CV-OPD: Adding Visual Evidence Constraints to On-Policy Self-Distillation¶

Paper information¶

Field	Value
Paper	arXiv 2610.13264
Affiliations	Xiamen University / Moxin Technology
First public release	2026-10-09 (arXiv v1)
Author code	Yes: authors' code.
Adapter	`cv-opd`
Local implementation	`src/auto_research/reproductions/cv_opd/`

The original v1 submission was 2026-10-09 09:15:34 UTC. The Apache-2.0 author repository is pinned at e9452a00ea50ddb776645b9acc49275e0180955e. The local equations were implemented independently, without running its installation script. Its paper_assets/paper.pdf is Vision-OPD 2605.18740v4, not this paper. The official v1 PDF confirms Haowei Wang's Xiamen University and Moxin Technology affiliations and Hao Lin's Xiamen University affiliation. Algorithm details follow the v1 full text and Appendix A.3.

source-verification.json records the official PDF's size, SHA256, first-page affiliations, and the upstream PDF discrepancy. Downloaded PDFs and temporary renders are not committed.

Original paper¶

Background and contribution¶

Crop-to-full-image OPD can teach answer distributions without teaching reliance on the right image regions. Vision-OPD remained above chance after answer-related regions were removed. CV-OPD introduces a shared-parameter negative teacher using a same-image, same-size, low-IoU crop, adding attraction to the positive teacher and repulsion from the negative teacher to JSD distillation.

Three constraints matter: the positive and negative teachers differ only in their crop and see the same student prefix; CDL uses a bounded-JSD margin; and DAG uses a strict greater-than-mean token threshold over the entire retained batch. A sample is retained only if at least one rollout is correct, and then all its rollouts are retained.

Architecture¶

Key figure from the paper¶

Figure 3 in arXiv v1 shows the full-image student, shared-parameter positive and negative crop teachers, JSD, discrepancy-aware gating, and contrastive divergence loss. The original PNG is uncropped and unredrawn; the filename is fig2.png, but the paper labels it Figure 3. The Qwen3.5-4B in the diagram is the authors' model, not the local SmolVLM diagnostic. See provenance and SHA256; the misplaced Vision-OPD PDF was not used. Copyright belongs to the authors.

Local implementation diagram:

flowchart TD
  A[Full image and question] --> B[Student generates M rollouts]
  B --> C[External correctness verifier]
  C --> D[Keep every rollout if at least one is correct]
  A --> E[Actual positive-region bounding box]
  E --> F[Same-size low-IoU negative crop]
  D --> G[Full-image student, same prefix]
  E --> H[EMA teacher on positive crop]
  F --> I[Same EMA teacher on negative crop]
  G --> J[Top-100 plus tail JSD]
  H --> J
  G --> K[Full-vocabulary JSD-margin CDL]
  H --> K
  I --> K
  K --> L[Above-batch-mean token gate]
  J --> M[Update student parameters]
  L --> M
  M --> N[EMA teacher update]

Core equations¶

For student and shared positive/negative teacher distributions \(p_t,\tau_t^+,\tau_t^-\) and \(m=(p+\tau)/2\):

\[ D_{JS}(p,\tau)=\tfrac12 KL(p\|m)+\tfrac12 KL(\tau\|m). \]

\[ d_t^\pm=D_{JS}(p_t,\tau_t^\pm),\qquad \ell_t^{CDL}=[\mu+d_t^+-d_t^-]_+,\quad 0<\mu<\log2. \]

\[ g_t=\mathbf1[d_t^+>\bar d_\mathcal B],\qquad b_x=\mathbf1[\sum_{i=1}^M r_i>0]. \]

Equation 14 writes per-rollout normalization. The local implementation follows Appendix A.3 Algorithm 1: main JSD is computed over the student-aligned top 100 and a complete tail bucket, weighted by \(1/T_i\) per rollout and renormalized. CDL uses full-vocabulary JSD averaged over gated tokens. Margin and CDL weight are both 0.1, and EMA update rate is 0.05. The teachers, gate, and generation are detached. The sample gate does not mean training on correct rollouts only.

Paper results¶

The authors used Qwen3.5-4B, 6.2K visual-reasoning examples, 8 rollouts of up to 1024 tokens per example, and 8 A100s.

Paper model	V* accuracy %	ZoomBench %	HR-4K %	HR-8K %	MME-RW EN %	CN %	Avg6 %
Qwen3.5-4B	82.20	48.40	86.38	82.00	63.94	63.80	71.12
Vision-OPD	89.53	59.41	81.75	80.12	74.51	70.17	75.92
CV-OPD	91.62	62.01	84.38	83.38	76.99	70.93	78.22

Avg6 improves by 2.30 percentage points, not 2.30% relative. CV-OPD still trails the undistilled baseline on HR-4K. The reported roughly 10% training-time overhead against Vision-OPD is not a local measurement. No online A/B test was reported.

Local reproduction¶

This core_mechanism / L1 diagnostic updates a real VLM, not an image-classification scoring proxy. GPU validation requires the committed sanitized receipt; CPU tests are not a substitute.

Compare the same checkpoint's validation accuracy before and after CV-OPD. A three-seed control sets CDL weight to zero, but still runs both crops, DAG, and teacher forwards. It is not original Vision-OPD or a matched-realized-compute capability comparison. Relative capability gains are not reported.

Data protocol¶

JSONL requires image, a uniform question, actual category answer, pixel bbox=[x0,y0,x1,y1], split, and verified=true. Train on a deterministic subset of a public train source; source images must not cross splits. Existing COCO-Caption2017 val captions lack boxes and cannot silently substitute for training inputs.

