acceptodds
Under review as a conference paper at ICLR 2027

Learning to See from Reading: Cross-Modal Self-Distillation for Coding

Abstract

A multimodal model can solve the same problem differently depending on whether the input is presented as text or as an image, even when both views preserve the same content. This asymmetry suggests a natural question: can the model use its stronger view to improve its weaker one? We study this question through same-model cross-modal on-policy distillation (OPD) for coding from rendered images. Specifically, an image-conditioned student generates its own responses, while a frozen copy of the same pretrained model reads the paired source text and provides next-token distributions along the student’s prefixes. Matching these distributions transfers supervision from the stronger text view without ground-truth solutions, task rewards, or a stronger external teacher. We further characterize this mechanism through projected view repair, which separates raw cross-modal disagreement from the component expressible by the student’s adaptation tangent and its alignment with task improvement. Empirically, mean image-input accuracy on MBPP+ rises from 46.6% to 52.6%, outperforming ground-truth supervised fine-tuning. Matched controls show that text-conditioned feedback yields a stronger image-student endpoint than image-conditioned feedback, while exact transcription also improves across rendering densities. The gains extend across views: image-trained checkpoints improve text-input coding beyond the frozen pretrained teacher. Together, these results show that a model can use a stronger representation of the same task as a source of self-supervision for a weaker one.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Cross-modal self-distillation, On-policy distillation, Multimodal language models, Modality gap, Visual code understanding

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_fyZxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates a prominent performance asymmetry in multimodal large language models: the same pretrained model performs substantially better on coding tasks when given source text than when given rendered images of the exact same task (e.g., Qwen3.5-4B achieves 75.1% pass@1 on text inputs vs. 55.2% on rendered image inputs across HumanEval and MBPP). Positing that this gap stems from uneven visual access rather than a lack of underlying coding knowledge, the authors propose Cross-Modal Self-OPD (On-Policy Distillation).

Critical strengths

The primary strength supporting acceptance is the clean problem formulation that isolates input representation privilege from confounding supervision sources (no external teacher capacity, no execution rewards, and no ground-truth targets), demonstrating that a multimodal model can bootstrap its weaker visual pathway purely by querying its own stronger text pathway on student-generated prefixes—and that doing so outperforms supervised fine-tuning on ground-truth solutions (60.9% vs. 56.5%).

Critical weaknesses
  1. Single-Model Evaluation and Small-Scale Synthetic Setup: The entire empirical validation is restricted to a single 4B model (Qwen3.5-4B) trained on a very small synthetic dataset (256 first-stage pairs and 360 second-stage tasks for 810 total updates) using clean, deterministically rendered monospaced text images. Without validation across different model families, parameter scales, or realistic visual coding environments, it remains unclear whether Cross-Modal Self-OPD scales as a general paradigm beyond repairing synthetic font-rendering bottlenecks in a single small VLM.
  2. Statistically Insignificant Headline Claims (Cross-View Transfer & Teacher Controls): Several central claims highlighted in the abstract and introduction are not supported by statistical significance under the paper's own exact McNemar tests. Specifically, the claim that image-trained checkpoints improve text-input coding beyond the frozen teacher (75.1% →\rightarrow 76.4%, Table 7) relies on per-seed gains that are all statistically indistinguishable from noise (p=0.49,0.46, and 0.23p = 0.49, 0.46, \text{ and } 0.23 in Table 8). Similarly, the matched teacher-view control (Image/text vs. start, p=0.13p = 0.13 in Table 2) and the teacher-solution ablation (p=0.10p = 0.10 in Table 4) do not reach conventional statistical significance (p<0.05p < 0.05).
Questions for authors
  1. Can you validate Cross-Modal Self-OPD on at least one additional VLM family or model scale beyond Qwen3.5-4B?
  2. Evaluate on realistic, noisy IDE screenshots (e.g., SWE-bench Multimodal) and explore unfreezing the vision-language projector to compare input-specific vs. shared-layer repair.
  3. Given that the per-seed McNemar pp-values for text-view transfer (Table 8, p≥0.23p \ge 0.23), teacher-view control (Table 2, p=0.13p = 0.13), and solution context (Table 4, p=0.10p = 0.10) are all >0.05> 0.05, can you provide pooled significance tests or temper these from headline claims to exploratory observations?
5 comments
AuthorsOpenReview ↗

