{"id":"71d10a0d-e422-465b-b02c-220ee38949a8","arxiv_id":"2509.01656","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ReVPT uses GRPO reinforcement learning with a cold-start SFT phase to make Qwen2.5-VL models call visual tools, improving perception benchmarks over SFT and text-only RL baselines.","lead":"This paper trains small vision-language models to use four visual tools (detection, depth, edge, zoom) via reinforcement learning, reporting gains on perception benchmarks. The abstract's headline numbers conflict with the main table, and the training set is filtered to the base model's errors, so the results need scrutiny.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on internally inconsistent benchmark numbers: abstract CV-Bench deltas don't match Table 2, and the abstract's SAT result is missing from the evaluation table.","rationale":"I read the paper as an empirical systems contribution: the value is in the recipe (cold-start SFT + GRPO + tool rewards) and the measured gains. The load-bearing evidence is therefore the benchmark table and its consistency with the abstract. The reader's weakest assumption about filtered training data is plausible, but I found a more direct problem: the reported numbers are internally inconsistent. This does not prove the method fails; it means the central claim cannot currently be checked. The correct scientific response is conditional acceptance pending a consistent numerical report. I keep the reader's CONDITIONAL verdict. I agree only partially with the reader's weakest_assumption: the data-filter issue is real, but the most immediate blocker is the evidence table itself.","tokens_in":21841,"tokens_out":18166,"duration_ms":179638,"concrete_test":"Using the released code/checkpoints, reproduce Table 2 with the exact evaluation scripts, add the SAT 1,000-subset score, and recompute the abstract's CV-Bench deltas from the same table. If the reproduced CV-Bench deltas equal 9.03/9.44 and SAT is reported with a clearly disjoint split, the concern is resolved; if the reproduced numbers match Table 2 (e.g., +1.78/+9.14) or SAT cannot be reported, the abstract's headline claim is not supported by the paper's own evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—ReVPT achieves state-of-the-art on SAT, CV-Bench, BLINK, and MMStar—is supported only by the benchmark table, but that table is not self-consistent. The abstract reports CV-Bench improvements of 9.03% (3B) and 9.44% (7B) over instruct models. In Table 2, ReVPT-3B scores 70.43 on CV-Bench versus 68.65 for Qwen2.5-VL-3B-Instruct (+1.78), and ReVPT-7B's row shows +9.14 over Qwen2.5-VL-7B-Instruct, not 9.44. SAT is named in the abstract as a key benchmark, yet Table 2 has no SAT column and Appendix B.1's evaluation list covers 'five benchmarks' (CV-Bench, Blink, Blink-Hard, MMVP, and a 1,000 SAT subset) while Section 3.1 says eight benchmarks are evaluated. If the headline numbers cannot be reproduced from the paper's own table, the SOTA claim is not verifiable; the missing SAT evaluation also leaves open the possibility of train/eval overlap since the RL data are filtered from the SAT training set. This is an internal-evidence problem, not a matter of disagreeing with consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ReVPT, a two-stage training procedure (cold-start SFT followed by GRPO-style RL) that teaches Qwen2.5-VL models to invoke four visual tools (object detection, zoom-in, edge detection, depth estimation) in interleaved reasoning traces. The authors claim state-of-the-art results on perception-heavy benchmarks, including SAT, CV-Bench, BLINK, and MMStar, with ReVPT-3B/7B outperforming instruct models by 9.03%/9.44% on CV-Bench. They also provide ablations on cold-start data construction and tool selection, and release code and data.","tokens_in":22138,"tokens_out":5579,"duration_ms":60113,"significance":"If the reported gains hold, ReVPT would be a useful, low-cost recipe for adding tool-use to small multimodal models; the use of a rule-based binary reward (Section 2.4) is a strength because it avoids learned reward models and reward hacking. The paper is empirical and does not make circular claims: the rewards are based on ground-truth answers, and evaluations are external benchmarks (though the SAT evaluation itself is missing). The main obstacle is that the headline numbers are internally inconsistent and the SAT evidence is absent; until those are fixed, the SOTA claim is not verifiable.","major_comments":[{"comment":"The headline CV-Bench gains are mutually inconsistent. The abstract reports ReVPT-3B and ReVPT-7B outperform instruct models by 9.03% and 9.44% on CV-Bench. Table 2 shows +1.78 for ReVPT-3B (70.43 vs 68.65) and +9.14 for ReVPT-7B (74.11 vs 64.97). Section 3.2 instead states '+8.65% and +9.82%' for CV-Bench, and also cites overall gains of +5.81% and +5.73% while Table 2's AVG delta row shows +6.69 for the 3B model. The same comparison is reported with three different numbers. The central empirical claim is not reproducible from the paper's own tables and must be reconciled.","section":"Abstract / §3.2 / Table 