acceptodds
Under review as a conference paper at ICLR 2027

Visual Contrastive Self-Distillation

Abstract

On-policy self-distillation (OPSD) removes the external teacher required by on-policy distillation (OPD), but still needs asymmetric information between teacher and student so that the self-teacher provides a stronger learning signal. Existing methods create this asymmetry through privileged answers or visual evidence. We ask whether both can be removed, yielding a simpler form of OPSD driven purely by input conditioning. We propose Visual Contrastive Self-Distillation, namely VCSD, which converts image-content removal into an on-policy self-distillation signal. At each student-generated response prefix, the EMA teacher produces two next-token distributions under the same prompt and prefix: one conditioned on the original image and the other on a content-erased control. Their token-wise log-probability difference highlights candidates whose likelihood is specifically increased by the instance-level visual content. We use this contrast to sharpen the teacher's original-image distribution within its plausible support, and distill the resulting full-distribution target into the student. On ViRL39K, VCSD consistently outperforms matched OPSD across Qwen3-VL and Qwen3.5 models. It improves the seven-benchmark aggregate from 72.51% to 76.26% on Qwen3-VL-8B and from 74.97% to 79.24% on Qwen3.5-9B, with gains at all six evaluated scales. VCSD requires no external teacher, privileged answers, visual evidence signals, reasoning traces, or additional inference-time cost.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
VLM, self-evolving

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_PRMFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper proposes Visual Contrastive Self-Distillation (VCSD), an on-policy self-distillation method for vision-language models that requires no external teacher, privileged answers, or additional visual-evidence annotations. At each student-generated response prefix, the same EMA teacher predicts next tokens under the original image and a same-size black-image control. Their log-probability difference reweights the original-image distribution within a plausibility support, and the student learns the resulting target through forward KL. Only the updated student is used at inference. Experiments on ViRL39K cover six Qwen3-VL/Qwen3.5 configurations and seven benchmarks, with additional training on Vision-OPD6K and ablations of the target construction and teacher updates. The seven-benchmark mean improves from 72.51 to 76.26 for Qwen3-VL-8B and from 74.97 to 79.24 for Qwen3.5-9B relative to the corresponding base models.

Critical strengths

Teacher conditioning on answers or extra visual evidence can introduce preferences that depend on information unavailable to the student. VCSD constructs its target from paired visual conditions of the same teacher, anchored by the original-image prediction. It reduces supervision requirements by avoiding reference-answer rewards and additional evidence annotations, while retaining a single student model at inference.

Across all six model configurations, Table 1 reports aggregate gains of 1.86–4.80 percentage points over the base models and 1.76–4.61 points over answer-hint OPSD. The no-contrast, sharpening-only, inference-only, frozen-teacher, and alternative-control-image experiments examine the roles of the main components. On Qwen3-VL-2B, forward KL reaches 67.49 versus 64.43 for reverse KL, while the frozen teacher reaches 65.03 versus 67.49 for the full method. Table 6 also reports three independent training seeds at this scale.

Critical weaknesses

The core idea of this paper is to extract learning signals related to instance-specific visual information by comparing the model's predictions under the original image and content-erased/control image conditions, and then use these signals for self-distillation. However, using changes in model predictions under different visual conditions to measure visual dependency is not first proposed in this paper. For example, prior work such as VPPO in the RL setting and the more recent VA-OPD adopt similar ideas, which weakens the novelty of this work. In comparison, the main distinction of VCSD appears to lie in using visual contrast to directly reconstruct a plausible full-distribution teacher target, rather than performing advantage shaping, loss reweighting, or token selection. Therefore, I believe the paper should more clearly articulate its methodological novelty and distinguish its contributions from existing visual-dependency and visual-contrast approaches.

The main training experiments in this paper are conducted on ViRL39K, a dataset containing a substantial amount of data involving mathematical, geometric, and spatial reasoning, as well as chart understanding, all of which are closely related to multimodal reasoning. However, the current evaluation benchmarks are primarily focused on visual perception, fine-grained visual understanding, and hallucination. Among the seven main benchmarks, MathVista is the only one that directly evaluates visual mathematical reasoning. Given the close connection between ViRL39K and multimodal reasoning capabilities, I suggest including additional multimodal reasoning benchmarks, such as MMMU and MMMU-Pro, to provide a more comprehensive evaluation of the proposed method.

