{"id":"1b8e66b0-7040-43a2-900f-850be1ed5078","arxiv_id":"2506.14821","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A 3B VLM trained with GRPO to call a zoom tool improves V*Bench accuracy by 5.7% over its base model but degrades TextVQA and HR-Bench performance.","lead":"This paper trains a small vision-language model to use a zoom tool via reinforcement learning, improving fine-grained visual accuracy on one benchmark while hurting others. It shows that simple tool interfaces and careful training data selection matter for keeping such models efficient.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No no-tool RL control means the V*Bench gain cannot be attributed to the zoom tool; the causal claim in the abstract is unsupported without that control.","rationale":"The reader's weakest assumption focuses on the resolution gap between training (<=1024 px) and inference (up to 8K) and the resulting poor HR-Bench transfer. That is a legitimate limitation, but the paper explicitly scopes its claim to 'some VQA tasks' and transparently reports no HR-Bench improvement, so the resolution gap does not by itself undermine the V*Bench result. The more load-bearing concern is the missing no-tool RL control: the paper's contribution is not merely that the model improves, but that the tool drives the improvement. Without a control, the +5.7% could arise from GRPO on the hard-data mix or from the answer-format reward, which would weaken the claimed mechanism. This concern does not overturn the empirical result; it identifies a specific experiment needed to support the causal attribution. Since the reader already issued a CONDITIONAL verdict based on missing evidence, the verdict should remain CONDITIONAL, and this stress-test identifies a different, arguably more central missing control.","tokens_in":11015,"tokens_out":3895,"duration_ms":44271,"concrete_test":"Train a no-tool control: identical GRPO setup, same filtered TextVQA training data, same correctness and format rewards, but with the zoom tool removed from the system prompt and no tool results injected. Run for the same 800 steps with the same rollout budget, and evaluate on V*Bench with the same protocol (temperature 0.1, max 2048 tokens). Repeat both the with-tool and no-tool runs with at least three seeds. If the no-tool control reaches roughly 80% V*Bench accuracy, the +5.7% gain is not caused by the zoom tool; if it stays near the 74.4% base-model level, the tool attribution is supported. Report overall and per-subtask numbers for both conditions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the V*Bench improvement comes from the zoom tool: the abstract says the gain is 'thanks to the detailed visual information gathered from the external tool.' However, the paper never reports a control trained with the same GRPO recipe, reward structure, and filtered TextVQA data mix but without the zoom tool. This control is essential because the training recipe contains several interventions that could improve V*Bench independently of tool use: (i) RL fine-tuning on a hard-data mix from TextVQA, (ii) answer-formatting regularization, (iii) the simplified YAML-like tool interface, and (iv) crop upscaling. Figure 3 shows successful tool calls receive higher advantage during training, but that is an in-training signal, not end-task attribution. Figure 4 compares RL-with-zoom against the base model, not against RL-without-zoom. The crop-upscale ablation does show that the tool output representation matters, but both arms still called the tool. Consequently, the observed +5.7% on V*Bench could be a general effect of GRPO on the curated data mix rather than a specific effect of the zoom tool. The resolution-gap issue on HR-Bench, central to the reader's verdict, is a real but separate limitation that the authors explicitly acknowledge and that does not affect the V*Bench claim, since the claim is scoped to 'some VQA tasks.'","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an RL recipe for teaching a 3B vision-language model (Qwen2.5-VL-3B-Instruct) to call a zoom tool before answering visual questions. Training uses GRPO with a reward combining answer correctness, format compliance, and tool-use success, a simplified YAML-like tool interface, a single tool call per query, crop upscaling, and a TextVQA training subset filtered to examples where the base model scores below 0.5. On V*Bench the method reports 80.1% overall accuracy versus 74.4% for the base model (+5.7%), matching Qwen2.5-VL-7B (79.1%) and beating SEAL (75.4%) at the 3B scale. On TextVQA validation the method degrades by 5.9 points, and on HR-Bench 4K/8K it shows