2"},{"comment":"SAT is named in the abstract and introduction as a key benchmark where ReVPT achieves state-of-the-art performance, but Table 2 has no SAT column. Section 3.1 says eight benchmarks are evaluated, while Appendix B.1 says the evaluation covers 'five benchmarks' including a 1,000-question SAT subset. Moreover, the RL training data are filtered from the SAT training set (Section 2.3 and Appendix A), so a held-out SAT evaluation is essential both to support the abstract's claim and to rule out train/eval overlap. Please add the SAT subset results, specify its provenance, and clarify the benchmark count.","section":"§3.1 / Appendix B.1 / Table 2"},{"comment":"The RL dataset is constructed by retaining only SAT and TACO examples that Qwen2.5-VL-7B-Instruct answers incorrectly. Because this selection is tied to one base model's specific error patterns, the reported transfer to held-out benchmarks could be an artifact of self-selected training data rather than a general improvement in visual tool-use. The paper provides no analysis of the filtered distribution or an ablation with unfiltered or differently filtered data. Please add such an ablation or an error-category analysis to justify the claim that the filtering is representative of general perception failures.","section":"Section 2.3 / Appendix A"},{"comment":"The paper states that all results are reported as averages of three runs, but no variance, per-run values, or significance tests are provided anywhere in Table 2 or the text. Several headline differences are small (e.g., ReVPT-3B vs Qwen2.5-VL-3B-Instruct on CV-Bench is +1.78), so without variance information it is impossible to tell whether the claimed improvements are reliable. Please report standard deviations or the three individual run scores.","section":"Section 3.1 / Table 2"}],"minor_comments":[{"comment":"Some case-study annotations are internally inconsistent: Figure 12 lists 'Answer: (E)' while the choices only include (A) and (B), and the helper model boxes (A); Figure 13 lists 'Answer: (E)' but the assistant concludes with \\(\\boxed{C}\\). These errors reduce confidence in the qualitative examples and should be corrected.","section":"Figure 12 / Figure 13"},{"comment":"The figure or its caption contains garbled text ('/uni00000026/...'), apparently a rendering artifact. Please replace it with a clean figure.","section":"Figure 6"},{"comment":"There are several typos and formatting inconsistencies: 'core funtions' in Table 1, inconsistent capitalization of 'ReVPT' (e.g., 'REVPT' in Appendix B.2 and GitHub URL), and malformed LaTeX in the case studies (e.g., '<//think>', 'Let's').","section":"Table 1 / text"},{"comment":"The exact filtering criterion for the cold-start data is stated loosely: Section 2.3 says 'we filter out rollout trajectories generated by the GPT model that contain errors,' while Appendix A says 'filtering out any rollouts that lead to incorrect answers.' Please specify whether this is answer-correctness only or also includes format-based filtering.","section":"Section 2.3 / Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The paper's open-sourcing and straightforward RL recipe are strengths. My main concern is verifiability: the three different CV-Bench deltas and the absent SAT numbers must be reconciled before I can sign off. If the authors add the SAT evaluation, clarify the benchmark count, and report variance or per-run values, I would view the contribution favorably."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a good idea—teaching small VLMs to call perception tools via GRPO instead of SFT—but the evidence as written is unreliable. The abstract's headline numbers for CV-Bench (9.03% and 9.44%) don't appear anywhere in Table 2; the actual deltas are +1.78 and +9.14. That's too large to be a typo. SAT is named in the abstract as a key benchmark, but no SAT scores are reported anywhere, even though the appendix says a 1,000-question SAT subset was evaluated. This also leaves open the possibility of train/eval overlap, since the RL data are filtered from the SAT training set.\n\nWhat's genuinely new here is the application of rule-based RL to visual tool-use with a cold-start SFT phase. The ablations in Tables 3 and 4 are useful—they show that tool hints and TACO data preserve general ability, and that removing object detection or depth hurts specific tasks. The failure-mode analysis (Figure 7) is also honest; it shows tools misleading the model, which is exactly the kind of limitation a technical report should document.\n\nSoft spots beyond the number mismatch: the \"state-of-the-art\" claim is overstated. GPT-4.1 and Gemini-2.0-Flash outperform ReVPT-7B on CV-Bench (76.06 and 76.09 vs 74.11), so if they mean open-source SOTA, they should say that. The training data selection—keeping only questions Qwen2.5-VL-7B-Instruct gets wrong—is a clever way to focus RL on hard cases, but it biases the distribution toward the base model's specific errors; the paper doesn't show that gains transfer to easier or unrelated examples. There are also no error bars, despite claiming three runs, and the code link has no commit hash.