The baseline comparisons are also insufficient. Regarding methods related to OPSD, the main results table only includes a comparison with standard OPSD. Considering that several multimodal OPD/OPSD methods have recently been proposed, I suggest including at least a few closely related and competitive baselines under the same backbone, training data, and optimization budget. This would provide a more convincing demonstration of the advantages of the proposed method over existing approaches.

Questions for authors

The paper presents thorough experiments and demonstrates substantial performance improvements. However, I would like the authors to more clearly articulate the methodological novelty of their approach relative to existing visual-dependency and visual-contrast methods, particularly explaining why this work goes beyond simply transferring existing visual-contrast approaches from other domains to OPSD. In addition, I encourage the authors to include more evaluations on benchmarks closely related to the training data, such as MMMU, and to provide additional comparisons with existing OPSD-related methods.

If all these concerns are adequately addressed, I would consider raising my score to 3/4.

4 comments

We thank the reviewer for the constructive feedback and suggestions regarding the methodological novelty of VCSD, evaluation on additional benchmarks, and comparisons with related OPD-family methods. We have conducted several new experiments to address these concerns and incorporated the results into Appendix B of the revised manuscript.

W1. Methodological novelty relative to visual-dependency methods (VPPO, VA-OPD)

We appreciate the opportunity to clarify the methodological distinction between VCSD and prior methods that exploit visual dependency signals.

The central distinction is how the visual contrast is used to construct the learning objective.

Prior visual-dependency methods, such as VPPO and VA-OPD, generally use visual contrast to modulate an existing learning signal, for example through advantage shaping, loss reweighting, or token selection based on an underlying reward or supervision target.

In contrast, VCSD uses the visual contrast to construct the supervision target itself. Specifically, the contrast determines a full-distribution self-distillation target, rather than serving as an auxiliary weight on an existing reward- or label-based objective.

Consequently, VCSD does not require reward labels, reference answers, or external evidence annotations to construct its training supervision.

We agree that sharing a visual contrast signal does not, by itself, establish a substantive methodological difference. To investigate whether the distinction in its utilization matters empirically, we conducted an additional controlled comparison.

Controlled study: Different usages of the same visual contrast signal

We evaluated the same visual contrast signal Δ under five different applications, using matched backbone, training data, and experimental settings (Qwen3-VL-2B, seven-benchmark average).

Usage of visual contrast ΔRepresentative approach7-benchmark average
Decoding-time adjustment only (no training)Contrastive decoding62.49
Binary trajectory selection (0/1 gate)V-Zero-style gating63.48
Grouped token-level loss weightingVA-OPD-style reweighting64.40
Continuous token-level loss weightingVPPO-style reweighting64.69
No contrast (reward-based reference)GRPO65.59
Full-distribution target constructionVCSD (ours)67.49

The results reveal a consistent difference between the evaluated ways of utilizing the same visual contrast.

  1. Simply incorporating the contrast signal into an existing objective does not reproduce VCSD's improvement. The trajectory-gating and token-reweighting variants achieve 63.48, 64.40, and 64.69, respectively, all below the GRPO reference of 65.59.

  2. Constructing a full-distribution distillation target yields the strongest result among the evaluated approaches. VCSD achieves 67.49, exceeding both GRPO and the alternative contrast-utilization variants.

These results suggest that the benefit of VCSD is not attributable solely to the availability of a visual contrast signal; the way this signal is incorporated into the learning objective is an important factor.

We emphasize that these experiments compare representative implementations of different contrast-utilization mechanisms under controlled settings, rather than claiming to reproduce every detail of the original methods.

We have added this study to Appendix B to better substantiate the methodological distinction.

W2. Evaluation on MMMU and MMMU-Pro

We thank the reviewer for suggesting additional evaluation on MMMU and MMMU-Pro, which provide complementary assessments of multimodal reasoning performance.

Following this suggestion, we evaluated all five methods (Base, SFT, GRPO, OPSD, and VCSD) on MMMU (validation) and MMMU-Pro using Qwen3.5-4B and Qwen3.5-9B.

The evaluations follow the same decoding and evaluation settings used in the main paper, where applicable.