Response to Reviewer fyZx

We sincerely thank the reviewer for the constructive feedback. We appreciate the recognition of our clean problem formulation, which isolates input representation privilege, and the gains over ground-truth supervised fine-tuning. We have addressed the concerns about cross-model generalization, realistic visual environments, and statistical evidence with additional experiments and clarifications, as detailed below.

Critical Weaknesses

W1. Single-Model Evaluation and Small-Scale Synthetic Setup: The entire empirical validation is restricted to a single 4B model (Qwen3.5-4B) trained on a very small synthetic dataset (256 first-stage pairs and 360 second-stage tasks for 810 total updates) using clean, deterministically rendered monospaced text images. Without validation across different model families, parameter scales, or realistic visual coding environments, it remains unclear whether Cross-Modal Self-OPD scales as a general paradigm beyond repairing synthetic font-rendering bottlenecks in a single small VLM.

Response.

Different model. To evaluate whether cross-modal self-OPD generalizes beyond Qwen3.5-4B, we conducted additional experiments with Molmo2-O-7B, a different VLM family with an OLMo3-7B-Instruct language backbone and a SigLIP 2 vision backbone. We retained the original training data, two-stage procedure, and update budget, and evaluated three independent training seeds.

Table R1. Image-input coding performance on Molmo2-O-7B.

ModelHumanEval /164MBPP /427Combined /591Gain vs. Base
Molmo2-O-7B5984143 (24.20%)—
+ Self-OPD (seed 17)82122204 (34.52%)+10.32 pp
+ Self-OPD (seed 18)75102177 (29.95%)+5.75 pp
+ Self-OPD (seed 19)75101176 (29.78%)+5.58 pp

As shown in Table R1, cross-modal self-OPD consistently improves performance on both benchmarks across all three seeds. Exact paired McNemar tests yield p = 1.37 × 10⁻⁸, 0.000987, and 0.00203, respectively. These results provide evidence that our method extends to a different VLM family without increasing the original training scale.

Manuscript revision: We have added the Molmo2-O-7B experiments, three-seed results, and paired tests to Section 6, Observation 4, and Table 3 (p. 9), marked in blue.

Realistic IDE screenshots. To evaluate performance beyond synthetic rendering, we tested the original Qwen3.5-4B checkpoints on all 427 MBPP tasks presented as screenshots from a running code-server IDE. The screenshots include light and dark editor themes. We also evaluated image degradation using 0.9× downsampling and JPEG compression (quality 85). No IDE-specific retraining was performed.

Table R2. Image-input coding performance on IDE screenshots.

MethodIDE ScreenshotsIDE + Degradation
Qwen3.5-4B282/427 (66.04%)270/427 (63.23%)
+ GT SFT289/427 (67.68%, +1.64 pp)282/427 (66.04%, +2.81 pp)
+ Self-OPD (seed 17)296/427 (69.32%, +3.28 pp)279/427 (65.34%, +2.11 pp)
+ Self-OPD (seed 18)301/427 (70.49%, +4.45 pp)289/427 (67.68%, +4.45 pp)
+ Self-OPD (seed 19)296/427 (69.32%, +3.28 pp)285/427 (66.74%, +3.51 pp)

As shown in Table R2, cross-modal self-OPD achieves mean improvements of +3.67 pp on IDE screenshots and +3.36 pp under image degradation across three seeds. These results extend our evaluation to realistic IDE screenshots without additional training.

Manuscript revision: We have added the IDE screenshot and degradation experiments to Section 6, Observation 5, and Table 4 (p. 9). These additions are marked in blue.

