{"id":"bd2a5327-fc0a-46f1-bcc6-a904e9241331","arxiv_id":"2505.16192","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"VLM-R3 uses reinforcement learning to teach a multimodal model when and where to crop an image during its chain-of-thought, improving scores on several visual reasoning benchmarks.","lead":"This paper trains a vision-language model to pause during reasoning, crop and zoom into image regions, and then continue thinking with the new visual details. The goal is to reduce errors on visual questions that need fine-grained, location-specific information.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Train/eval overlap is probably broader than DocVQA: M3CoT is a composite benchmark that can carry ScienceQA-derived items into R-GRPO, threatening the largest reported gain (87.9 vs 73.6).","rationale":"I tried to find a more internal weakness in the method itself. The mechanism is plausible: the ablation removing interleaved images shows drops, and the grounding-accuracy experiment in Section 4.5.1 shows sensitivity to region quality. These do not, however, rule out training-set contamination. The reader's DocVQA concern is correct as far as it goes; my read extends it because the highest-impact result (ScienceQA) can be reached through the unexamined M3CoT sampling route, making the overlap concern more load-bearing than the reader's narrower DocVQA instance. The verdict remains CONDITIONAL: the paper should be accepted only if the authors provide split provenance, release the relevant sample IDs, and report scores on a disjoint held-out set. This is not a rejection because the concern is resolvable and the other benchmarks provide partial support for the core idea.","tokens_in":18694,"tokens_out":6348,"duration_ms":51424,"concrete_test":"Obtain the exact instance IDs and split provenance for the VLIR corpus and the ~5,000 R-GRPO samples. Intersect them with the ScienceQA and DocVQA evaluation sets used in Table 1, and recompute all table entries after removing every overlapping instance. If the ScienceQA gap (87.9 vs 73.6) or the DocVQA gap (96.8 vs 95.7) shrinks beyond noise, the generalization claim fails; also rerun on a fully disjoint benchmark such as MathVista test to confirm the remaining gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that R-GRPO teaches region selection that generalizes—requires the evaluation benchmarks to be disjoint from the data used at either training stage. The paper does not establish this. Section 3.1.1 lists DocVQA as a VLIR source and Section A.3 lists DocVQA as an R-GRPO sampling source, yet Table 1 reports DocVQA as a benchmark without stating that the evaluation split was excluded. More importantly, Section A.3 also samples from M3CoT [9], a composite multi-domain benchmark that inherits items from existing VQA/CoT datasets, so ScienceQA instances can enter the R-GRPO stage through that route. The paper never discloses the composition of the M3CoT sample or the splits used. Since ScienceQA contributes the single largest gain (87.9 vs 73.6) and is the empirical keystone of the 'new state of the art' claim, the possibility of train/eval leakage through either DocVQA or M3CoT is the most load-bearing unresolved issue. If the gains on ScienceQA and DocVQA are partly memorization, the generalization claim collapses to smaller margins on MathVista/MMMU, and the headline SOTA claim is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VLM-R3, a framework built on Qwen2.5-VL 7B that performs interleaved textual and visual chain-of-thought reasoning. During inference, the model can emit crop/zoom commands (as JSON bounding boxes), after which the cropped image tokens are injected back into the context and the model continues reasoning. The method has two training stages: supervised fine-tuning on a newly constructed VLIR dataset of interleaved rationales, and a reinforcement learning stage called R-GRPO that rewards accuracy, format compliance, region validity, and reasoning length. Experiments compare VLM-R3 against open- and closed-source models on MathVista, MathVision, MMMU, MME, ScienceQA, DocVQA, and HallusionBench, reporting gains over the base model on all benchmarks and an overall claim of state-of-the-art performance.","tokens_in":18947,"tokens_out":6335,"duration_ms":50957,"significance":"If the reported results are reliable, the work addresses a genuine and timely limitation of multimodal LLMs: the inability to dynamically re-focus on specific image regions during long reasoning chains. The design of masking environment-injected image tokens in the policy gradient is a thoughtful and sensible detail, and the ablation study attempts to attribute gains to the interleaved framework, VLIR fine-tuning, and R-GRPO. The region-grounding accuracy analysis in Section 4.5.1 also provides useful evidence that the model's performance depends on the quality of grounding. However, the significance is currently conditional on resolving train/evaluation overlap concerns, the correctness of the R-GRPO objective as written, and the lack of statistical reliability measures. The reported gains on ScienceQA and DocVQA are the empirical keystone of the claims, and both are threatened by possible leakage through the training