MMMU (validation)

BackboneBaseSFTGRPOOPSDVCSD (ours)
Qwen3.5-4B68.7870.2269.0070.5670.89
Qwen3.5-9B69.0072.2272.1167.0073.78

MMMU-Pro

BackboneBaseSFTGRPOOPSDVCSD (ours)
Qwen3.5-4B58.5559.1957.6955.0360.46
Qwen3.5-9B60.2361.7962.2555.8465.20

VCSD achieves the highest accuracy among the evaluated methods across all four backbone-benchmark combinations.

At the 9B scale, VCSD obtains 73.78 on MMMU and 65.20 on MMMU-Pro, compared with 72.11 and 62.25 for GRPO, corresponding to improvements of +1.67 and +2.95 points, respectively.

At the 4B scale, VCSD also achieves the best results on both benchmarks, with 70.89 on MMMU and 60.46 on MMMU-Pro.

The absolute improvements over the base model are generally smaller than those observed on the perception-oriented benchmarks in the main paper. This is consistent with the nature of the ViRL39K training data and VCSD's emphasis on visually grounded prediction rather than explicit reasoning supervision.

Nevertheless, the consistent ranking across both additional benchmarks provides evidence that VCSD's improvements are not restricted to the original evaluation suite and can extend to more reasoning-oriented multimodal tasks.

Both result tables have been included in Appendix B of the revised manuscript.

W3. Comparisons with additional OPD-family baselines

We agree with the reviewer that comparisons with closely related OPD-family approaches are important for establishing the empirical contribution of VCSD.

We would like to clarify the existing comparisons and report an additional baseline evaluated in response to this concern.

1. Answer-hint OPSD (Table 1)

The main experiments already include the standard answer-hint OPSD baseline across all six evaluated backbone configurations.

This baseline uses privileged reference answers to condition the teacher, whereas VCSD constructs its supervision using only image-question pairs.

All methods are trained on the same ViRL39K dataset with the same 90-step optimization budget, and method-specific hyperparameters are provided in Appendix A.

2. Vision-OPD (Table 2)

We also compare VCSD with Vision-OPD, a recent multimodal OPD approach that constructs visual-evidence supervision using bounding-box-annotated cropped teacher views.

To ensure a controlled comparison, we reproduced Vision-OPD using the same training data and the same number of optimization steps across three backbone configurations.

BackboneVision-OPD (same-data reproduction)VCSD (ours)Improvement
Qwen3-VL-4B70.3473.82+3.48
Qwen3.5-4B74.2376.96+2.73
Qwen3.5-9B77.1579.26+2.11

VCSD outperforms the reproduced Vision-OPD baseline across all three configurations, with improvements ranging from +2.11 to +3.48 points.

An important difference is that VCSD does not require the bounding-box annotations used by Vision-OPD.

For completeness, Table 2 also reports results from the officially released Vision-OPD checkpoints under our evaluation protocol, together with a joint variant combining the two supervision signals.

3. Additional VA-OPD-style baseline

Following the reviewer's suggestion, we additionally evaluated a VA-OPD-style baseline under matched backbone, data, and optimization budget.

Specifically, we adapted the grouped token-reweighting mechanism to the self-distillation setting using a degraded visual control view.

We emphasize that this is a VA-OPD-style adaptation, rather than an exact reproduction of the original method, which uses an external 8B teacher. This adaptation allows us to examine the effect of contrast-based loss reweighting under a controlled training setup.

Method (Qwen3-VL-2B)7-benchmark average
Answer-hint OPSD (privileged reference answers)63.63
VA-OPD-style grouped reweighting (self-distillation adaptation)64.40
GRPO (reward-based reference)65.59
VCSD (ours; image-question pairs only)67.49

Under these matched settings, VCSD achieves the highest accuracy, outperforming both the answer-hint OPSD baseline and the VA-OPD-style reweighting adaptation.

Together with the Vision-OPD comparisons in Table 2 and the contrast-utilization study in W1, these results further support the effectiveness of VCSD's full-distribution target construction over the evaluated alternatives.

We hope these additional comparisons and clarifications address the reviewer's concerns regarding methodological novelty and baseline coverage.

As the discussion period is coming to a close, we wanted to briefly follow up on our rebuttal. If any of your concerns remain, we’d be happy to clarify. If our response addresses your main concerns, we’d appreciate it if you would consider revisiting your rating. Thank you!