AuthorsOpenReview ↗

Response to Reviewer fyZx

W2. Statistically Insignificant Headline Claims (Cross-View Transfer & Teacher Controls): Several central claims highlighted in the abstract and introduction are not supported by statistical significance under the paper's own exact McNemar tests. Specifically, the claim that image-trained checkpoints improve text-input coding beyond the frozen teacher (75.1% → 76.4%, Table 7) relies on per-seed gains that are all statistically indistinguishable from noise (p = 0.49,0.46, and 0.23 in Table 8). Similarly, the matched teacher-view control (Image/text vs. start, p = 0.13 in Table 2) and the teacher-solution ablation (p = 0.10 in Table 4) do not reach conventional statistical significance (p < 0.05).

Response. First, we would like to clarify that our primary finding is the consistent improvement in image-input coding. On Qwen3.5-4B, cross-modal self-OPD improves accuracy from 55.2% to 60.2–61.6% across three independent seeds, with statistically significant gains over both Pretrained and GT SFT under exact paired McNemar tests.

We have removed the claims about text-input transfer and the matched teacher-view comparison from the abstract, while retaining the statistically supported image-input improvements as our primary finding. The auxiliary code-stage teacher-view comparison, text-input transfer, and teacher solution context are presented as exploratory observations in Section 6.

Manuscript revision: We have removed the text-input transfer and matched teacher-view claims from the abstract (p. 1). Section 6 now explicitly qualifies the original code-stage teacher-view comparison in Observation 2 (pp. 7–8), text-input transfer in Observation 6 (p. 10), and teacher solution context in Observation 7 (p. 10) as auxiliary exploratory.

AuthorsOpenReview ↗

Response to Reviewer fyZx

Questions For Authors

Q1. Can you validate Cross-Modal Self-OPD on at least one additional VLM family or model scale beyond Qwen3.5-4B?

Response. To evaluate whether cross-modal self-OPD generalizes beyond Qwen3.5-4B, we conducted additional experiments on Molmo2-O-7B, a different VLM family with an OLMo3-7B-Instruct language backbone and a SigLIP 2 vision backbone. As shown in Table R1, cross-modal self-OPD consistently improves performance on both benchmarks across all three seeds. We have added the Molmo2-O-7B experiments, three-seed results, and paired tests to Section 6, Observation 4, and Table 3 (p. 9), marked in blue.

Q2. Evaluate on realistic, noisy IDE screenshots (e.g., SWE-bench Multimodal) and explore unfreezing the vision-language projector to compare input-specific vs. shared-layer repair.

Response. (a) Realistic IDE screenshots. As shown in Table R2, cross-modal self-OPD achieves mean improvements of +3.67 pp on IDE screenshots and +3.36 pp under image degradation across three seeds. These results extend our evaluation to realistic IDE screenshots without additional training. We have added the IDE screenshot and degradation experiments to Section 6, Observation 5, and Table 4 (p. 9).

(b) Projector-unfreezing ablation. We further investigated whether updating the vision-language projector provides additional benefits. On Qwen3.5-4B (seed 17), we compared the original language-layer LoRA adaptation with jointly training language-layer LoRA and the projector under the same training budget.

Table R3. Projector-unfreezing ablation on image-input coding.

Trainable ComponentsHumanEval /164MBPP /427Combined /591
Language LoRA (Projector Frozen)115243358 (60.58%)
Language LoRA + Projector119235354 (59.90%)

As shown in Table R3, unfreezing the projector improves HumanEval by four tasks but reduces MBPP by eight, yielding no additional overall improvement. This suggests that language-layer adaptation alone can effectively improve image-input coding without requiring projector updates.

Manuscript revision: We have added this ablation to Appendix C.9 and Table 17 (p. 22), marked in blue, with a pointer in Section 6's Additional diagnostics paragraph (p. 10).