data sources.","major_comments":[{"comment":"The paper does not establish that the evaluation benchmarks are disjoint from the data used in either training stage. DocVQA appears as a VLIR training source (1,497 samples in Section 4.2), as an R-GRPO sampling source (Section A.3), and as an evaluation benchmark (Table 1). The paper never states that the DocVQA evaluation split was excluded from both training stages, so the reported 96.8% is uninterpretable as a generalization result unless such a split is explicitly documented. In addition, Section A.3 lists M3CoT as an R-GRPO source; M3CoT is a composite benchmark that inherits items from existing datasets, potentially including ScienceQA-derived questions. Since ScienceQA shows the largest gain (87.9 vs. 73.6), the authors must disclose the exact composition of the M3CoT sample and verify that no evaluation item from ScienceQA or any other reported benchmark was present in VLIR or R-GRPO. Without this disjointness guarantee, the central generalization claim is not supported.","section":"Sections 3.1.1, 4.2, A.3, Table 1"},{"comment":"The importance-sampling ratio in the R-GRPO objective is written as πθ(ci|Q) / πθ(ci|Q)|no grad, which is identically 1 and therefore makes the first term in the loss constant. As written, this cannot provide the intended advantage-weighted policy gradient. The ratio should presumably be πθ/πθ_old or involve the reference policy πref. Please correct the equation and clarify how the 'no grad' annotation is meant to be interpreted. This is a load-bearing technical detail of the R-GRPO contribution.","section":"Section 3.3, Eq. (2)"},{"comment":"The claim of 'state of the art in zero-shot and few-shot settings' is not supported by the experimental protocol. All main results in Table 1 are obtained after supervised fine-tuning on VLIR and subsequent R-GRPO training; no zero-shot or few-shot evaluation procedure is described anywhere in the paper. Furthermore, Table 1 itself shows that Gemini-2 Flash and o1 exceed VLM-R3 on MathVista, MMMU, and DocVQA. The global 'new state of the art' wording is therefore inaccurate unless it is explicitly restricted to open-source reasoning MLLMs, and the zero-shot/few-shot claim should either be removed or supported by dedicated experiments.","section":"Abstract and Section 1"},{"comment":"All accuracy numbers are single-run point estimates without error bars, confidence intervals, or significance tests. Several of the reported improvements over the base model are small (MathVista +2.2, DocVQA +1.1, HallusionBench +0.7), so it is impossible to assess whether these differences are stable or within run-to-run noise. The authors should report results over multiple seeds (at least three) with means and variances, and state the decoding protocol (e.g., greedy vs. sampling) for each benchmark.","section":"Section 4.3, Tables 1 and 2"}],"minor_comments":[{"comment":"The benchmark is referred to as 'DocQA' in Section 4.1 but 'DocVQA' in Table 1 and elsewhere; please unify the naming.","section":"Section 4.1 and Table 1"},{"comment":"The parenthetical deltas in Table 2 are relative to the 'Full VLM-R3' row, not to the 'Base Model' row, but this is not stated. The Avg column (e.g., 74.4 for 'w/o Interleaved Chain-of-Thought' vs. 74.0 for 'Base Model') can otherwise appear contradictory. Please add a note explaining the reference for the deltas.","section":"Table 2"},{"comment":"The text says 'As shown in the lower portion of Figure 4' when discussing the attention comparison; the correct reference is Figure 5.","section":"Section 4.5.2"},{"comment":"The paper describes VLIR as a 'benchmark dataset,' but it is used only for training and no evaluation split is defined. Please clarify that VLIR is a training corpus, and consider releasing it with clear train/validation splits to support reproducibility.","section":"Section 1 and 3.1"},{"comment":"No code, data, or model checkpoints are released. Given that the method is data- and RL-heavy, the lack of release plans makes the exact filtering, training, and evaluation pipeline difficult to reproduce. A supplementary reproducibility statement or release plan would strengthen the paper.","section":"A.3 and overall reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The paper comes from a strong group and the core idea is plausible and timely. However, the train/evaluation overlap issue is the first thing the community will check, and it is currently unresolved both for DocVQA and for the M3CoT path into R-GRPO that may affect ScienceQA. The authors should be asked to provide explicit split-level guarantees, correct the R-GRPO equation, and moderate the SOTA wording. If the disjointness cannot be established, the ScienceQA and DocVQA results should be removed or re-run on genuinely held-out data. I see this as fixable within revision and therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: there is a real idea here, but the evaluation is not clean enough to trust as written. The core of the paper—make the model emit crop commands, inject the cropped image back into the context, and mask gradients for those injected image tokens—is a sensible and novel adaptation of GRPO. The ablations support that each component contributes, and the gains on MathVista, MathVision, and MMMU are on benchmarks not listed among the training sources. The VLIR dataset, if released, would be a useful resource. So this is not a throwaway.