JoshuaW1997/mm-cot-results

MM-CoT results

• outputs/: published results, compact CoT records, final analysis artifacts and required CPU readers.
• recovery/: the server's active consistent Qwen3/Gemma feature re-extraction from saved CoTs. This folder stays in place until those runs finish; its successful results will then be merged into the canonical outputs.

Read the saved results

outputs/records/index.json lists the compact archive and per-model/dataset indexes. Each indexed CoT retains complete text, exact token/step identities, spectra/traces, selected visual IDs, pooled attention and retained token scalar sequences. The records do not contain source images or hidden-state tensors.

The archive contains 210367 retained records. This counts available archived records, not completed replacement experiments. Retired SmolVLM trial records are excluded. New server uploads in recovery/ remain separate until the final merge.

Final analysis summaries and verification files remain under each model's published analysis scope. outputs/metadata/scopes/ maps those scopes to their exact frozen encoders, splits, population transitions and calibration, including files shared between scopes. outputs/metadata/accepted-exclusions.json records the final coverage exclusions.

Use outputs/tools/README.txt for CPU reading and metric reconstruction. Scientific pilot CoTs used to fit the query atlas remain part of their experiments.

Worker logs, scheduling plans, migration histories, deduplication reports and retired-model trial folders are excluded. Saved generation inputs that the active re-extraction still needs remain with their model outputs; redundant Qwen2.5 generation copies are removed because its full final records retain those IDs and text.

Application experiments (production-v1, intermediate)

Five downstream applications of the coupled text–visual Markov process are run on the Qwen3-VL-2B and Gemma-3-4B cells; Qwen2.5-VL-3B cells are unavailable (no corrected frozen scope). Code, configs and every per-cell artifact live under applications/production-v1/ (reproducibility/ holds the exact code). Protocol: applications/production-v1/reproducibility/docs/experiments/PROTOCOL.md.

Status. These are intermediate results from runs that were interrupted on the previous server (snapshot of 2026-10-21 18:05 UTC). The study is not complete and is not the final matched report: allocation and LoRA training have no results yet, several cells are missing for some applications, and VisualPRM and process-error annotations are unavailable. The run is being resumed on a new server after the source features are re-extracted and hash-verified against source-freeze.json; live status is in applications/production-v1/EXPERIMENT_STATUS.md.

Reading the tables. Numbers are accuracies or rates in %. Differences are paired percentage points with a 95% bootstrap CI and a two-sided sign-randomization p-value. The bootstrap unit is the image/question group, and repeated seeds are averaged within each group. p-values are uncorrected; the protocol applies Holm correction in the final report. Scorer arms (protocol §Shared design):

• length: CoT length only.
• text: textual-chain features.
• independent: text + attention + visual features + product-chain (uncoupled) surprise.
• coupled: independent + paired occupancy/edges + query-specific coupled kernel K_q and conditional marginals.
• Every arm uses the same capacity-limited logistic head.

Missing cells.

• open_r1: no correctness head can be fitted, because every training label has the same value (the base held-out accuracy below is 0%).
• gemma3_4b/mpcc and qwen3/open_r1: CPU preparation failed before ranking.
• qwen3/mpcc repair was 194/229 queries complete, and gemma3_4b/visual_cot repair 11/194, when the snapshot was taken; neither has a results file yet.

1. Answer selection from multiple CoTs

Setup: for each test query, choose one of K independently sampled target CoTs using each arm's predicted correctness. All arms see the same target pool. Oracle@K is the fraction of queries with at least one correct candidate.