Reviewer_BXFzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposed Visual Contrastive Self-Distillation (VCSD), a novel method that improves a VLM with using only images and questions without external supervision. VCSD trains a student model repeatedly by comparing the predicted results of an EMA teacher on the original image and a black image; then it constructs visually grounded training objectives based on the differences. The learner learns from the targets and the teacher is updated via EMA. The experiments and ablation studies fully support the proposed method and show improvements on the Qwen3.5 base models range from 2.8% to +4.3% points.

Critical strengths

(1) the self-distillation in VCSD requires less supervision. It compares the same teacher's predictions under the original image and a content-erased control, and keeping the prompt and response prefix same. by this way, it creates informative train targets, so the reference answers, annotated regions, or an outside instructor are not needed. The typical inference process remains unchanged because this contrast is distilled into the student. This approach is the main contribution: it provides a practical way to enhance models through image-question pairings and makes visually informed self-distillation accessible when richer supervision is not available.

(2) The experiments and ablations are enough to support the proposed VCSD. It comes out on top of the seven-benchmark average in all six Qwen3-VL/Qwen3.5 configurations tested, beating SFT, GRPO, and answer-hint OPSD. the Figure3b shows that removing the contrast while retaining self-distillation and plausibility filtering reduces Qwen3-VL-2B’s accuracy. This supports the claim of the paper.

Critical weaknesses

(1) The method makes a strong assumption that the difference of predictions with original and masked image is representative of visual grounding. But more image sensitivity does not necessarily mean better visual understanding. A more systematic discussion is needed: whether the contrastive signal always emphasizes correct visual evidence rather than reinforcing wrong predictions or misleading connections.

(2) The evidence of computational efficiency and generalizability is limited. The method adds extra forward passes for the teacher, however the paper does not offer a compute-matched comparison to existing methods in detail. The studies are limited to the Qwen model family, the usefulness across alternative VLM designs is unclear. Also, table 3 show that introducing the original-image anchor is not consistently a win.

Questions for authors

Please see the weakness above.

3 comments

We thank the reviewer for the detailed feedback and thoughtful concerns regarding the interpretation of visual contrast, computational efficiency, generalization across model families, and the role of the original-image anchor. We have conducted additional experiments and analyses to further investigate these questions.

W1. Visual sensitivity versus visual understanding

We agree that image sensitivity alone does not imply correct visual understanding. We therefore evaluated whether visual contrast distinguishes correct from incorrect rollouts of the same prompt, using training rewards to label correctness. We measure the gap in top-1 token agreement between correct and incorrect rollouts on length-matched samples (within 15%).

Correctness alignment. The contrast direction shows substantially stronger correctness discrimination than the original-image teacher:

Token scopeOriginal teacherContrast directionImprovement
All tokens+0.93 pp+2.63 pp+1.70 pp (p = 0.0047)
Reasoning tokens+0.47 pp+2.17 pp+1.71 pp (p = 0.0047)

Increasing the contrast weight α further strengthens discrimination, from +0.00 pp at α = 0 to +0.20 pp at α = 1, +0.97 pp at α = 4, and +1.70 pp in the pure-contrast limit. This supports the contribution of visual contrast beyond the original teacher, although α = 1 achieves the best downstream accuracy (Fig. 3b).

Mitigating incorrect reinforcement. We also measured the contrast-induced log-probability adjustment difference between correct and incorrect rollouts:

TeacherAll tokensReasoning tokens
Initial−0.009 nats−0.022 nats
Converged+0.055 nats+0.045 nats

Negative values indicate greater relative reinforcement of incorrect rollouts. The β-plausibility constraint limits unreliable adjustments, while the evolving EMA teacher shifts the relative reinforcement toward correct rollouts.

Together with downstream accuracy and hallucination-related improvements across six model scales, these results support improved visually grounded prediction rather than increased image sensitivity alone. Full analyses are provided in Appendix B.

W2. Compute matching, backbone diversity, and the original-image anchor

We thank the reviewer for highlighting these three important issues. We address them separately below.

1. Compute-matched comparison

We measured steady-state per-step training cost for all four methods on identical hardware (Qwen3-VL-2B, 8 GPUs), averaging steps 5–10 to exclude warm-up effects.

MethodTime / step (s)Relative to GRPOPeak memory (GB)
SFT22.90.52×132
GRPO44.41.00×141
OPSD50.01.13×146
VCSD (ours)65.41.47×214 (1.51×)

VCSD incurs 1.47× the per-step training cost of GRPO, primarily due to additional teacher forward passes and full-vocabulary target construction.