\n\nThe problem is benchmark hygiene. DocVQA is explicitly a training source: Section 3.1.1 lists it in VLIR construction and A.3 lists it in R-GRPO sampling, yet Table 1 reports it as a test benchmark with no statement about splits. That row should not be there. The bigger concern is ScienceQA, where the paper claims the largest gain (87.9 vs 73.6). R-GRPO sampling uses M3CoT, a composite benchmark that commonly inherits items from existing datasets, and ScienceQA is one of the usual constituents. The stress-test note is right: the paper does not disclose the composition of the M3CoT sample or which split was used. That makes the headline \"state of the art\" claim load-bearing on an unresolved overlap. If part of the ScienceQA gain is memorization, what remains is smaller margins on MathVista/MMMU—still positive, but not the story the abstract tells.\n\nOther soft spots are minor in comparison: no error bars, no code or data release, and no random-crop baseline. A random-crop control would isolate whether the benefit comes from selecting the region or merely from having an extra view. The region-validity reward also mirrors the VLIR filtering criterion; as a training mechanism that is fine, but it should be acknowledged.\n\nMy take: this deserves a serious referee, not a desk reject. The method is substantive and the masking adaptation is correct in principle. But the authors need to re-evaluate on truly held-out benchmarks, state exactly what was in the M3CoT sample, and release code/data before I would trust the numbers. As written, I would treat the ScienceQA and DocVQA rows as unverified.","headline":"Real training idea, shaky evaluation: DocVQA overlap is explicit, ScienceQA gains are plausibly contaminated via M3CoT, yet the core R-GRPO method still deserves a careful referee.","tokens_in":19493,"tokens_out":3981,"would_cite":false,"duration_ms":32561,"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":"VLM-R3 claims that a vision-language model which interleaves textual reasoning with region crops and zoom-ins outperforms its base model on every benchmark tested, with the largest gain on ScienceQA (87.9 vs. 73.6).","keywords":["multimodal chain-of-thought","region grounding","reinforcement learning","GRPO","visual reasoning","crop and zoom","interleaved reasoning","VQA"],"falsifier":"Re-run DocVQA evaluation on the subset of test examples that were never used to build VLIR rationales or R-GRPO rollouts; if accuracy drops materially below 96.8, the reported DocVQA gain is training exposure, not generalization. A second check: fix all bounding boxes to a random region of the same size; if ScienceQA stays near 87.9, region quality is not actually what the reward is optimizing.","tokens_in":18469,"feed_emoji":"🔍","tokens_out":6110,"duration_ms":49673,"temperature":0.7,"pith_summary":"The paper sets out to show that complex visual reasoning fails when a model looks at the image once and then reasons only in text. Its proposed fix is an interleaved chain-of-thought in which the model can write text, emit a crop command with a bounding box, receive the zoomed-in sub-image, and continue. Training is bootstrapped on a curated corpus of interleaved rationales (VLIR) and then optimized by R-GRPO, a reinforcement policy that rewards correct answers, well-formed region commands, and non-redundant reasoning. On the reported benchmarks the trained 7B model beats its base model everywhere, and the largest gain is on ScienceQA (87.9% vs. 73.6%). A sympathetic reader would take this as evidence that dynamic, iterative grounding is a practical way to get finer-grained visual reasoning without scaling the model.","feed_headline":"Crop-and-zoom reasoning lifts ScienceQA by 14 points","feed_subtitle":"A 7B vision-language model interleaves text reasoning with region crops; R-GRPO rewards picking the right place to look.","key_machinery":"The load-bearing object is R-GRPO, a variant of group-relative policy optimization that treats the model's own tokens and bounding-box commands as actions and the injected cropped images as environmental state updates; gradients are masked on image tokens so the policy is rewarded for choosing informative regions, not for generating pixels. The reward mixes a terminal accuracy reward, a format reward, a region-validity reward capped at 0.5 per episode, and a small reasoning-length reward, which together encourage the model to look again only when it helps. The companion VLIR corpus cold-starts the policy with step-level demonstrations, and a piecewise zoom-scaling rule maps the selected bounding-box area to a zoom factor between 1.0 and 2.0. The paper's grounding-accuracy sweep, which degrades performance as bounding boxes are perturbed, is the direct evidence that the selected region, not the reasoning text alone, carries the benefit.","core_discovery":"The central claim is that region recognition and region reasoning belong inside the generation loop, not before it. VLM-R3 is the paper's demonstration: a 7B vision-language model whose output stream