Cell	K	Test queries	Majority	Mean log-prob	Independent	Text	Coupled	Oracle@K	Coupled − majority	Coupled − independent
gemma3_4b/mathvista	4	191	61.8	60.7	61.8	66.0	60.7	80.6	-1.8 [-7.1, +3.5], p=0.587	-1.0 [-6.2, +4.0], p=0.753
gemma3_4b/mathvista	15	191	64.9	64.4	66.0	62.3	66.0	88.5	+0.4 [-5.8, +6.7], p=0.929	+0.3 [-5.0, +5.6], p=1.000
gemma3_4b/scienceqa	4	190	79.5	78.9	78.4	79.5	76.8	92.1	-2.7 [-6.6, +0.9], p=0.228	-1.2 [-4.2, +1.5], p=0.551
gemma3_4b/scienceqa	15	190	83.2	81.6	79.5	77.9	78.9	95.8	-4.5 [-9.3, +0.3], p=0.094	-1.8 [-5.8, +2.2], p=0.425
gemma3_4b/visual_cot	4	194	8.2	6.7	9.3	8.8	7.2	10.8	-1.1 [-4.3, +2.2], p=0.697	-2.5 [-5.1, -0.4], p=0.123
gemma3_4b/visual_cot	15	194	8.2	8.8	10.3	9.8	8.2	13.4	+0.2 [-2.5, +3.1], p=1.000	-2.0 [-5.1, +0.7], p=0.262
qwen3/mathvista	4	191	76.4	72.3	73.3	74.3	73.3	87.4	-3.0 [-8.3, +2.3], p=0.289	+0.0 [-4.7, +4.7], p=1.000
qwen3/mathvista	15	191	75.4	67.5	73.3	73.3	75.9	94.2	+1.1 [-4.9, +7.0], p=0.832	+2.6 [-2.9, +7.9], p=0.428
qwen3/mpcc	4	228	6.1	5.3	5.3	6.1	7.0	11.4	+1.3 [-2.7, +5.2], p=0.651	+1.4 [-1.7, +4.6], p=0.356
qwen3/mpcc	15	228	7.9	3.5	7.9	6.6	8.3	25.0	-0.1 [-4.0, +4.2], p=1.000	-0.5 [-4.9, +3.8], p=0.942
qwen3/scienceqa	4	190	87.9	85.8	88.9	89.5	86.3	96.8	-1.2 [-4.7, +2.2], p=0.560	-1.3 [-5.2, +2.5], p=0.573
qwen3/scienceqa	15	190	91.1	85.3	90.5	88.9	87.4	97.9	-3.6 [-8.1, +0.3], p=0.132	-2.7 [-7.5, +1.8], p=0.346
qwen3/visual_cot	4	194	20.6	16.0	19.6	20.6	21.1	26.3	+0.4 [-4.2, +4.3], p=0.928	+2.4 [-0.7, +6.0], p=0.226
qwen3/visual_cot	15	194	19.6	13.9	18.6	21.6	20.1	33.0	-0.7 [-5.4, +3.6], p=0.819	+1.6 [-2.0, +5.1], p=0.414

Conclusions.

• The coupled scorer does not improve answer selection in any available cell. Paired differences against majority voting and against the independent-chain control range from −4.5 to +2.6 pp, and no test reaches p<0.05. The single bootstrap CI that excludes zero is Gemma/Visual-CoT at K=4 versus independent (−2.5 pp, p=0.12), where coupled is worse.
• There is headroom: Oracle@15 exceeds the best scorer by 7–23 pp on MathVista and ScienceQA. None of the process-feature scorers captures that headroom.
• Majority voting is consistently competitive. Mean token log-probability is the lowest or tied-lowest selector in 10 of 14 rows.
• Visual-CoT and MPCC are near floor (<25% accuracy), which leaves little signal for ranking.

2. Early failure triage

Setup: from a CoT prefix ending at step 2, 4 or 8, predict whether the final answer will be wrong. "Error capture @20%" is the share of eventual failures caught when the top 20% riskiest CoTs are flagged for review. The last column is the coupled − text difference in error capture at that equal 20% budget. Fewer CoTs remain eligible at later checkpoints because short trajectories end before the step. First-error localization (process diagnosis) is unavailable because no independent step annotations have been supplied.