To account for this difference, we extended GRPO training beyond the original 90-step budget. GRPO reaches 65.46 at step 100 and 65.70 at step 180, compared with 67.49 for VCSD at step 90. The latter GRPO checkpoint uses approximately 1.36× the total training compute of VCSD.

Thus, VCSD maintains its accuracy advantage even when GRPO is allocated greater total training compute.

We also emphasize that VCSD introduces no additional inference-time computation, as only the student model is used for deployment.

The complete computational profiling and additional training results are provided in Appendix B. 2. Generalization beyond the Qwen family

We agree that evaluation on non-Qwen backbones would provide stronger evidence of cross-family generalization. We therefore ported VCSD unchanged (the exact Qwen recipe, zero hyperparameter adaptation) to two non-Qwen backbones, InternVL3-2B and Gemma-3-4B, and report the results transparently: training is healthy on both (stable loss, no divergence), but under this zero-adaptation port the gains do not transfer, with the seven-benchmark average remaining within ±0.2 points of the respective base models (−0.04 and +0.16).

  • This is an open problem of the method class, not a VCSD-specific limitation. To our knowledge, no published multimodal OPD/OPSD method reports results on any non-Qwen backbone, including the methods discussed in the reviews. Notably, in the text-only setting, self-distillation has been shown to transfer across families (e.g., Gemma and Llama), suggesting the gap lies in multimodal-specific components rather than in the self-distillation principle itself.
  • The result is not explained by a saturated base. Gemma's base score is more than 10 points below InternVL's, yet shows the same pattern.

From a methodological perspective, VCSD's construction remains architecture-agnostic: it requires only the teacher's token distributions under two visual conditions, with no dependence on a particular vision encoder, resolution scheme, attention layout, or special-token design. These two ports were full engineering efforts rather than quick trials: each required re-implementing the vision-input path and correcting family-specific tokenizer and template components, with training verified healthy end to end. We report the results transparently, and we are actively extending VCSD to additional backbones, with cross-family adaptation as our immediate next step.

3. Role of the original-image anchor

We appreciate the reviewer pointing out that the original-image anchor does not consistently improve aggregate benchmark accuracy in Table 3.

We have clarified that the primary purpose of the anchor is training stability rather than direct accuracy improvement. Its role is to retain a training signal from the original-image distribution and reduce undesirable generation behavior.

To examine this more directly, we compared paired runs differing only in the anchor coefficient.

Without the anchor, the fraction of training rollouts that exhaust the generation budget without producing an answer increases from approximately 33% to 60% at step 90, with a similar pattern observed across both evaluated seeds (reported permutation test: p = 0.0083).

In contrast, aggregate benchmark accuracy remains comparable with and without the anchor.

These results support a more specific interpretation: the anchor primarily improves generation stability, while its effect on aggregate accuracy is limited under the evaluated settings.

We have revised the anchor ablation discussion accordingly and included the additional rollout-level analysis in the revised manuscript.

We hope these additional experiments and clarifications address the reviewer's concerns and provide a more precise characterization of VCSD's benefits and limitations.

As the discussion period is coming to a close, we wanted to briefly follow up on our rebuttal. If any of your concerns remain, we’d be happy to clarify. If our response addresses your main concerns, we’d appreciate it if you would consider revisiting your rating. Thank you!

Reviewer_FT68OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a new version of on-policy distillation for the vision domain. Here, a student generates an answer to a question based on an image. Instead of standard RL, the student's generation is reinforced using information from two teachers. The teachers share the same model (EMA of the student) but act on different image inputs: either the complete image or a black image. The student distribution is sharpened according to the difference between the two teachers. The idea is to sharpen or upweight the tokens that are relevant for the image compared to a blank image. On multiple datasets, the proposed method gives better results than GRPO or OPSD baselines.

Critical strengths

The proposed method is simple and doesn’t require additional privileged information, such as boxes.

The method achieves good results.

Critical weaknesses

The computational overhead of the method is not given compared to OPSD or GRPO. VCSD needs to do two additional teacher forwards, as opposed to 1 forward for OPSD or none for GRPO. The time per step should be measured and reported. Where are the methods compared using the same amount of training time or FLOPs?

