{"id":"8ee46fc7-f77e-4845-b108-f80e085330cd","arxiv_id":"2605.29496","paper_version":1,"verdict":"UNVERDICTED","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Post-training of VLMs exhibits perception-reasoning asymmetry from token imbalance in SFT and reward coupling in RL, mitigated by loss reweighting (up to 18.2 gain) and perception-aware rewards (up to 6.0 gain).","lead":"The paper introduces a diagnostic framework using synthetic tasks to show that post-training improves reasoning more than perception in vision-language models due to token imbalance in supervised fine-tuning and reward coupling in reinforcement learning. Interventions like dynamic loss reweighting and perception-aware rewards can reduce this asymmetry and boost end-to-end performance.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Whether the two synthetic tasks truly disentangle perception from reasoning without confounding factors","rationale":"The reader's weakest_assumption matches the load-bearing precondition for the strongest_claim. Because the full manuscript is referenced but the separation claim is the point at which the causal attribution could break, the concern is already correctly identified and the UNVERDICTED status is appropriate pending further validation of the diagnostic tasks.","tokens_in":1748,"tokens_out":302,"duration_ms":18489,"concrete_test":"Vary perception difficulty independently (e.g., by altering visual noise or object count) while holding reasoning depth fixed across matched task instances; measure whether post-training deltas in end-to-end accuracy remain consistent with the reported asymmetry. If perception-only changes produce unexpected shifts in reasoning metrics, the disentanglement assumption fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim attributes the observed asymmetry (and its differing mechanisms under SFT vs. RL) to token imbalance or reward coupling, which in turn rests on the diagnostic framework's ability to isolate perception and reasoning. If the synthetic tasks contain unintended correlations (e.g., perception errors propagating into reasoning steps, or task artifacts that couple the two), then the measured improvements from reweighting or perception-aware rewards cannot be cleanly attributed to the proposed mechanisms. The abstract asserts the tasks 'disentangle' perception from reasoning, but this separation is the least-secured precondition for the rest of the argument.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper introduces a controlled diagnostic framework with two synthetic tasks designed to disentangle perception from reasoning in vision-language models. It reports a consistent perception-reasoning asymmetry under post-training: SFT improves reasoning more due to token imbalance in chain-of-thought supervision, while RL does so due to reward coupling with outcome signals. Interventions (dynamic loss reweighting for SFT; perception-aware rewards for RL, including a surrogate) are claimed to close the gap and improve end-to-end accuracy by up to 18.2, 6.0, and 3.2 points.","tokens_in":1851,"tokens_out":426,"duration_ms":21541,"significance":"If the diagnostic tasks isolate the factors as claimed, the work supplies a mechanistic account of why perception lags reasoning in VLM post-training and supplies concrete, paradigm-specific fixes. The controlled synthetic setup and the SFT-vs-RL distinction are methodological strengths that could inform future training recipes.","major_comments":[{"comment":"§3 (Synthetic Tasks): The central attribution of the asymmetry to token imbalance (SFT) or reward coupling (RL) rests on the claim that the two synthetic tasks cleanly disentangle perception from reasoning. The manuscript provides no explicit ablation or correlation analysis ruling out confounds such as perception errors propagating into reasoning steps or task artifacts that couple the two. This is load-bearing for the causal claims and the reported gains from the interventions.","section":"§3"}],"minor_comments":[{"comment":"Abstract and results sections: quantitative gains (18.2, 6.0, 3.2) are stated without accompanying error bars, number of runs, or statistical tests; these should be added for reproducibility.","section":"Abstract and §5"},{"comment":"Notation for the perception-aware reward and the surrogate reward should be defined explicitly with equations rather than prose descriptions.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback and for recognizing the methodological strengths of the controlled synthetic setup and the SFT-vs-RL distinction. We address the single major comment below.","responses":[{"response":"We agree that explicit validation strengthens the causal attribution. The tasks were constructed so that perception (object identification in isolated visual queries) and reasoning (relational inference over provided symbolic inputs) operate on disjoint token sequences and evaluation metrics, with no shared intermediate states that would induce propagation. Nevertheless, the manuscript does not currently include the requested correlation or ablation analyses. We will add them in revision: (i) Pearson correlation between per-example perception and reasoning accuracies across checkpoints, and (ii) a controlled error-injection ablation that perturbs only perception outputs while holding reasoning inputs fixed, quantifying downstream reasoning degradation. These additions will directly test for residual coupling.","revision_made":"yes","referee_comment":"[§3] §3 (Synthetic Tasks): The central attribution of the asymmetry to token imbalance (SFT) or reward coupling (RL) rests on the claim that the two synthetic tasks cleanly disentangle perception from reasoning. The manuscript provides no explicit ablation or correlation analysis ruling out confounds such as perception errors propagating into reasoning steps or task artifacts that couple the two. This is load-bearing for the causal claims and the reported gains from the interventions."