{"id":"e328bc6d-dabd-4a85-b50e-d541d66cd8d8","arxiv_id":"2606.08719","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Imagine-OPD internalizes 'thinking with images' via on-policy self-distillation with privileged teacher views, yielding top average performance and lower overhead on vision benchmarks.","lead":"The paper proposes Imagine-OPD, an on-policy self-distillation method that trains models to internally imagine visual cues from zoomed regions instead of calling external tools. This aims to retain the reasoning gains of explicit image manipulation while cutting inference cost.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Privileged supervision via annotated regions may fail to produce faithful, generalizable internal imagination trajectories at inference.","rationale":"The reader's weakest_assumption matches the load-bearing point exactly; the abstract-only review already flags the precise transfer assumption that would need empirical support in the full experiments for the performance claim to be secure.","tokens_in":1698,"tokens_out":270,"duration_ms":15394,"concrete_test":"On a held-out validation split, compute the pixel-level or feature-level overlap (e.g., IoU or cosine similarity) between the model's generated imagination trajectories and the ground-truth annotated regions used for the teacher; if overlap is low (<0.4) yet task performance remains high, the internalization claim does not hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that on-policy self-distillation from a teacher using privileged zoomed views (derived from annotations) causes the student to internalize accurate visual cue imagination without any external teacher or annotations at test time. This is the weakest link because the abstract provides no mechanism ensuring the imagined trajectories are faithful to the privileged evidence rather than learning a spurious correlation or benefiting from regularization effects alone; the 'no external teacher' phrasing holds only if annotations count as non-external, which is not obviously true for generalization.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper introduces Imagine-OPD, an on-policy self-distillation framework to internalize the benefits of 'Thinking with Images' (explicit zooming into relevant regions) into 'Thinking with Imagination' (internal decision of where to look and what visual cues would be revealed). A teacher model receives privileged zoomed evidence views derived from annotated regions and supervises the student's imagination trajectories during training. The method is presented as not requiring an external teacher or high-quality demonstrations, and experiments on vision-centric benchmarks are claimed to show the best average performance among compared models while significantly reducing inference overhead relative to explicit tool-use methods.","tokens_in":1783,"tokens_out":417,"duration_ms":21524,"significance":"If the results hold, the work would demonstrate a viable path to efficient visual reasoning by internalizing tool-use behaviors, lowering inference cost without sacrificing performance. The on-policy self-distillation setup that avoids external high-quality demonstrations is a positive design choice if the generalization from privileged training signals is shown to be robust.","major_comments":[{"comment":"The central claim that privileged supervision from annotated zoomed views produces faithful internal imagination trajectories that generalize at inference (without external teacher or annotations) is load-bearing but unsupported by any mechanism or analysis in the method description; the abstract provides no evidence that the student learns accurate cues rather than spurious correlations or regularization effects alone.","section":"Method"},{"comment":"The performance claim ('best average performance among compared models') cannot be evaluated because the manuscript text supplies no metrics, baselines, ablations, or statistical details; without these, it is impossible to determine whether gains are attributable to the self-distillation procedure.","section":"Experiments"}],"minor_comments":[{"comment":"The abstract would be clearer if it named the specific vision-centric benchmarks and the magnitude of the inference-overhead reduction (e.g., tokens or latency).","section":null}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the detailed and constructive feedback. We address each major comment below and indicate the revisions we will make to strengthen the manuscript.","responses":[{"response":"We agree that the method description would benefit from explicit analysis demonstrating that the self-distillation produces faithful trajectories rather than spurious correlations. The on-policy self-distillation mechanism transfers the teacher's privileged zoomed views to supervise the student's internal imagination process during training, enabling generalization without external tools at inference. In the revised manuscript we will add a dedicated analysis subsection with qualitative trajectory examples and an ablation isolating the distillation signal from generic regularization effects.","revision_made":"yes","referee_comment":"[Method] The central claim that privileged supervision from annotated zoomed views produces faithful internal imagination trajectories that generalize at inference (without external teacher or annotations) is load-bearing but unsupported by any mechanism or analysis in the method description; the abstract provides no evidence that the student learns accurate cues rather than spurious correlations or regularization effects alone."},{"response":"We acknowledge that the submitted manuscript version does not present the full quantitative results, baselines, ablations, or statistical details in the main text. The experiments were conducted on multiple vision-centric benchmarks and include comparisons showing the claimed average performance and reduced overhead, but these details were omitted from the provided text. We will expand the experiments section with complete tables, ablations, and statistical reporting in the revision.","revision_made":"yes","referee_comment":"[Experiments] The performance claim ('best average performance among compared models') cannot be evaluated because the manuscript text supplies no metrics, baselines, ablations, or statistical details; without these, it is impossible to determine whether gains are attributable to the self-distillation procedure."