The training time of 90 steps, each having 32 prompts, is quite a small training time, meaning not a lot of training signal. For example, in the Perception-aware-RLVR repository [1] the GRPO is trained with 200 steps each having 384 prompts, giving 25x more generated data, thus more training signal. GRPO / DAPO have their best checkpoint towards the end of training in this repository, meaning that a small training run doesn’t show them at their best. Sharpening methods like VCSD can make most of their gains early, giving them an advantage. This is not necessarily bad; it’s an advantage of the method, but we need to know if the gains are consistent across training or just for early steps. Presenting the performance of all methods at intermediate checkpoints across training will be beneficial.

It is unclear how the checkpoint selection was made. Is it after the 90 training steps? Is the best checkpoint of each method reported? Checkpoint selection and longer training time can change the results a lot, so it will be good to have a more complete view.

My main concern is that the baseline might be undertrained, and the proposed method might be doing just a small sharpening that can be achieved quickly. The authors need to clarify this.

[1] https://github.com/zsxm1998/awesome-perception-aware-rlvr/tree/main/examples/comparison/rlvr_qwen3_vl_4b

Questions for authors

It’s unclear if the OPSD and GRPO methods in Table 1 were trained by the authors. Can the authors confirm if they trained these methods? Also, what is the privileged information used in OPSD in this setting?
The paper mentioned “the published answer-hint OPSD method (Zhao et al., 2026)”. Is this baseline run by the authors, as the original paper didn’t have vision experiments, only math experiments? What is used as a hint in this baseline?

3 comments

We thank the reviewer for the constructive feedback and thoughtful suggestions. We have conducted additional experiments to address the concerns regarding computational overhead, training budgets, and checkpoint selection. All new results have been incorporated into Appendix B of the revised manuscript, with changes highlighted in blue.

W1. Computational overhead and compute-matched comparison

We agree that reporting computational overhead and comparing methods under matched training compute are important for assessing VCSD's practical efficiency.

We measured the steady-state per-step training cost of all four methods on identical hardware (Qwen3-VL-2B, 8 GPUs, same batch size), averaging steps 5–10 to exclude warm-up effects.

MethodTime / step (s)Relative to GRPOPeak memory (GB)
SFT22.90.52×132
GRPO44.41.00×141
OPSD50.01.13×146
VCSD (ours)65.41.47×214 (1.51×)

Computational overhead. VCSD incurs 1.47× the per-step training time and 1.51× the peak memory of GRPO. Rollout generation takes approximately 27–28 seconds per step across the on-policy methods. The additional cost of VCSD arises from the update phase, primarily due to two no-gradient teacher forward passes and full-vocabulary target construction. Importantly, VCSD introduces no inference-time overhead, as only the student model is deployed.

Compute-matched comparison. To evaluate whether the improvement persists when accounting for this additional training cost, we extended GRPO training and compared its performance at different total compute budgets.

MethodRelative training cost7-benchmark average
GRPO @ step 900.68×65.59
GRPO @ step 1801.36×65.70
GRPO @ step 2702.04×65.80
VCSD @ step 901.00×67.49

Even when GRPO is allocated substantially more training compute than VCSD (1.36×–2.04×), its accuracy remains below VCSD (65.70–65.80 vs. 67.49). Thus, the performance advantage of VCSD is not explained by its higher per-step computational cost.

Furthermore, the extended training experiments discussed in W2 demonstrate that this advantage persists over substantially longer GRPO training horizons.

Both the per-step profiling and compute comparison are included in Appendix B of the revised manuscript.

W2. Training budget and potentially undertrained baselines

We appreciate the reviewer's concern that the original training budget may be insufficient for some baselines, particularly GRPO.

To investigate this, we extended GRPO training by up to 12.5× the original 90-step budget and evaluated multiple intermediate checkpoints. All evaluations use the same seven-benchmark protocol as Table 1. For clarity, the results in Table 1 correspond to the final checkpoint of the fixed 90-step training budget for every method, without checkpoint selection.

BackboneMethodStep 90Step 100Step 300Step 600Step 1126
Qwen3.5-4BGRPO74.9774.5873.8673.9273.28
VCSD (ours)76.77————
Qwen3.5-9BGRPO76.3876.2576.3776.0976.57
VCSD (ours)79.24————

These results provide two observations.

First, extending GRPO training does not close the performance gap. On Qwen3.5-4B, the highest GRPO accuracy among the evaluated checkpoints is 74.97, compared with 76.77 for VCSD. On Qwen3.5-9B, GRPO remains within approximately 76.1–76.6 across the evaluated checkpoints, compared with 79.24 for VCSD. In both cases, GRPO remains below VCSD throughout the evaluated training horizon.

