{"id":"ea698bbe-a7bb-4953-b9d6-59874e8f5044","arxiv_id":"2607.23125","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A VLM improves itself by matching its own predictions on corrupted images to its predictions on clean images, without labels or external teachers, per 12-benchmark experiments.","lead":"This paper proposes NOPD, a training method in which a vision-language model learns from corrupted images while using its own predictions on clean images as the teaching signal, with no ground-truth answers. The reported effect is large: up to 20-point accuracy gains on geometry reasoning and consistent gains across 12 benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Self-consistency has no mechanism to track correctness: Eq. (6)'s fixed points include confidently wrong clean/corrupted pairs, and the paper never probes whether the online teacher's accuracy or calibration improves during training.","rationale":"The reader's weakest assumption—clean-image predictions as a reliable teacher—is indeed the most load-bearing condition. I agree that the absence of ground truth makes teacher-quality critical, but I sharpen the concern: the objective's fixed points are all self-consistent distributions, not necessarily correct ones, and the online sync-teacher update means the target itself can drift. This is a more precise failure mode than 'the teacher might be wrong.' The reader's conditional verdict already flags the need for robustness analysis and code release, and my concern reinforces that rather than changing it. I set agreement_with_reader to 'partial' because the reader identified the same assumption but did not formalize the fixed-point/drift argument or propose the specific teacher-accuracy-vs-KL probe. The verdict remains UNCHANGED: the paper is internally consistent and the reported experiments are extensive, but this unverified assumption justifies keeping the conditional accept.","tokens_in":18455,"tokens_out":9487,"duration_ms":115763,"concrete_test":"Run the Geometry3K NOPD training on Qwen2.5-VL-7B while holding out a labeled subset (e.g., 200 examples from the val split). Every 10 training steps, record (i) clean-teacher greedy accuracy on the held-out subset, (ii) corrupted-student accuracy, and (iii) the average DKL between clean and corrupted output distributions. If DKL decreases while clean-teacher accuracy does not improve (or decreases), the objective is being optimized by consistency collapse rather than by improved correctness, and the central self-boosting claim loses support. This check requires no change to the method and settles whether the teacher-quality assumption holds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Eq. (6), where the target πθ(·|q,x) is a stop-gradient copy of the same model's clean-image predictions. No ground-truth or external signal constrains this target, so the loss is minimized whenever πθ(·|q,T(x)) = πθ(·|q,x), regardless of whether that shared distribution is correct. The fixed-point family therefore includes all self-consistent clean/corrupted pairs, including confidently wrong ones. The paper's sync-teacher choice (Table 2c) makes this worse: because teacher and student share parameters, the target drifts as the student trains, and nothing prevents drift toward a wrong but self-consistent mode. The paper does not report the clean teacher's accuracy or calibration over training, nor does it track training loss against held-out accuracy (Figure 4 shows only final-benchmark accuracy curves). Without such a probe, the reported +20/+7.4 gains could reflect a beneficial consistency regularizer or the early phase of a self-consistency collapse; the method has no built-in guard against the latter. The gradient derivation in Eq. (8) is sound, so the concern is not internal inconsistency but an unverified load-bearing assumption about teacher quality.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes NOPD, a post-training method for vision-language models that uses no ground-truth answers, rewards, or external teachers. The training signal is a KL divergence between the model's distribution under a corrupted image T(x) (the 'student') and the same model's stop-gradient distribution under the clean image x (the 'teacher'), Eq. (6); responses are sampled on-policy from the corrupted branch. Experiments train on Geometry3K (2.1K pairs) and report gains on five reasoning benchmarks, with headline improvements of +20 on Geometry3K val and +7.4 on MathVista for Qwen2.5-VL-7B. Additional experiments on a larger mixed corpus report gains across 12 benchmarks for three VLM families, plus analyses of divergence choice, teacher update, generation length, corruption type, and an extension to captioning.","tokens_in":18757,"tokens_out":11953,"duration_ms":130727,"significance":"The contribution is potentially significant: if the reported effects are real, NOPD would be a rare example of a VLM post-training method that improves on external