Step 2:

Cell	Step	Test CoTs	Fail rate	AUROC length	AUROC text	AUROC independent	AUROC coupled	Error capture @20% (coupled)	Coupled − text capture
gemma3_4b/mathvista	2	2863	38.5	0.542	0.545	0.549	0.586	24.3	-1.1 [-5.0, +2.5], p=0.593
gemma3_4b/scienceqa	2	2850	20.5	0.618	0.676	0.710	0.712	43.9	+10.5 [+1.6, +18.5], p=0.012
gemma3_4b/visual_cot	2	2910	92.6	0.581	0.634	0.749	0.714	20.8	+0.1 [-1.1, +1.6], p=0.835
qwen3/mathvista	2	2855	29.2	0.581	0.551	0.589	0.663	39.1	+11.5 [+4.1, +18.2], p=0.003
qwen3/mpcc	2	3389	95.8	0.544	0.651	0.674	0.641	20.4	-0.1 [-0.2, +0.2], p=0.596
qwen3/scienceqa	2	2850	13.9	0.472	0.528	0.582	0.604	34.3	+10.1 [+0.0, +21.1], p=0.058
qwen3/visual_cot	2	2901	83.7	0.525	0.745	0.727	0.680	22.4	-0.1 [-1.5, +1.4], p=0.858

Step 4:

Cell	Step	Test CoTs	Fail rate	AUROC length	AUROC text	AUROC independent	AUROC coupled	Error capture @20% (coupled)	Coupled − text capture
gemma3_4b/mathvista	4	2278	38.6	0.570	0.601	0.646	0.631	28.2	-1.3 [-7.9, +6.7], p=0.735
gemma3_4b/scienceqa	4	2535	20.6	0.653	0.687	0.705	0.712	37.7	+2.1 [-7.8, +10.6], p=0.651
gemma3_4b/visual_cot	4	2602	93.0	0.583	0.641	0.764	0.676	21.0	+0.2 [-0.6, +1.1], p=0.698
qwen3/mathvista	4	2120	32.1	0.608	0.612	0.665	0.704	37.5	+4.6 [-0.6, +10.9], p=0.122
qwen3/mpcc	4	3231	95.8	0.539	0.654	0.669	0.626	20.6	+0.1 [-0.2, +0.3], p=0.640
qwen3/scienceqa	4	2528	14.7	0.481	0.516	0.622	0.632	35.3	+10.8 [-3.1, +24.1], p=0.128
qwen3/visual_cot	4	942	88.3	0.450	0.711	0.587	0.510	20.7	-0.5 [-3.0, +2.0], p=0.709

Step 8:

Cell	Step	Test CoTs	Fail rate	AUROC length	AUROC text	AUROC independent	AUROC coupled	Error capture @20% (coupled)	Coupled − text capture
gemma3_4b/mathvista	8	529	43.1	0.534	0.646	0.665	0.628	25.0	-3.5 [-10.7, +2.4], p=0.289
gemma3_4b/scienceqa	8	54	42.6	0.710	0.659	0.634	0.762	43.5	+17.4 [+3.6, +33.3], p=0.042
qwen3/mathvista	8	480	55.2	0.692	0.700	0.734	0.686	26.8	-1.1 [-4.6, +3.1], p=0.567
qwen3/mpcc	8	727	97.9	0.543	0.580	0.710	0.758	20.5	+0.3 [+0.0, +0.7], p=0.239
qwen3/scienceqa	8	323	30.0	0.523	0.555	0.593	0.617	27.8	-2.1 [-12.8, +13.0], p=0.771

Conclusions.