Q3. Given that the per-seed McNemar p-values for text-view transfer (Table 8, p ≥ 0.23), teacher-view control (Table 2, p = 0.13), and solution context (Table 4, p = 0.10) are all > 0.05, can you provide pooled significance tests or temper these from headline claims to exploratory observations?

Response. According to your suggestion, we have removed the claims about text-input transfer and the matched teacher-view comparison from the abstract, while retaining the statistically supported image-input improvements as our primary finding. The auxiliary code-stage teacher-view comparison, text-input transfer, and teacher solution context are presented as exploratory observations in Section 6.

Manuscript revision: We have removed the text-input transfer and matched teacher-view claims from the abstract (p. 1). Section 6 now explicitly qualifies the original code-stage teacher-view comparison in Observation 2 (pp. 7–8), text-input transfer in Observation 6 (p. 10), and teacher solution context in Observation 7 (p. 10) as exploratory.

Reviewer_fR5KOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper observes that Qwen3.5-4B solves coding tasks much better from source text than from a rendered image of the same text. It proposes cross-modal self-OPD: an image-conditioned LoRA student samples its own responses, and a frozen copy of the same model, which reads the paired text, supervises it with top-100 JSD. Over 3 seeds, image-input pass@1 improves from 55.2 to 60.9.

Critical strengths

Clean design with a reliable gain. The teacher and student are the same model solving the same task; the only difference is that the teacher reads text and the student sees an image. This rules out a stronger external teacher as the source of the gain. The improvement (image pass@1 55.2% → 60.9%) is significant in each of three independent runs.

Critical weaknesses

The evidence that the text view causes the gain is thin. The paper's central claim is that the student improves because the teacher reads text. The direct test is to repeat the same training with a teacher that sees the image instead. The paper runs this comparison only for Stage 2, in a single run. Stage 1, which accounts for about half of the improvement, has no such comparison at all. A simple training-free baseline is also missing: transcribe the image first, then solve from the transcription. Since the base model already reads rendered code quite accurately, this baseline could close part of the gap. Several claims are stronger than the evidence. The improvement on text-input coding is small and not statistically significant. The finding that adding the ground-truth solution to the teacher does not help comes from a single run. These claims should be softened or better supported.

Questions for authors

What is the accuracy of transcribe-then-solve with the pretrained model? Can you add image-teacher / no-teacher controls for the full two-stage pipeline and report the corresponding results? Can you provide a failure-type breakdown before vs. after training?

4 comments
AuthorsOpenReview ↗

Response to Reviewer fR5K

We sincerely thank the reviewer for the thoughtful and constructive feedback. We appreciate the recognition of our clean experimental design and the statistically significant image-input gains across three independent runs. In response to the concerns raised, we have added the full-pipeline teacher-view comparison, the transcribe-then-solve baseline, and the failure-type analysis, and removed the auxiliary claims about text-input transfer and the matched teacher-view comparison from the abstract. We address each point below.

Critical Weaknesses

W1. The evidence that the text view causes the gain is thin. The paper's central claim is that the student improves because the teacher reads text. The direct test is to repeat the same training with a teacher that sees the image instead. The paper runs this comparison only for Stage 2, in a single run. Stage 1, which accounts for about half of the improvement, has no such comparison at all.

Response. To address this concern, we conducted additional experiments comparing text-conditioned and image-conditioned teachers across the full training pipeline. Both configurations use the same image-conditioned student architecture, training budget, and evaluation protocol.

Table R1. Image-input coding performance with different teacher views.

ConfigurationHumanEval /164MBPP /427Combined /591
Pretrained (Baseline)122208330 (55.8%)
Image Teacher114216330 (55.8%)
Text Teacher (Ours)115243358 (60.6%)

Both teacher configurations complete the full two-stage training pipeline. Results are from a single-seed comparison.