can switch between natural language and structured crop commands, with the environment injecting the requested visual detail so the model can revise its reasoning mid-chain. The paper argues this yields a new state of the art in zero-shot and few-shot settings for fine-grained visually grounded inference, with consistent gains across MathVista, MathVision, MMMU, MME, ScienceQA, DocVQA, and HallusionBench. The mechanism it credits is not the extra text but the extra visual evidence: the ablation that removes the interleaved image while keeping the localized text drops sharply on fine-grained benchmarks such as ScienceQA.","pith_inferences":["Editorial extension: the paper does not fully separate the DocVQA benchmark from the DocVQA samples used to build VLIR and R-GRPO, so the reported DocVQA score of 96.8 should be read as an upper bound until a clean split is verified.","Editorial extension: one testable prediction of the grounding-sensitivity result is that an oracle region proposal (perfect bounding boxes) would lift scores further; the paper's sweep from 40% to 90% grounding accuracy implies a monotone continuation toward the oracle.","Editorial extension: because the region-validity reward is syntactic and non-redundancy based, a future comparison could measure whether the model's chosen regions match human-marked informative regions, which would test whether the reward is selecting semantically useful evidence or merely well-formed boxes."],"forward_implications":["If the claim is correct, interleaved region grounding is a transferable recipe: the same crop-command loop should work for chart reading, OCR-heavy documents, and spatial-relation questions without scaling parameters.","The masking of injected image tokens in the policy gradient is a general trick for any tool-augmented language model whose environment returns images.","The ablation results imply that textual descriptions of regions are a poor substitute for the actual cropped pixels, so future systems should keep image evidence in the context window.","The grounding-accuracy sensitivity result implies that improving region proposal quality, not just the reasoning text, is a direct lever on final accuracy."],"supporting_citations":[{"why":"Supplies the Qwen2.5-VL 7B base architecture and dynamic-resolution image encoding that the whole pipeline builds on.","marker":"[8]"},{"why":"Provides the GRPO policy-optimization framework that R-GRPO adapts with region-conditioned rewards and gradient masking.","marker":"[45]"},{"why":"Introduces the prior bounding-box-in-chain-of-thought paradigm that VLIR extends with multi-step, model-generated crops.","marker":"[44]"},{"why":"A representative CoT-based multimodal reasoning baseline whose fixed one-pass visual grounding the paper contrasts with its own.","marker":"[57]"},{"why":"ScienceQA, the benchmark showing the largest reported gain and the main evidence for fine-grained visual-cue reasoning.","marker":"[30]"},{"why":"MathVista, the visual-math benchmark used to demonstrate the method's gains on spatial and visual-detail reasoning.","marker":"[29]"},{"why":"DocVQA, both an evaluation benchmark and a source of VLIR/R-GRPO training data, making the train/test split the key caveat.","marker":"[32]"}],"fun_headline_variants":["Interleaved crop commands lift ScienceQA by 14 points","Mid-chain zoom-in: RL rewards region selection for visual CoT","VLM-R3: region recognition inside the CoT loop","Zero-shot visual reasoning gains from region-conditioned RL","Crop-and-zoom CoT: new SOTA on fine-grained VQA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that evaluation benchmarks, especially DocVQA, are cleanly separated from the data used to build VLIR and to sample R-GRPO rollouts, so the reported scores measure generalization rather than training exposure.","fun_headline_variants_meta":{"raw":{"variants":["Interleaved crop commands lift ScienceQA by 14 points","Mid-chain zoom-in: RL rewards region selection for visual CoT","VLM-R3: region recognition inside the CoT loop","Zero-shot visual reasoning gains from region-conditioned RL","Crop-and-zoom CoT: new SOTA on fine-grained VQA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000871,"raw_usage":{"total_tokens":3807,"prompt_tokens":1013,"completion_tokens":2794,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":2705}},"tokens_in":629,"tokens_out":2794,"duration_ms":18956,"temperature":1.0,"reasoning_tokens":2705,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:05:08.715936+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run DocVQA evaluation on the subset of test examples that were never used to build VLIR rationales or R-GRPO rollouts; if accuracy drops materially below 96.8, the reported DocVQA gain is training exposure, not generalization. A second check: fix all bounding boxes to a random region of the same size; if ScienceQA stays near 87.9, region quality is not actually what the reward is optimizing.","supporting_citations":[{"cited_title":"Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning, 2024","cited_arxiv_id":null,"evidence_quote":"Introduces the prior bounding-box-in-chain-of-thought paradigm that VLIR extends with multi-step, model-generated crops."}],"review_version":1}