The diagnostic selects real COCO train2017 main objects from person/cat/dog/car/bus/bird/horse/bicycle, with 2 train, 1 validation, and 1 test image per category (32 total). Validation and test are local holdouts derived from train2017, not official test. The shared question is Name the main object in this picture. Reply with one COCO category name and no other words. It supplies neither the real category nor the box to the full-image student. Constant-label training is rejected. The verifier normalizes case, punctuation, whitespace, and initial a/an, then requires an exact category: not a cat, cat or dog, and cats do not match cat.

verify_offline_manifest checks actual checkpoint and dataset SHA256s; the dataset manifest covers JSONL and each referenced image. This binds labels and boxes to source files, not just CLI claims. Integrity checks do not expose held-out content to the model.

The student sees the full image and each teacher sees its crop, all with the same question and no answer. Only the external verifier reads the answer. If no same-size low-IoU negative crop exists, record a failure and skip. If all 8 rollouts are wrong, make exactly zero updates; never insert a gold completion. Rejected rollouts still count toward token cost. Low IoU ensures geometric separation, not absence of another same-class object in the negative crop.

Entry point¶

PYTHONPATH=src python scripts/run_cv_opd.py \
  --checkpoint /path/to/smolvlm256 \
  --checkpoint-id HuggingFaceTB/SmolVLM-256M-Instruct \
  --checkpoint-revision 7e3e67edbbed1bf9888184d9df282b700a323964 \
  --records /path/to/coco-train-boxes.jsonl \
  --dataset-id COCO2017/train2017 --dataset-revision <official-annotation-archive-sha256> \
  --device cuda --steps 16 --rollouts 8 --max-new-tokens 16 \
  --seed 42 --output runs/cv-opd-seed42.json

The adapter reads an explicit CLI-parameter map from dataset_dir/cv-opd.json, without seed/output. Missing configuration fails instead of falling back to a fixture.

CPU contracts and limits¶

cpu-contracts.json records 7 rerun behavioral tests covering box size/IoU, JSD/top-K tail, sample and token gates, teacher stop-gradient, zero gradient without correct samples, strict multiclass verification, label isolation, and image/split hashes. Tests fix seed 42, with seed 2 inside dedicated crop/gate cases. These are unit cases, not multi-seed capability runs or GPU training.

The small path uses SmolVLM-256M and rank-8 language q/v LoRA, not full Qwen3.5-4B training. Public main-object classification with an exact-category verifier differs from the six fine-grained benchmarks. Changing external gold affects correctness scoring only, not image, question, or budget supplied to the policy. No Evolve integration.

A100 measurements, 2026-10-14¶

The sanitized GPU receipt and six-run measurements bind executable snapshot 4496283ac8e37c070b8f71b20af6138526cbba15, six actual commands, 24 raw file hashes, and five core source hashes. The receipt filename retains 20261013; completion was 2026-10-13 17:21:37 UTC (October 14 in China). Both arms successfully updated for all three seeds, without test-based reselection.

Pinned checkpoint: HuggingFaceTB/SmolVLM-256M-Instruct@7e3e67edbbed1bf9888184d9df282b700a323964. COCO official annotation archive SHA256: 113a836d90195ee1f884e704da6304dfaaecff1f023f49b6ca93c4aaae470268. Derived records SHA256: f23de926af9f803f1c5568eae1bb150b2e55b919fce82a66167e961e764acb00. The receipt includes complete checkpoint and 32-image hashes, not weights, images, or raw logs.

The fixed budget was 16 sample attempts, batch 1, 8 rollouts per image, and at most 16 tokens. CDL weight was 0.1 vs. 0; other settings matched. All accuracies are local L1 correct/8, not paper benchmark scores.

Seed	Main val	CDL-off val	Main test	CDL-off test	Updates per arm	Main/control sampled tokens	Main/control seconds
42	2/8	2/8	0/8	0/8	8	520 / 520	147.99 / 146.34
43	3/8	4/8	1/8	2/8	9	512 / 491	156.34 / 152.62
44	2/8	2/8	0/8	1/8	10	538 / 513	168.57 / 165.13

Test mean ± sample SD was 4.17% ± 7.22% for main and 12.50% ± 12.50% for control; paired difference was −8.33 ± 7.22 percentage points. This does not show a CDL advantage. Initial validation was 3/8, 3/8, and 2/8. Initial test was not measured, so no gain over initial test can be inferred.

Each arm and seed sampled 128 rollouts, counting rejected tokens. Parameter L2 changes were nonzero: 0.003335/0.003482/0.003787 for main, and 0.003192/0.003484/0.003722 for control. Update counts matched within seeds, but realized tokens, gradient trajectories, and wall-clock costs did not. Crop failures were zero and maximum IoU ≤ 0.1. The control still computes unweighted CDL; a nonzero CDL value does not mean it enters the loss.

The evidence establishes executable image rollouts, sample retention, crop teachers, and CUDA updates. Three seeds do not upgrade L1 to capability evidence or establish comparison with Vision-OPD or the reported 10% overhead. scripts/curate_oct14_cv_opd.py recalculates from the pinned source bundle without modifying it. Fixture tests reject wrong budgets, weights, zero updates, or source/public-input identities.

Code map¶

Component	Local implementation
Actual positive/negative crops, JSD, CDL, DAG, EMA	cv_opd.py
Image checkpoint rollouts and updates	run_cv_opd.py
Behavioral tests	test_cv_opd.py