As shown in Table R1, the text-conditioned teacher improves image-input coding accuracy from 55.8% to 60.6%, solving 28 additional tasks compared with both the pretrained baseline and the image-conditioned teacher. In contrast, the image-conditioned teacher does not improve the aggregate accuracy over the pretrained model. These results provide further evidence that the stronger text view offers more effective supervision for image-conditioned coding.

Manuscript revision: We have added this full-pipeline comparison to Section 6, Observation 2 (pp. 7–8), and Table 2(a) (p. 8). The added experiment and discussion are marked in blue.

W2. A simple training-free baseline is also missing: transcribe the image first, then solve from the transcription. Since the base model already reads rendered code quite accurately, this baseline could close part of the gap.

Response. We thank the reviewer for suggesting this important baseline. To examine whether the benefits of cross-modal self-OPD extend beyond direct image-input coding, we conducted additional experiments using a transcribe-then-solve procedure, where the model first transcribes the rendered task image and then solves the problem from the generated text. We evaluated both the pretrained model and the image-trained checkpoint under this setting.

Table R2. Coding performance under different inference procedures.

ModelInference ProcedureHumanEval /164MBPP /427Combined /591
PretrainedDirect Image122208330 (55.8%)
Cross-modal self-OPDDirect Image115243358 (60.6%)
PretrainedTranscribe-Then-Solve117296413 (69.9%)
Cross-modal self-OPDTranscribe-Then-Solve124300424 (71.7%)

As shown in Table R2, cross-modal self-OPD improves coding performance under both inference procedures. Direct image-input accuracy increases from 55.8% to 60.6%, while transcribe-then-solve accuracy increases from 69.9% to 71.7%. These results suggest that the benefits of cross-modal self-distillation are not restricted to direct image-input coding but can also extend to an alternative inference procedure.

Importantly, the effectiveness of transcribe-then-solve does not contradict our central contribution. Our work investigates the representation asymmetry of the same pretrained model across different input views and how its stronger text view can provide supervision for improving image-conditioned coding. Transcribe-then-solve addresses this asymmetry through an explicit intermediate transcription, whereas our method uses the stronger view as a source of self-supervision during training. The two approaches therefore address the same underlying challenge through complementary mechanisms.

Manuscript revision: We have added the transcribe-then-solve evaluation and its interpretation to Section 6, Observation 3 (pp. 8–9), and Table 2(b) (p. 8), alongside the direct-image results in Table 2(a). These additions are marked in blue.

AuthorsOpenReview ↗

Response to Reviewer fR5K

W3. Several claims are stronger than the evidence. The improvement on text-input coding is small and not statistically significant. The finding that adding the ground-truth solution to the teacher does not help comes from a single run. These claims should be softened or better supported.

Response. First, we would like to clarify that our primary finding is the consistent improvement in image-input coding. On Qwen3.5-4B, cross-modal self-OPD improves accuracy from 55.2% to 60.2–61.6% across three independent seeds, with statistically significant gains over both Pretrained and GT SFT under exact paired McNemar tests.

Accoring to your suggestion, we have removed the claims about text-input transfer and the matched teacher-view comparison from the abstract, while retaining the statistically supported image-input improvements as our primary finding. The auxiliary code-stage teacher-view comparison, text-input transfer, and teacher solution context are presented as exploratory observations in Section 6.

Manuscript revision: We have removed the text-input transfer and matched teacher-view claims from the abstract (p. 1). Section 6 now explicitly qualifies the original code-stage teacher-view comparison in Observation 2 (pp. 7–8), text-input transfer in Observation 6 (p. 10), and teacher solution context in Observation 7 (p. 10) as auxiliary exploratory.

Questions For Authors

Q1. What is the accuracy of transcribe-then-solve with the pretrained model?

Response. The pretrained model achieves 413/591 (69.9%) with transcribe-then-solve: 117/164 on HumanEval and 296/427 on MBPP (Table R2 above). We have added the transcribe-then-solve evaluation and its interpretation to Section 6, Observation 3 (pp. 8–9), and Table 2(b) (p. 8).