}],"tokens_in":1357,"tokens_out":377,"duration_ms":17395,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main contribution is Imagine-OPD, which trains a model to imagine zoomed visual cues internally instead of calling external tools at inference. A teacher reasoner gets privileged zoomed views derived from annotations during training and supervises the student's own on-policy imagination trajectories. This setup is presented as not needing an external teacher or high-quality demonstrations.\n\nThe approach is new in its specific combination of on-policy distillation with privileged visual evidence to replace tool calls. It directly targets the issues of redundant invocations and noisy crops in prior thinking-with-images work, and the claim of lower inference overhead while matching or exceeding performance on vision benchmarks is a practical goal.\n\nThe soft spots are clear from the abstract. No experimental details, baselines, metrics, or ablations are given, so the performance claims cannot be checked. The central assumption that supervision from annotated privileged views will produce faithful internal imagination at test time is not backed by any described mechanism, leaving open the possibility that the model learns spurious patterns or benefits from regularization alone. That matches the stress-test concern.\n\nThis paper is aimed at researchers working on efficient multimodal reasoning and tool internalization in vision-language models. Readers interested in reducing runtime costs for fine-grained visual tasks would get the most from seeing whether the numbers hold.\n\nIt deserves peer review because the direction addresses a real efficiency problem and the method is outlined clearly enough for referees to evaluate the experiments and generalization claims.","headline":"The paper tries to internalize explicit image zooming via on-policy self-distillation from privileged annotations, but the abstract leaves the faithfulness of the resulting internal trajectories unproven.","tokens_in":2288,"tokens_out":361,"would_cite":false,"duration_ms":14094,"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":"Models internalize the benefits of zooming into image regions by learning to imagine visual cues through on-policy self-distillation.","keywords":["visual reasoning","self-distillation","on-policy learning","image zooming","internal imagination","computer vision","fine-grained reasoning"],"falsifier":"A test where the Imagine-OPD model shows no improvement or worse performance than non-imagination baselines on fine-grained visual reasoning tasks, or where its internal predictions do not correlate with actual zoomed image contents.","tokens_in":2608,"feed_emoji":"👁️","tokens_out":563,"duration_ms":17310,"temperature":0.7,"pith_summary":"The paper establishes that the reasoning advantages of explicitly zooming into relevant image parts can be captured internally without tool calls at test time. It introduces Imagine-OPD, where a teacher using privileged zoomed views from annotations supervises the student's imagined reasoning paths. This self-distillation requires no external teacher or demo data. A sympathetic reader would care because it promises accurate fine-grained visual reasoning at lower inference cost than methods that repeatedly invoke image tools.","feed_headline":"Models learn to imagine zoomed views instead of calling tools","feed_subtitle":"On-policy self-distillation from privileged annotations cuts inference cost while preserving reasoning accuracy on vision tasks","key_machinery":"Imagine-OPD, the on-policy self-distillation framework that transfers supervision from privileged zoomed evidence views to internal imagination trajectories.","core_discovery":"Imagine-OPD is an on-policy self-distillation framework in which a teacher plays the role of a 'Thinking with Images' reasoner during training: it receives privileged zoomed evidence views derived from annotated regions, and supervises the model's own imagination reasoning trajectories. The resulting model achieves the best average performance among compared models while significantly reducing inference overhead compared with 'Thinking with Images' methods.","pith_inferences":["Similar self-distillation could internalize other tool-using behaviors in agents beyond visual zooming.","Training with privileged information might enable better generalization in multimodal reasoning systems.","Inspection of imagined trajectories could reveal what the model has learned to focus on."],"forward_implications":["Best average performance on vision-centric benchmarks compared to other models.","Significantly lower inference overhead than methods that explicitly think with images.","Internalization works without external teachers or high-quality imagination demonstrations.","Imagined trajectories become faithful to task-relevant visual evidence."],"fun_headline_variants":["Self-distillation internalizes thinking with images","Imagine-OPD replaces tool calls with internal visual cues","On-policy teacher supervision cuts inference overhead","Models learn internal zoomed views from annotated regions"],"cache_read_input_tokens":64,"weakest_assumption_plain":"Supervision from privileged zoomed evidence views derived from annotated regions during training will produce faithful internal imagination trajectories that generalize without an external teacher or high-quality demonstrations.","fun_headline_variants_meta":{"raw":{"variants":["Self-distillation internalizes thinking with images","Imagine-OPD replaces tool calls with internal visual cues","On-policy teacher supervision cuts inference overhead","Models learn internal zoomed views from annotated regions"]},"model":"grok-4.3","cost_usd":0.004997,"raw_usage":{"total_tokens":2357,"prompt_tokens":663,"num_sources_used":0,"completion_tokens":55,"cost_in_usd_ticks":49965500,"prompt_tokens_details":{"text_tokens":663,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1639,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":663,"tokens_out":55,"duration_ms":9264,"temperature":1.0,"reasoning_tokens":1639,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T18:21:19.166285+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A test where the Imagine-OPD model shows no improvement or worse performance than non-imagination baselines on fine-grained visual reasoning tasks, or where its internal predictions do not correlate with actual zoomed image contents.","supporting_citations":[],"review_version":1}