• Prefix process features carry real eventual-failure signal. At step 2, coupled AUROC is 0.59–0.71 and above the length baseline in every cell.
• Coupled triage catches significantly more failures than the text-only chain at a 20% review budget in some cells:
  • Qwen3/MathVista, step 2: +11.5 pp, p=0.003.
  • Gemma/ScienceQA, step 2: +10.5 pp, p=0.012.
  • Gemma/ScienceQA, step 8: +17.4 pp, p=0.042, but on only 54 CoTs.
• Coupling is not consistently better than the independent (uncoupled) chain. The independent arm matches or exceeds coupled AUROC in about half the cells, most clearly on Visual-CoT, so the gain comes mostly from adding visual/attention dynamics rather than from coupling itself.
• In high-failure cells (Visual-CoT and MPCC, ≥83% wrong), review-budget metrics are near their trivial values and uninformative.
• None of these comparisons is multiplicity-corrected; treat them as exploratory until the final Holm-corrected report.

3. Targeted visual repair

Setup: on the first archived target CoT of each test query, at the first checkpoint where coupled risk crosses its validation threshold, re-inject two image crops (25% total area) and continue generation, over 3 continuation seeds. All five arms are paired on the same queries and seeds: keep the original answer, generic full-image recheck, random crops, current-attention crops, and coupled-atlas forecast crops. Untriggered queries keep their original answer and stay in the denominator.

Cell	Queries×seeds	No repair	Full image	Random crop	Attention crop	Atlas crop	Atlas − no repair	Atlas − attention	Atlas − random
gemma3_4b/mathvista	191×3	60.7	59.9	58.6	59.7	59.3	-1.7 [-4.8, +1.5], p=0.356	-0.5 [-1.7, +0.7], p=0.503	+0.6 [-1.0, +2.6], p=0.524
gemma3_4b/scienceqa	190×3	78.9	79.1	80.0	79.6	79.5	+0.3 [-2.8, +3.6], p=0.910	-0.1 [-1.1, +1.0], p=0.871	+0.0 [-2.0, +2.1], p=1.000
qwen3/mathvista	191×3	68.6	68.8	69.6	67.7	67.4	-1.3 [-3.9, +0.9], p=0.328	-0.4 [-2.0, +1.0], p=0.704	-2.4 [-4.6, -0.6], p=0.025
qwen3/scienceqa	190×3	83.7	84.4	83.3	84.4	84.4	+0.4 [-2.0, +3.2], p=0.873	+0.2 [-1.2, +1.8], p=0.898	+1.5 [-0.6, +3.7], p=0.215
qwen3/visual_cot	194×3	16.5	17.2	16.7	16.2	16.2	-0.5 [-1.4, +0.0], p=1.000	+0.0 [+0.0, +0.0], p=1.000	-0.5 [-1.6, +0.0], p=0.510

Conclusions.

• No repair action improves accuracy over keeping the original answer in any completed cell. Atlas − no-repair ranges from −1.7 to +0.4 pp, and every CI covers zero.
• Atlas-forecast crops are no better than current-attention crops. On Qwen3/MathVista they are significantly worse than random crops: −2.4 pp, p=0.025.
• In this archived-prefix setting, re-injecting targeted crops does not rescue failing trajectories. This measures incremental repair from saved prefixes, not online generation savings.

3b. Repair timing

Setup: the crop rule is fixed to attention crops and only the trigger varies. The triggers are coupled risk, text-chain risk, raw token entropy (from native teacher-forced scoring), periodic (step 4) and random. Each trigger is calibrated on validation to intervene on about 20% of queries.

Cell	Coupled	Text	Entropy	Periodic	Random	Coupled − text	Coupled − entropy	Coupled − random
gemma3_4b/mathvista	59.9	59.2	62.7	59.7	61.3	+0.8 [-1.2, +2.9], p=0.602	-2.7 [-6.3, +0.6], p=0.111	-1.7 [-5.2, +1.8], p=0.356
gemma3_4b/scienceqa	78.8	78.2	80.4	78.9	78.8	+0.1 [-2.3, +2.5], p=1.000	-2.3 [-5.5, +0.8], p=0.178	-0.4 [-3.8, +3.0], p=0.882
qwen3/mathvista	67.4	68.4	69.6	68.4	70.2	-1.2 [-3.1, +0.8], p=0.340	-2.4 [-5.1, +0.2], p=0.085	-3.0 [-6.5, +0.6], p=0.099
qwen3/scienceqa	85.3	84.6	84.4	85.8	83.9	+0.5 [-2.6, +3.8], p=0.801	-0.0 [-2.9, +2.9], p=1.000	+0.5 [-2.4, +3.6], p=0.799
qwen3/visual_cot	–	–	16.3	14.6	12.9	–	–	–

Conclusions.