Q2. Can you add image-teacher / no-teacher controls for the full two-stage pipeline and report the corresponding results?

Response. We conducted additional experiments comparing text-conditioned and image-conditioned teachers across the full training pipeline. Both configurations use the same image-conditioned student architecture, training budget, and evaluation protocol. The image-conditioned teacher solves 330/591 (55.8%), while the text-conditioned teacher solves 358/591 (60.6%) (Table R1 above; Section 6, Observation 2, and Table 2(a)).

Q3. Can you provide a failure-type breakdown before vs. after training?

Response. We thank the reviewer for this valuable suggestion. To better understand the improvements introduced by cross-modal self-OPD, we conducted an additional analysis of the execution outcomes from the pretrained model and all three image-trained checkpoints. Each of the 591 coding tasks was assigned to a mutually exclusive outcome category.

Table R3. Execution outcome breakdown before and after cross-modal self-OPD.

Execution OutcomePretrainedSelf-OPD (s17)Self-OPD (s18)Self-OPD (s19)
Solved326356359364
Test-Assertion Failure179147152144
Runtime Exception81837581
Syntax / Indentation Error4552
Timeout1000

As shown in Table R3, test-assertion failures decrease consistently across all three runs, from 179 to 147, 152, and 144, respectively. In contrast, runtime exceptions remain broadly comparable, while syntax errors and timeouts are relatively rare. The reduction in failed assertions therefore accounts for much of the observed improvement in coding accuracy. Together with the improvements in exact transcription reported in our paper, these results provide complementary evidence for understanding the effects of cross-modal self-distillation on rendered-code reading and downstream coding performance. We have incorporated the failure-type breakdown into Appendix C.8 and Table 16 (p. 22), marked in blue, and clarified its interpretation as a descriptive error-profile analysis rather than evidence of a single causal mechanism. Section 6's Additional diagnostics paragraph (p. 10) points readers to this analysis.

AuthorsOpenReview ↗

Hugely appreciate the suggestions and looking forward to engaging further!

Thank you again for your constructive feedback. We have responded to your concerns point by point and hope our responses help clarify the issues you raised. If you have any further questions, we would be happy to provide additional clarification. Thank you again for your time and consideration!

AuthorsOpenReview ↗

Follow-up on Our Rebuttal and Request for Reassessment

Dear Reviewer,

We would appreciate your feedback on our rebuttal when you have an opportunity. Please let us know if any concerns remain or further clarification would be helpful.

If our responses address your concerns, we would be grateful if you could reconsider your assessment and score.

Thank you for your time and consideration.

Reviewer_Jsu8OpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper uses a model’s stronger text view to improve coding from images. A frozen copy reads the source text and gives token-level feedback on responses generated by the image-conditioned model. This provides training without correct solution targets or a larger teacher, and the experiments show gains in image-input coding and code transcription.

Critical strengths

The paper presents an interesting idea, backed by substantial theoretical analysis and solid experiments. The experimental results show clear performance of the proposed approach.

Critical weaknesses

I did not identify any notable weaknesses.

Questions for authors

I am unfamiliar with the research topic of this paper. But I think the contribution of the paper may be solid.

1 comment
AuthorsOpenReview ↗

Response to Reviewer Jsu8

We sincerely thank the reviewer for the positive and encouraging assessment of our work. We appreciate the recognition of the paper's central idea, theoretical analysis, and experimental results.

Our key contribution is demonstrating that a multimodal model can leverage its stronger text view to improve image-conditioned coding through cross-modal self-distillation, without relying on a stronger external teacher, ground-truth solutions, or task rewards.

We are grateful for the reviewer's thoughtful feedback and positive evaluation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Additional Evidence

Dear Area Chair and Reviewers,

We thank all reviewers for their constructive feedback, which strengthened our experiments and clarified our contributions. We summarize the discussion and additional rebuttal evidence.

Strengths recognized by the reviewers.