}],"tokens_in":1335,"tokens_out":297,"duration_ms":18536,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core takeaway is that SFT and RL both widen the gap between perception and reasoning in vision-language models, but for different reasons that the authors can measure and partially correct. They build two synthetic tasks meant to separate the two skills, then track how each training stage affects them. On SFT the fix is reweighting the loss to give perception tokens more signal, which lifts end-to-end accuracy by 18 points on their benchmark. On RL they add a perception-aware term to the reward and get 6 points; even a surrogate version without ground truth still adds 3 points.\n\nWhat stands out is the controlled setup itself. The tasks let them isolate token counts in SFT and reward correlations in RL, and the interventions follow directly from those measurements. That is more concrete than the usual “we added more data” papers.\n\nThe soft spot is still the synthetic tasks. If perception errors bleed into the reasoning steps or if the tasks contain hidden shortcuts, the measured asymmetry and the size of the fixes could be overstated. The abstract reports the gains without error bars or comparisons to strong non-synthetic baselines, so it is hard to judge how much carries over to real images and questions. The mechanisms are plausible, but the evidence is still tied to the authors’ own testbed.\n\nThis is worth a serious referee for groups working on VLM post-training. Anyone trying to improve visual reasoning will want to see whether the reweighting and perception rewards hold up on standard benchmarks. I would send it out rather than desk-reject.","headline":"The paper gives a clean diagnostic for why post-training helps reasoning more than perception in VLMs and shows workable fixes on synthetic tasks.","tokens_in":2343,"tokens_out":385,"would_cite":false,"duration_ms":14110,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Post-training of vision-language models improves reasoning more than perception because of token imbalance in SFT and reward coupling in RL.","keywords":["vision-language models","post-training","perception-reasoning asymmetry","supervised fine-tuning","reinforcement learning","chain-of-thought","token imbalance","reward coupling"],"falsifier":"If reweighting the loss in SFT or adding a perception-aware reward in RL fails to close the performance gap on the synthetic tasks while leaving reasoning gains intact, the diagnosed mechanisms would be refuted.","tokens_in":2632,"feed_emoji":"","tokens_out":465,"duration_ms":19150,"temperature":0.7,"pith_summary":"The paper sets out to explain why post-training on vision-language models lifts reasoning performance much more than perceptual performance, leaving a bottleneck for tasks that require both. It builds a diagnostic framework of two synthetic tasks that isolate the two abilities and tracks how each training stage affects them separately. The analysis shows the imbalance has different roots in supervised fine-tuning versus reinforcement learning. Targeted changes to the loss weights or the reward signal reduce the gap and raise combined accuracy.","feed_headline":"Post-training improves reasoning more than perception in vision models","feed_subtitle":"Token imbalance during SFT and reward coupling during RL create the gap; loss reweighting and perception rewards close it.","key_machinery":"A controlled diagnostic framework built from two synthetic tasks that separate perception from reasoning.","core_discovery":"Post-training improves reasoning more substantially than perception. For supervised fine-tuning this stems from token imbalance in chain-of-thought supervision, where perception occupies fewer tokens and receives a weaker training signal. Dynamically reweighting the loss mitigates this imbalance and boosts end-to-end performance by up to 18.2. For reinforcement learning the asymmetry arises from reward coupling, where outcome rewards correlate more strongly with reasoning than with perception. Adding a perception-aware reward alleviates the imbalance and improves end-to-end accuracy by up to 6.0; even a reliable surrogate reward yields gains of 3.2 points.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Reasoning outpaces perception after VLM post-training","Token imbalance limits perception in supervised VLM fine-tuning","Outcome rewards favor reasoning over perception in VLM RL","Perception rewards reduce asymmetry in vision-language models"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The two synthetic tasks truly isolate perception from reasoning without introducing other factors that drive the observed asymmetry.","fun_headline_variants_meta":{"raw":{"variants":["Reasoning outpaces perception after VLM post-training","Token imbalance limits perception in supervised VLM fine-tuning","Outcome rewards favor reasoning over perception in VLM RL","Perception rewards reduce asymmetry in vision-language models"]},"model":"grok-4.3","cost_usd":0.005567,"raw_usage":{"total_tokens":2608,"prompt_tokens":709,"num_sources_used":0,"completion_tokens":59,"cost_in_usd_ticks":55665500,"prompt_tokens_details":{"text_tokens":709,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1840,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":709,"tokens_out":59,"duration_ms":15428,"temperature":1.0,"reasoning_tokens":1840,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T07:36:54.721737+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"If reweighting the loss in SFT or adding a perception-aware reward in RL fails to close the performance gap on the synthetic tasks while leaving reasoning gains intact, the diagnosed mechanisms would be refuted.","supporting_citations":[],"review_version":1}