• Coupled-risk timing is not significantly better than text, entropy, periodic or random timing in any cell; its point estimates range from −3.0 to +0.8 pp.
• The raw token-entropy trigger is at least as accurate as coupled in all four cells and ahead by 2.3–2.7 pp in three, though no difference is significant (p ≥ 0.085).
• For Qwen3/Visual-CoT, the coupled and text triggers are unavailable because their step-8 heads could not be fitted.

4. Online compute allocation

Setup: real online generation under fixed time budgets (4×, 8× and 16× the profiled single-CoT cost). It compares fixed sampling, random/entropy restart and text/independent/population/coupled continue-or-restart controllers, with the atlas pilot cost charged to atlas policies.

Status: no results yet. Train-device profiling had finished, and validation threshold tuning had started, for 5 cells (Qwen3 MathVista/ScienceQA/Visual-CoT and Gemma MathVista/ScienceQA) when the previous server stopped. No test-split allocation has completed, so no conclusion can be drawn.

5. Verified-CoT selection and LoRA SFT

Setup: among independently verified correct CoTs, select supervision data with random, text, independent, population or coupled facility-location coverage under an equal token budget. Then train a language-only LoRA (rank 8, 1M supervised tokens per arm and seed) and evaluate on held-out queries.

Status: only the untrained base model has been evaluated. Curation and adapter training had not started (both wait on verified labels and heads), so no selection arm can be compared yet. Base held-out accuracy (3 sampling seeds):

Cell	Arm	Seed	Held-out queries	Accuracy
gemma3_4b/mpcc	base	seed_101	229	7.0
gemma3_4b/mpcc	base	seed_202	229	9.2
gemma3_4b/mpcc	base	seed_303	229	9.6
qwen3/mathvista	base	seed_101	191	66.0
qwen3/mathvista	base	seed_202	191	66.0
qwen3/mathvista	base	seed_303	191	67.5
qwen3/mpcc	base	seed_101	229	4.8
qwen3/mpcc	base	seed_202	229	3.9
qwen3/mpcc	base	seed_303	229	3.5
qwen3/open_r1	base	seed_101	31	0.0
qwen3/open_r1	base	seed_202	29	0.0
qwen3/open_r1	base	seed_303	32	0.0
qwen3/scienceqa	base	seed_101	190	89.5
qwen3/scienceqa	base	seed_202	190	88.4
qwen3/scienceqa	base	seed_303	190	87.9
qwen3/visual_cot	base	seed_101	194	17.0
qwen3/visual_cot	base	seed_202	194	17.5
qwen3/visual_cot	base	seed_303	194	16.0

Conclusions. None yet for data selection. The base evaluations fix the reference each adapter will be compared against.

• They show near-floor accuracy on MPCC (3–10%) and Visual-CoT (16–18%).
• They show 0% graded accuracy on Open-R1, on only 29–32 graded queries. That is consistent with the unavailable Open-R1 correctness heads, and Open-R1 grading should be audited before any Open-R1 claim.

Overall (interim)

• Across the completed applications, coupled text–visual Markov features help most for early failure triage. There they improve on text-only and length baselines in several cells, although the explicit coupling term does not consistently beat the uncoupled visual chain.
• They give no measurable benefit for answer selection, visual repair, or repair timing in the available cells.
• Allocation and SFT conclusions are pending.