  • Problem formulation. The work identifies a performance asymmetry between text and image representations of the same coding task, isolating input modality from stronger external teachers, ground-truth solutions, and task rewards (fyZx, fR5K).
  • Method. Cross-modal self-OPD enables a multimodal model to improve its weaker image-conditioned coding ability using supervision from its own stronger text view, without requiring an external teacher or ground-truth solutions (fyZx, fR5K, Jsu8).
  • Experiments. Reviewers recognized the clean experimental design and consistent improvements across three independent runs. On Qwen3.5-4B, image-input coding accuracy improves from 55.2% to 60.9% on average, outperforming ground-truth supervised fine-tuning (fyZx, fR5K).
  • Analysis. Theoretical and experimental analyses support cross-modal self-distillation and clarify the relationship between representation differences and task improvement (Jsu8).

Reviewer Jsu8 recognized the interesting central idea, substantial theoretical analysis, and solid experimental results. The reviewer identified no notable weaknesses and assigned a rating of 3 (Weak Acceptance).

Reviewer fR5K. Our rebuttal provides additional experiments and analyses addressing the reviewer's major concerns:

  • Teacher-view comparison (W1, Q2). We evaluated text-conditioned and image-conditioned teachers across the complete two-stage training pipeline under identical training budgets. The text teacher achieves 60.6% accuracy, compared with 55.8% for both the image teacher and pretrained baseline. This additional single-seed comparison supports the benefit of supervision from the stronger text view (Section 6, Table 2(a)).
  • Transcribe-then-solve baseline (W2, Q1). We added the requested inference baseline. Transcribe-then-solve achieves 69.9% accuracy with the pretrained model and 71.7% after cross-modal self-OPD, showing that our training also improves performance under this alternative inference procedure (Section 6, Table 2(b)).
  • Failure analysis (Q3). Across three independent runs, test-assertion failures decrease from 179 to 147, 152, and 144, respectively, while runtime exceptions remain broadly comparable. We added this descriptive analysis to Appendix C.8 (Table 16).
  • Statistical claims (W3). We removed unsupported auxiliary claims from the abstract and explicitly qualified text-input transfer, the original code-stage teacher-view comparison, and teacher-solution context as exploratory observations. The primary image-input improvements remain statistically significant across three seeds.

Reviewer fyZx. Our rebuttal extends the evaluation beyond the original model and synthetic rendering setting:

  • Cross-model generalization (W1, Q1). We evaluated Molmo2-O-7B, a different VLM family, using the original training data and update budget. Accuracy improves from 24.20% to 29.78–34.52% across three independent seeds, with statistically significant gains in every run (Section 6, Table 3).
  • Realistic IDE screenshots (W1, Q2). We evaluated all 427 MBPP tasks using screenshots from a running IDE, including different editor themes and image degradation. Without IDE-specific retraining, self-OPD achieves average improvements of 3.67 points on IDE screenshots and 3.36 points under downsampling and JPEG compression (Section 6, Table 4).
  • Projector-unfreezing ablation (Q2). We compared language-layer LoRA adaptation with jointly training the vision-language projector. The original configuration achieves 60.58% accuracy, compared with 59.90% when additionally unfreezing the projector, indicating no overall improvement from projector updates in this experiment (Appendix C.9, Table 17).
  • Statistical evidence (W2, Q3). Following the reviewer's suggestion, we narrowed the abstract to statistically supported findings and explicitly presented auxiliary results as exploratory observations in Section 6.

All additional experiments, analyses, and revisions are incorporated into the revised submission and highlighted in blue.

We believe these results strengthen the evidence that a multimodal model can leverage its stronger text representation to improve image-conditioned coding without external teacher supervision or ground-truth solutions. The additional cross-model and realistic IDE evaluations further support the applicability of cross-modal self-distillation beyond the original experimental setting. We thank the reviewers and Area Chair again for their time and engagement, and hope our responses and revisions will be considered in the final assessment.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.