Second, longer training does not necessarily improve GRPO performance. On the 4B backbone, GRPO achieves its highest evaluated accuracy at the original 90-step checkpoint, with lower accuracy at subsequent checkpoints. On the 9B backbone, its performance fluctuates within a relatively narrow range without a sustained upward trend.

We also examined the training trajectories of the other methods. SFT achieves its highest observed performance at its first evaluated checkpoint and subsequently declines from 63.86 to 62.13 between steps 10 and 90. In contrast, the VCSD and OPSD trajectories shown in Figure 4 do not exhibit comparable degradation over their evaluated training windows.

Together with the computational profiling in W1, these experiments indicate that simply increasing GRPO's training budget does not eliminate VCSD's advantage. The extended training results are included in Appendix B.

W3. Checkpoint selection protocol

We thank the reviewer for pointing out that the checkpoint selection protocol should be stated more explicitly.

All methods in Table 1 are evaluated using the final checkpoint of the same fixed 90-step training budget. No method-specific best-checkpoint selection is performed, including for VCSD. We have clarified this protocol in the Setup section of the revised manuscript.

We additionally examined whether alternative checkpoint selection could affect the comparison.

  • Intermediate checkpoint evaluation. As shown in our W2 response, GRPO does not exceed VCSD at any evaluated checkpoint on either Qwen3.5-4B or Qwen3.5-9B, even when training is extended substantially beyond the original budget.
  • VCSD training stability. The VCSD training trajectory in Figure 4 remains stable near the final checkpoint, indicating that the reported performance is not attributable to an unusually favorable stopping point.

These additional results suggest that the main performance comparison is robust to the considered checkpoint selection alternatives.

Q1/Q2. Baseline implementation and privileged information in OPSD

We thank the reviewer for requesting clarification of the baseline implementations and the privileged information used by OPSD.

All baselines (SFT, GRPO, and OPSD) were trained by us, using the same backbone models, training dataset (ViRL39K), and fixed 90-step optimization budget as VCSD. Method-specific hyperparameters were tuned separately and are detailed in Appendix A.

For OPSD, the privileged information is the ground-truth reference answer. Following the answer-hint construction of Zhao et al., the teacher is conditioned on the reference answer, while the student receives only the image and question. We adapted this answer-hint mechanism to the vision-language setting.

This distinction is important because VCSD does not require reference answers or other privileged annotations: its supervision is constructed directly from the model's predictions under different visual conditions.

We hope these clarifications and additional experiments address the reviewer's concerns regarding computational efficiency, training budgets, and evaluation fairness.

As the discussion period is coming to a close, we wanted to briefly follow up on our rebuttal. If any of your concerns remain, we’d be happy to clarify. If our response addresses your main concerns, we’d appreciate it if you would consider revisiting your rating. Thank you!

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their constructive and detailed feedback. In response to the reviewers' concerns, we have conducted extensive additional experiments to further examine VCSD's computational efficiency, robustness to training budgets, effectiveness of the visual contrast signal, and generalization to additional benchmarks.

All new evaluations follow the main paper's evaluation protocol, using the same seven benchmarks, judge, and decoding settings where applicable. We summarize the key findings below.

Key takeaways.

  • Compute efficiency: Although VCSD incurs 1.47× the per-step training cost of GRPO, GRPO remains below VCSD even when allocated up to approximately 2× the total training compute.
  • Extended training: On the 4B and 9B backbones, GRPO does not match VCSD at any evaluated checkpoint, including runs extended to 7–12.5× the original training budget.
  • Additional benchmarks: VCSD achieves the highest accuracy among the evaluated methods on both MMMU and MMMU-Pro at the 4B and 9B scales.
  • Contrast utilization: Using the visual contrast to construct a full-distribution distillation target outperforms alternative applications of the same signal, including decoding, gating, and loss reweighting.
  • Visual grounding: Additional token-level analyses characterize the early-stage failure mode of visual contrast and show how the plausibility constraint and EMA teacher mitigate it.

Together, these results provide further empirical support for our main conclusions while clarifying the computational trade-offs and the distinct role of contrast-based target construction.

Detailed point-by-point responses follow. The new experiments and analyses have been incorporated into Appendix B of the revised manuscript, with changes highlighted in blue.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.