benchmarks without any external supervision, which could lower the cost of adapting VLMs. The paper is also unusually broad empirically: three model families, twelve benchmarks, several ablations, and a comparison against RLVR/OPD/OPSD. The self-referential training signal is a genuine correctness risk, and the current manuscript does not contain the diagnostics needed to rule out self-consistency collapse. I therefore view the contribution as interesting but not yet established.","major_comments":[{"comment":"Hyperparameters and design choices are selected on the same benchmarks that later serve as the headline results. Table 2 reports accuracy on MathVista to choose the loss type, full-vocab vs sampled-token computation, teacher update rule, and generation length; Table 3 chooses the corruption schedule on the Geometry3K test set. Table 1 then presents MathVista and Geometry3K as held-out evidence, including the headline +7.4 and +20 gains. This is model selection on the test set. Please either use a separate validation split for ablations or report the full set of ablated configurations on all benchmark suites so the reader can see that the design choices do not flip the conclusions.","section":"§4.1, §4.2 (Tables 1–3)"},{"comment":"Eq. (6) writes L = E_{y~πθ(·|q,T(x))} DKL(πθ(y|q,T(x)) || πθ(y|q,x)), but the KL divergence is a deterministic function of θ, so the outer expectation over y is not meaningful as written. The Appendix C derivation only goes through for a per-token KL with a fixed prefix, and even then Eq. (8) drops the dependence of the prefix distribution on θ. Please define the exact token-level objective used in the implementation, state clearly whether Eq. (8) is an approximation, and provide a derivation that matches the stated objective.","section":"§3 Eq. (6) and Appendix C"},{"comment":"The stop-gradient clean prediction is the only training signal, so the fixed points of Eq. (6) include any θ for which the clean and corrupted distributions coincide, regardless of whether that shared distribution is correct. The paper does not report the clean teacher's accuracy or calibration over training, nor the training loss trajectory; Figure 4 shows only benchmark accuracy curves. Because the teacher is fully synchronous (Table 2c), the target can drift during training. Please add a diagnostic tracking (i) the consistency loss, (ii) clean-teacher accuracy on a held-out set, and (iii) output diversity during training for at least the Geometry3K and MMK12 setups. If the clean accuracy does not improve or degrades, the claimed self-boosting effect would need to be reframed as a regularizer rather than evidence of teacher-driven improvement.","section":"§3 (Eq. 6) and §4.2 (Table 2c, Fig. 4)"},{"comment":"No multiple seeds or error bars are reported anywhere. Several Table 4 entries are small negative/positive deltas (e.g., GLM-4.6V-Flash MMBench-EN -0.2, ChartQA-Pro -0.2; Qwen3-VL-8B BLINK -0.6), and Qwen3.5-2B results use avg@16 random sampling. Without variance estimates, 'consistent gains' cannot be distinguished from noise. Please provide at least three seeds or confidence intervals for the main comparisons, or a bootstrapped significance analysis.","section":"Tables 1, 4; §4.1–4.3"}],"minor_comments":[{"comment":"The column header 'Off-polocy KD' is a typo; it should read 'Off-policy KD'.","section":"Table 7"},{"comment":"The name 'Noisy Student' coincides with a well-known semi-supervised learning method (Xie et al., 2020). The related-work section should cite and explicitly differentiate NOPD from that line of work.","section":"Related Work / Title"},{"comment":"The header 'testmini' for MathVision should be expanded or defined in the table caption.","section":"Table 1"},{"comment":"The evaluation uses GPT-OSS-120B as an answer judge for some tasks. Since the paper's selling point is 'no external models,' it would help to state clearly that this judge is used only at evaluation time, not during training.","section":"§4.3 / Appendix B.1"}],"recommendation":"major_revision","confidential_remarks":"The most serious threat to publication is the hyperparameter selection on the headline benchmarks (MathVista and Geometry3K). The teacher-quality diagnostic is also essential; without it, the self-referential training loop remains a correctness risk. If the authors can provide a clean validation split and the requested training-time diagnostics, I would be supportive; as it stands, the central empirical claim is not yet adequately isolated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: NOPD is a novel and unusually simple post-training method for VLMs—use the model's own predictions on clean images as token-level targets for the same model on corrupted images, reverse KL, stop gradient—and the paper backs it with a lot of experiments. I think the core claim is probably true, and the paper deserves a serious referee.