no improvement or a 5.7-point drop, which the authors attribute to a training/inference resolution gap.","tokens_in":11359,"tokens_out":6006,"duration_ms":60635,"significance":"If the main result is reproducible, the paper offers a practical and relatively cheap recipe: a 3B model, four A100s, one zoom call, and no tool-use trajectory supervision. The design contains several credible components—masking tool-generated image tokens in the GRPO loss, upscaling the crop, a compact tool interface, and a hard-data mix—and the efficiency comparison against SEAL is favorable. However, the central causal claim that the gain comes from the tool is not currently established, because the recipe changes several things at once and the reported numbers come from a single best run. The paper is a useful systems contribution in its current form, but it needs a control and variance reporting before the attribution can be accepted.","major_comments":[{"comment":"The causal attribution in the abstract ('thanks to the detailed visual information gathered from the external tool') is not supported by the experimental design. The recipe contains at least four interventions that could improve V*Bench without tool access: GRPO on a filtered hard subset of TextVQA, the correctness/format reward structure, the simplified tool-calling interface, and crop upscaling. Figure 4 compares the RL-with-zoom model to the base model, not to an RL-trained model without the zoom tool, and the crop-upscale ablation still calls the tool in both arms. I request a control trained with the same rewards, data mix, and rollouts but with the zoom tool disabled (or replaced by a no-op tool); without it the paper cannot distinguish 'GRPO on this data helps' from 'the zoom tool helps.'","section":"Section 3, Table 1, and Abstract"},{"comment":"The paper states, 'We present the results from the training run that performed the best,' meaning Table 1 reports a single run with no variance information. GRPO training with no KL regularization and clip-higher can be noisy, and selecting the best run can overstate the +5.7% V*Bench gain. Report mean and standard deviation across at least three independent seeds, or otherwise justify why the selected run is representative.","section":"Section 3, Training details"},{"comment":"The statement that downscaled training 'helps the model generalize well on high-resolution visual tasks at inference time' is contradicted by the HR-Bench results in Table 1: HR-Bench 4K shows no improvement over the base model and HR-Bench 8K shows a 5.7-point overall drop. The authors' resolution-gap conjecture in Section 3 is reasonable, but the claim in Section 2 should be scoped to the resolutions actually tested (V*Bench, roughly 2K), or the paper should include high-resolution training to support the broader generalization claim.","section":"Section 2, 'Image resolution,' and Table 1"}],"minor_comments":[{"comment":"The JSON example contains the typo 'boudning box'; it should read 'bounding box'.","section":"Figure 2"},{"comment":"The regular expression in the paragraph beginning 'Once the tool is used' is garbled in the manuscript ('[ˆ<]*)</think>...'); please present it in monospace and verify that it matches the intended grammar.","section":"Section 2, Training"},{"comment":"The table uses 'HR-Bench' and 'HR-bench' inconsistently; choose one capitalization throughout.","section":"Table 1"},{"comment":"The text repeatedly renders 'YAML-like' as 'Y AML-like' (e.g., Figure 2 and Section 2); fix the spacing.","section":"General"},{"comment":"The phrase 'RL substantially improves' should be qualified as 'on V*Bench,' because Table 1 shows decreases on TextVQA and HR-Bench 8K.","section":"Section 3, Results"},{"comment":"Label the x-axis with units (pixels of the long side) and add a marker for the training resolution of 1024 pixels to make the resolution-gap discussion easier to follow.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The requested no-tool RL control and variance reporting are feasible and should be added before publication. The overlap with DeepEyes is acknowledged in the manuscript, and I do not see a novelty concern; the contribution is the resource-constrained recipe and its component analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid, honest engineering contribution: it shows a 3B VLM can be trained with GRPO to call a zoom tool and gain +5.7% on V*Bench, matching the 7B base. The practical insights are real: the YAML-like tool interface, keypoint-based zoom, crop upscaling, and the hard-example data mix. The authors also deserve credit for reporting their failures on HR-Bench and TextVQA and for including the crop-upscale ablation, which shows that how the tool output is represented matters.