)	Exceeds point tolerance
Grid Accuracy ↑	0.0468683	0.0495036	0.00263536	0.00263536	yes
Grid ADE ↓	1.09641	1.0862	-0.0102074	0.0102074	yes
Grid FDE ↓	1.07772	1.06721	-0.010501	0.010501	yes
Transition Accuracy ↑	0.0056841	0.00534038	-0.000343729	-0.000343729	no
Token F1 ↑	0.0418983	0.0450525	0.00315414	0.00315414	yes

Legacy High-Motion results remain withheld

The old target/reference hold and immutable 27-run historical protocol audit remain intact. No old High-Motion score is mixed into the corrected v2 table or its CSV cohort. Reference, scorer, prediction and release provenance are available in the versioned numeric audit.

GRT vs archived and matched controls

All three promoted educational candidates exceed every contracted Open MOS and Token F1 floor, with 11.43–15.23% fewer patch projections than the matched all-patch route. LLaVA-OneVision 7B failed its MOS gate and is not promoted. These are observed point estimates, not statistical-significance claims. Archived Qwen baselines differ from stronger matched controls: their entire gap must not be attributed to GRT.

All 12 educational comparison rows; 634 items and eight sampled frames per method
Family	Control	Method ID	Items	Open MOS ↑	Token F1 ↑	Patch recompute ↓	Mean throughput (fps) ↑	Mean request time (s) ↓
LLaVA-OneVision 0.5B Route31	Archived public baseline	llava_onevision_0_5b	634	0.116719	0.0140879	—	—	—
LLaVA-OneVision 0.5B Route31	Matched quality baseline	llava_onevision_0_5b_hf_route31_base	634	0.116719	0.0139815	1	4.81928	2.0401
LLaVA-OneVision 0.5B Route31	Matched all-patch control	llava_onevision_0_5b_hf_route31_exact	634	0.116719	0.0139815	1	4.8239	2.00908
LLaVA-OneVision 0.5B Route31	GRT candidate	grt_llava_onevision_0_5b_hf_route31_t0001	634	0.119874	0.0141964	0.867154	4.86514	2.00911
Qwen2.5-VL 3B t03	Archived public baseline	qwen2_5_vl_3b	634	1.13565	0.0343388	—	—	—
Qwen2.5-VL 3B t03	Matched all-patch control	qwen2_5_vl_3b_grt_all	634	1.55836	0.0991358	1	1.74994	5.42461
Qwen2.5-VL 3B t03	Matched quality baseline	qwen2_5_vl_3b_quality_base	634	1.55363	0.0988916	1	1.63172	5.88014
Qwen2.5-VL 3B t03	GRT candidate	grt_qwen2_5_vl_3b_t03	634	1.58991	0.0995837	0.885659	1.67264	5.70157
Qwen2.5-VL 7B route floor	Archived public baseline	qwen2_5_vl_7b	634	1.28391	0.0347846	—	—	—
Qwen2.5-VL 7B route floor	Matched all-patch control	qwen2_5_vl_7b_floor_grt_all	634	1.57571	0.0487774	1	2.46645	3.39562
Qwen2.5-VL 7B route floor	Matched quality baseline	qwen2_5_vl_7b_floor_quality_base	634	1.57571	0.0487774	1	2.4373	3.4355
Qwen2.5-VL 7B route floor	GRT candidate	grt_qwen2_5_vl_7b_dual_floor_s080_o055_cap48	634	1.59148	0.0489087	0.847719	2.48043	3.38126

Not all metrics improve. Qwen 3B GRT reports mean throughput 1.67264 fps versus 1.74994 for its all-patch control (about 4.42% lower), even though its Open MOS, Token F1 and patch reuse improve. Route31 mean request time is essentially unchanged versus its all-patch control. Throughput is the mean of per-request sampled-frame rates, not total frames divided by total campaign time, and these single historical runs do not establish repeated speedup. Patch ratios measure patch projection only, not end-to-end FLOPs.

Original immutable 32-row historical snapshot remains byte-identical to the archived numerical bundle; it is not the current release-policy view. The hold does not rewrite historical evidence or change Educational scores, ranks or GRT gates. Full dataset access, GPU/judge reproduction and manuscript alignment remain separate release checks.