\n\nWhat's actually new is the specific combination. OPSD needs ground-truth solutions; NoisyRollout needs verifiable rewards. NOPD needs neither. The paper is careful to distinguish these, and the gradient derivation in Appendix C is correct. The main experiments are extensive: three base models, twelve benchmarks, ablations on loss type, teacher update, corruption type, generation length, and a sensible rollout-temperature/diversity control showing the gain is not just added diversity. The implicit reasoning compression (fewer tokens, higher accuracy) is a nice byproduct and consistent with the method teaching the model to rely less on spurious visual detail.\n\nSoft spots, in order:\n\nNo error bars or multiple seeds anywhere. That's the biggest concrete weakness. The ablations in Table 2 are run once on MathVista, and design choices (reverse KL, sync teacher, gen length 2048) are made on that one benchmark. It would be easy to overfit—the reported 0.1–0.3 point differences between 2048 and 4096, or reverse KL vs. JSD, are within what seed noise would produce. I would want multi-seed numbers for at least the main comparison table before believing the specific margins.\n\nThe self-referential teacher is the conceptual worry. The fixed-point set of Eq. (6) includes confidently wrong clean/corrupted pairs, and with a sync teacher nothing prevents drift toward a wrong self-consistent mode. The stress-test note makes this point well. But the paper does partially address it: Figure 4 tracks benchmark accuracy over training steps for Qwen3.5-2B, so we know the model doesn't collapse immediately on those tasks. What's missing is a direct probe of the teacher's accuracy/calibration on clean inputs during training, especially for Qwen2.5-VL-7B, and any analysis of the training loss trajectory. That's a fair referee request, not a fatal flaw.\n\nMinor: 'without any external models' is slightly overstated. An external model filters the training data in Sec. 4.3, and GPT-OSS does answer extraction in evaluation. That's not part of the NOPD loss, but the abstract's phrasing is loose.\n\nNo code or data released. For a method this cheap and simple, that's a bigger deal than usual—people will want to reproduce the 20-point jump.\n\nBottom line: the central argument holds as an empirical claim. The math is coherent, the related work is handled honestly, and the evidence is broad. It's not a methods breakthrough—consistency regularization is old—but as a practical VLM post-training recipe it looks useful and cheap. I'd send it to review, ask for multi-seed results, code/data, and an analysis of teacher quality over training, and otherwise expect a solid conditional accept.","headline":"A simple self-distillation recipe that seems to work—clean-image predictions as teacher for corrupted inputs—with real experimental support, but reproducibility gaps and the absence of any correctness signal keep it from being a clean accept.","tokens_in":19237,"tokens_out":3488,"would_cite":true,"duration_ms":37447,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a vision-language model can improve itself without ground-truth answers or external teachers by minimizing the divergence between its own predictions on corrupted and clean images.","keywords":["vision-language models","self-distillation","noisy student","self-improvement","image corruption","visual reasoning","post-training","on-policy learning"],"falsifier":"Train NOPD on a set where the base model's clean-image answers are known to be systematically wrong (e.g., a set with available labels where the model scores near chance), and check whether accuracy improves or degrades. Additionally, track the training loss against validation accuracy: if the consistency loss falls while accuracy falls, the model is being trained into stable, confident errors—falsifying the claim that clean predictions are a safe teacher.","tokens_in":18351,"feed_emoji":"🖼️","tokens_out":6606,"duration_ms":66629,"temperature":0.7,"pith_summary":"This paper tries to establish that a vision-language model can improve its reasoning without any ground-truth answers, external teacher models, or verifiable rewards. The proposed method, NOPD, corrupts the input image (noise, downsampling, gamma changes), samples an answer from the model on the corrupted view, and trains the model to make that answer's token distribution match the distribution the same model produces on the clean image, with gradients blocked on the clean branch. Because the clean prediction is never checked against a correct answer, the method's success depends on the model's clean outputs being a trustworthy teacher. The authors report large gains: for example, training a 7-billion-parameter VLM on 2.1K geometry samples lifts its Geometry3K validation score by 20 points and MathVista by 7.4, with consistent gains across three models and twelve benchmarks. The paper notes the mechanism is specific to vision-language models because it requires corrupting images.","feed_headline":"A