\n\nThe soft spot is the one the stress-test flags: there is no RL-without-tool control. The training recipe bundles several interventions—GRPO on a hard-example mix, formatting rewards, the simplified interface, crop upscaling—and any of these could improve V*Bench independently of the zoom tool. The paper's abstract says the gain is \"thanks to the detailed visual information gathered from the external tool,\" but the evidence does not rule out a general GRPO-on-curated-data effect. Figure 3 shows successful tool calls get higher advantage during training, which is an in-training signal, not end-task attribution. This is a load-bearing gap for the causal claim, though not for the weaker claim that the whole recipe works on V*Bench.\n\nTwo smaller issues: the paper reports only the best training run with no variance across seeds, which matters for a recipe paper; and code and data are not released, so the engineering insights are hard to verify. The resolution gap on HR-Bench is real but cleanly acknowledged and does not undermine the V*Bench result, since the claim is scoped to \"some VQA tasks.\"\n\nWho is this for? People working on RL for tool use in small VLMs, and anyone building resource-constrained multimodal agents. The recipe is useful, the writing is clear, and the limitations are stated. It deserves a serious referee, but the referee should ask for the no-tool RL control and a variance estimate. I would accept it for peer review, with the expectation of a revision.","headline":"A useful engineering recipe for teaching small VLMs to zoom via GRPO, but the headline causal claim about the tool is not proven without an RL-without-tool control.","tokens_in":11865,"tokens_out":1142,"would_cite":true,"duration_ms":13132,"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":"GRPO and one zoom call lift a 3B VLM by 5.7 points on V*Bench","keywords":["reinforcement learning","vision-language models","tool use","zoom","GRPO","visual question answering","fine-grained visual reasoning","resource efficiency"],"falsifier":"A decisive test is to ablate the learned keypoint selection: run the same GRPO recipe but feed the zoom tool a fixed center or random keypoint. If V*Bench accuracy stays near 80%, the RL policy is not what selects the region and the claim that the model 'learns to zoom where it matters' collapses.","tokens_in":10842,"feed_emoji":"🔍","tokens_out":7423,"duration_ms":65755,"temperature":0.7,"pith_summary":"The paper shows that a small vision-language model can be taught to call a zoom tool through reinforcement learning, closing part of the gap with much larger models on fine-grained visual question answering. Training a 3-billion-parameter model with GRPO, a structured reward, and a deliberately simple tool interface raises its accuracy on V*Bench from 74.4% to 80.1%, matching a 7-billion-parameter baseline and beating a hierarchical visual search system, all with a single targeted zoom per image. The recipe works without any ground-truth tool-use trajectories and under a modest four-GPU budget, but it does not transfer to 4K/8K images, where the model was trained only on downscaled inputs. The paper identifies the training data mix, the reward shape, and the tool-interface simplicity as the decisive factors.","feed_headline":"3B VLM with one zoom call gains 5.7 points on V*Bench","feed_subtitle":"Reinforcement learning teaches a small vision-language model to zoom in on the right spot, matching a 7B baseline.","key_machinery":"The load-bearing mechanism is a GRPO training loop with a one-call zoom tool accepting a single keypoint coordinate, a structured reward $R = \\alpha R_c + \\beta R_f + \\gamma R_t$ mixing a hard VQA score, a soft edit-distance partial credit, a format reward, and a small tool-use reward, plus masking of tool-generated tokens during the policy loss. A YAML-like tool invocation format replaces JSON because the small model can emit it reliably, and the 400x400 crop is upscaled to the input image's longer dimension so the model sees detail at a comparable scale. The training mix filters TextVQA to examples where the base model scores below 0.5 in an 8-shot evaluation, roughly 7,800 of 34,600 examples, so the model is pushed to use the tool where it matters.","core_discovery":"The central claim is that group-relative policy optimization can teach a small VLM to invoke a zoom tool productively for fine-grained visual reasoning, provided the tool interface is easy to emit, the reward separates correctness from formatting from tool