CV-OPD

When the evidence disappears, the answer should reflect it.

Figure 1

01The shortcutCorrect predictions can persist without the answer-relevant pixels.

02The interventionContrast relevant and irrelevant crops under the same student prefix.

03The outcomeMore evidence-sensitive responses and higher benchmark accuracy.

01 / Motivation

Answer matching leaves
a visual shortcut open.

Vision-OPD gives a crop-conditioned teacher privileged visual evidence. Matching its output distribution does not, by itself, require the full-image student to find that evidence.

Figure 2: accuracy under black-mask and Gaussian-noise corruption, and Vision-OPD attention examples for a lock's color and a woman's pose. — Figure 2 · Corrupted-input evaluation and attention analysis, paper §2.2.

Intervention

Remove the relevant evidence.

Replace the answer-related region with a black mask or Gaussian noise.

Observation

The expected answer can remain.

The model may still predict the original answer even though its visual basis is missing.

Implication

Accuracy alone cannot establish visual grounding. We also need to test whether predictions depend on the available evidence.

02 / Method

Same prefix.
Different visual evidence.

CV-OPD adds an unrelated crop-conditioned teacher. The student learns to move toward the positive teacher and away from the negative teacher, at selected token positions.

Figure 3 / Training pipeline

What is the color of the little girl's shirt?

Visual input

Policy

Token distributions

Supervision

Original full image from Figure 3, containing the little girl.

V Full image

Studentπ_θQwen3.5-4B

Student rollout π

It'sablackshirtEOS

Illustrative error token from Figure 3

Figure 3 positive crop containing the girl's pink shirt.

R⁺ Relevant crop

Positive teacherτ⁺Shared parameters ϕ

Positive distribution τ⁺

TheapinkshirtEOS

Correction supported by the shirt crop

R⁻ Unrelated crop

Negative teacherτ⁻Shared parameters ϕ

Negative distribution τ⁻

TheaceramicgroundEOS

Unrelated content from the ground crop

Both teachers reuse C_i,t = (Q, A_<t)Same student prefix · only the crop changes

JSDPositive alignment

D_JS(π, τ⁺)

Retained samples receive distillation.

DAGToken selection

d_i,t > d̄_B

d = D_JS(student, positive teacher)

CDLMargin constraint

[μ + d⁺ − d⁻]₊

τ⁺πτ⁻

← toward τ⁺away from τ⁻ →

Sample gate Retain the example if at least one of its rollouts is correct.

Update student θ→EMA teacher update ϕ ← (1 − η)ϕ + ηθ

01 / 05

The student starts with the full image.

It samples its own responses. Keep a training example when at least one rollout is correct; incorrect rollouts of that retained example are still available for correction.

Eq. 6, 13

Controlled comparison. R⁻ comes from the same image, matches R⁺ in size, and satisfies IoU(R⁺, R⁻) ≤ δ. One negative crop is shared across a sample's rollouts.

Distribution-level supervision. The words above are the paper's illustrative token outputs, not three independently generated answers. Both teacher distributions are evaluated on the student's prefix.

Original Figure 3

The training objective

Align. Select. Contrast.

L = L_OPD + λL_CDL^gated

JSD aligns with the positive teacher. DAG selects positions for CDL. The contrastive loss becomes zero once d⁻ − d⁺ ≥ μ.

03 / Results

Better grounding.
Measurable gains.

Paper-reported evaluation on six fine-grained visual benchmarks. The comparison below uses the same Qwen3.5-4B backbone and the paper's non-thinking evaluation setting.

78.22%

Average accuracy across six benchmarks

CV-OPD · Table 1

+2.30 pp

Average gain over Vision-OPD

78.22 − 75.92 · Table 1

+2.60 pp

Gain on ZoomBench

62.01 − 59.41 · Table 1

Table 1 · Selected Qwen3.5-4B baselines · Accuracy (%)
Method	V*	Zoom	HR 4K	HR 8K	MME EN	MME CN	Avg6
Base model	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
VAD	91.62	60.12	85.13	83.38	76.98	71.44	78.11
CV-OPD Ours	91.62	62.01	84.38	83.38	76.99	70.93	78.22

Avg6 is the unweighted mean. CV-OPD improves on Vision-OPD across all six benchmarks; it does not outperform every comparator on every individual benchmark. On HR-Bench 4K, the base model remains higher.

Table 2 / Component ablation

Both components contribute.

Removing either the contrastive objective or the gate reduces average accuracy.

Neither

75.92

DAG only

76.81

CDL only

76.88

CDL + DAG

78.22

Avg6 (%)0100

04 / Visual evidence

Look at the object.
Respect what is missing.

Attention and response examples make the distinction concrete: identifying a lock's color is different from recognizing when the relevant pixels are unavailable.

Figure 4: Vision-OPD predicts the mailbox's dominant red; CV-OPD attends to the lock and predicts silver. — Figure 4 · Appendix A.4, p. 15

Evidence present

The lock is silver.
The mailbox is red.

Vision-OPD follows the mailbox's dominant red color. CV-OPD focuses on the lock itself and identifies its metallic silver color.

The relevant object, not the surrounding color, should determine the answer.

A qualitative example of attention and response behavior, not a causal proof of grounding.

Implementation

Contrastive Vision On-Policy Distillation

Code

CV-OPD: Contrastive Vision On-Policy Distillation