VLM trains itself by matching answers on clean vs noisy images","feed_subtitle":"No labels or teachers: 2.1K geometry examples add 20 points on Geometry3K validation.","key_machinery":"The load-bearing object is the asymmetric-input reverse-KL objective, Eq. (6): L = E_{q,x, y~πθ(·|q,T(x))} DKL(πθ(y|q,T(x)) ∥ πθ(y|q,x)), with a stop-gradient on the clean branch. The same network is both noisy student and clean teacher; the corruption operator T (Gaussian noise, downsampling, gamma) creates the asymmetry that makes the self-supervision non-trivial. The stop-gradient is essential: without it, the model could drive the loss to zero by producing identical outputs for any input, which the paper calls collapse.","core_discovery":"The central claim is that the gap between a VLM's predictions on corrupted and clean inputs is a complete self-supervision signal. NOPD samples an on-policy answer from the model on a corrupted image, then minimizes the reverse KL divergence between the corrupted-input token distribution and the clean-input distribution, with stop-gradient on the clean branch. The resulting gradient is a contrastive policy gradient whose log-ratio term rewards tokens that differ between the two views. In the authors' experiments this single-model scheme matches or beats RLVR and teacher distillation on five visual-reasoning benchmarks and improves accuracy out of distribution.","pith_inferences":["The consistency objective should be measurable as robustness: a model trained with NOPD should become less sensitive to the specific corruptions used (noise, resolution, gamma), and one could test this by evaluating on held-out corruption types; if robustness does not track accuracy, the mechanism is less about perception than about output regularization.","For a model whose clean outputs are already near-perfect, NOPD should produce little gain; the largest improvements should occur where the clean view is right and the corrupted view is wrong. One can test this by splitting training questions by base-model correctness and checking where the loss focuses.","The paper itself scopes the approach to VLMs, since the teacher-student asymmetry is created by corrupting images; extending the idea to text or audio would require finding a corruption that degrades the input without destroying the answer."],"forward_implications":["Post-training a VLM can in principle be done with raw image-question pairs alone: no human labels, no external teacher, no verifiable rewards, and only one rollout per sample.","The method simultaneously shortens reasoning: on MathVista, output tokens dropped by roughly 58 percent while accuracy rose, suggesting the learning signal is compressing verbose reasoning rather than merely memorizing.","Gains transfer beyond the training distribution: training on a small geometry set improved scores on out-of-distribution benchmarks like MathVista and MMMU-Pro.","Scaling the training set (15.6K samples) continues to improve accuracy, so the reported gains are not an artifact of one small dataset."],"fun_headline_variants":["VLM self-distills by comparing clean vs noisy predictions","No teachers, no labels: VLM boosts itself on noisy views","Self-supervised VLM gains 20 points from 2.1K geometry examples","On-policy self-distillation lifts VLM by 7.4 on MathVista","Clean vs noisy images teach VLMs without external data"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that a model's own clean-image predictions are trustworthy enough to serve as a teacher: the clean branch is detached and never compared with a ground-truth answer, so if the model is confidently wrong on clean images, NOPD can amplify the error rather than fix it.","fun_headline_variants_meta":{"raw":{"variants":["VLM self-distills by comparing clean vs noisy predictions","No teachers, no labels: VLM boosts itself on noisy views","Self-supervised VLM gains 20 points from 2.1K geometry examples","On-policy self-distillation lifts VLM by 7.4 on MathVista","Clean vs noisy images teach VLMs without external data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00016,"raw_usage":{"total_tokens":1074,"prompt_tokens":754,"completion_tokens":320,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":225}},"tokens_in":498,"tokens_out":320,"duration_ms":3810,"temperature":1.0,"reasoning_tokens":225,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T01:45:47.572900+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train NOPD on a set where the base model's clean-image answers are known to be systematically wrong (e.g., a set with available labels where the model scores near chance), and check whether accuracy improves or degrades. Additionally, track the training loss against validation accuracy: if the consistency loss falls while accuracy falls, the model is being trained into stable, confident errors—falsifying the claim that clean predictions are a safe teacher.","supporting_citations":[],"review_version":2}