success, and the training set over-represents examples the base model fails. The trained 3B model learns to emit a single keypoint coordinate, receives a 400x400 crop upscaled to the input size as extra image tokens, and answers with access to both the original and the crop. On V*Bench this yields 80.1% overall accuracy, a 5.7-point gain over the untrained 3B baseline, roughly matching a 7B model, and beating the SEAL visual search system while using far fewer crop evaluations. The same model does not improve on HR-Bench 4K/8K or on in-domain TextVQA, and the authors trace the failure to the resolution gap between low-resolution training and high-resolution inference.","pith_inferences":["The gain on V*Bench may be partly attributable to the zoom tool providing additional image tokens rather than to the RL policy's keypoint selection; ablating keypoint choice with random or center crops would separate these effects.","A straightforward extension is to train with mixed-resolution inputs or multi-scale augmentation to close the resolution gap; if that recovers HR-Bench performance, the recipe would generalize beyond V*Bench's resolution range.","The one-call zoom limit means the method is suited to single-ROI questions; tasks requiring multiple distinct regions would need iterative calls or a multi-crop tool, which the current interface does not support.","The same GRPO recipe should transfer to other visual tools (e.g., OCR or depth estimators) by swapping the tool description and keeping the keypoint interface, provided the reward still gives dense enough feedback."],"forward_implications":["A small VLM can learn to call an external tool from a reward signal alone, without any supervised tool-use trajectories, under a four-GPU budget.","A single keypoint-based zoom call is enough to match or exceed systems that evaluate multiple crops per query, cutting inference cost.","The paper identifies the simple tool interface, token masking, crop upscaling, and a difficult-example-heavy data mix as the decisive components of the recipe.","The failure to improve on HR-Bench 4K/8K is attributed to the training/inference resolution gap, so gains on V*Bench should not be expected to transfer to higher native resolutions without bridging that gap."],"supporting_citations":[{"why":"Supplies the GRPO algorithm that the training loop is built on.","marker":"Guo et al. 2025"},{"why":"Introduces the GRPO/DeepSeekMath lineage that the method extends to VLMs.","marker":"Shao et al. 2024"},{"why":"Provides the TextVQA dataset used for training and the VQA score used as the hard reward component.","marker":"Singh et al. 2019"},{"why":"Defines V*Bench, the primary out-of-domain benchmark, and the SEAL visual search baseline the method is compared against.","marker":"Wu and Xie 2024"},{"why":"Provides the HR-Bench 4K/8K benchmark that reveals the resolution gap limitation.","marker":"Wang et al. 2025c"},{"why":"Supplies the Qwen2.5-VL base model and its instruction-tuned baselines.","marker":"Bai et al. 2025"}],"fun_headline_variants":["3B VLM learns to zoom, gains 5.7 points on V*Bench","GRPO + zoom: 3B VLM matches 7B on V*Bench","One zoom call lifts a tiny VLM by 5.7 points on V*Bench","Small VLM, big zoom: 80.1% on V*Bench"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a keypoint-selection policy learned on images downsized to at most 1024 pixels on the long side transfers to original images at much higher resolutions at inference time.","fun_headline_variants_meta":{"raw":{"variants":["3B VLM learns to zoom, gains 5.7 points on V*Bench","GRPO + zoom: 3B VLM matches 7B on V*Bench","One zoom call lifts a tiny VLM by 5.7 points on V*Bench","Small VLM, big zoom: 80.1% on V*Bench"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000485,"raw_usage":{"total_tokens":2364,"prompt_tokens":884,"completion_tokens":1480,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":1387}},"tokens_in":500,"tokens_out":1480,"duration_ms":10871,"temperature":1.0,"reasoning_tokens":1387,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:53:57.396570+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test is to ablate the learned keypoint selection: run the same GRPO recipe but feed the zoom tool a fixed center or random keypoint. If V*Bench accuracy stays near 80%, the RL policy is not what selects the region and the claim that the model 'learns to zoom where it matters' collapses.","supporting_citations":[{"cited_title":"Towards vqa models that can read","cited_arxiv_id":null,"evidence_quote":"Provides the TextVQA dataset used for training and the VQA score used as the hard reward component."}],"review_version":1}