\n\nOverall, the approach is worth taking seriously, but the reporting inconsistencies undermine confidence in the headline results. A revision that fixes the abstract/table mismatch, reports the SAT scores (and checks for overlap with training), and adds confidence intervals would substantially strengthen this paper.\n\nI'd send it to peer review—the idea deserves a careful referee—but only with major revisions. For a reading group, it's a maybe; the discussion about RL and tool selection is valuable, but the inconsistent data might distract.","headline":"A promising RL-for-visual-tool-use idea undercut by sloppy reporting—the headline numbers don't match the tables and the key SAT result is missing.","tokens_in":22690,"tokens_out":4037,"would_cite":false,"duration_ms":41167,"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":"ReVPT claims that group-relative reinforcement learning with four visual tools—object detection, zoom-in, edge detection, depth estimation—lets small multimodal models beat supervised fine-tuning and text-only RL on perception-heavy benchma","keywords":["reinforcement learning","visual tool use","GRPO","multimodal language models","visual perception","cold-start fine-tuning","perception benchmarks","depth estimation"],"falsifier":"Run ReVPT's pipeline unchanged but replace the error-filtered RL set with a random sample of SAT and TACO questions the base model answers correctly, keeping everything else fixed. If CV-Bench and BLINK gains disappear or reverse, the claim that RL on hard, error-filtered perception questions produces general tool-using visual reasoning is falsified; the gains would instead trace to the specific filtered distribution.","tokens_in":21727,"feed_emoji":"🔍","tokens_out":7480,"duration_ms":73173,"temperature":0.7,"pith_summary":"The paper tries to establish that reinforcement learning, not more supervised data, is the missing ingredient for teaching multimodal language models to use visual tools. The proposed two-stage method, ReVPT, first cold-starts the model with a small set of verified tool-use traces, then runs GRPO—group-relative policy optimization—on 20k questions the base model answers incorrectly. The reward is binary and rule-based: format plus final-answer correctness, with no learned reward model. On perception-heavy benchmarks (CV-Bench, BLINK, BLINK-Hard, MMVP, MMStar), ReVPT-3B and ReVPT-7B outperform SFT and text-only RL baselines, including 9.03% and 9.44% gains over the instruct models on CV-Bench. If this is right, a small model can acquire generalizable tool-using perception without expensive, carefully filtered trajectory data.","feed_headline":"RL plus four vision tools lifts small vision models by 9 percent","feed_subtitle":"Reward-trained 3B and 7B models beat supervised and text-only baselines on CV-Bench, BLINK, MMVP, and MMStar.","key_machinery":"The load-bearing machinery is the GRPO training loop with rule-based rewards over a tool-augmented rollout. In each rollout the model may answer directly or call one of four tools—open-vocabulary object detection, region zoom-in, edge detection, and monocular depth estimation—and then reason over the rendered tool output. The policy is updated by normalizing each rollout's reward against its group, so the model learns which tools help without a learned reward model. Two stabilizers carry the argument: the cold-start SFT phase that teaches the tool-calling format and prevents tool-use collapse, and the error-filtered 20k-question RL set that provides a dense signal of hard perceptual failures","core_discovery":"The central claim is that rewarding the model for correct final answers while letting it freely choose among object detection, zoom-in, edge detection, and depth estimation is enough to induce strategic tool use. ReVPT uses a binary rule-based reward: +1 if the response matches the required <think>/<tool_call>/<answer> format and the final answer is correct, −1 otherwise; GRPO converts these scores into group-relative advantages and updates the policy toward better tool-selection strategies. A cold-start supervised phase on GPT-4.1-synthesized traces prevents the tool-call collapse observed when RL starts from scratch. The paper reports state-of-the-art results on perception-heavy benchmarks","pith_inferences":["Extension: the same pipeline should work with a different tool suite (e.g., segmentation or OCR); because cold-start data fixes the tool repertoire, swapping tools is a direct way to test whether the RL benefit is tool-agnostic.","Extension: adding a small penalty for tool calls, or a separate reward for correct tool choice when the final answer is wrong, would test whether the binary reward is the binding constraint; the case studies show distinct failure modes the binary reward cannot tell apart.","The paper's discussion implies tool benefit is non-monotonic with model scale: small models gain most, mid-size models gain less, and the largest models may use tools differently. Testing ReVPT on a range of base-model sizes would map where the benefit crosses zero."],"forward_implications":["ReVPT-3B and ReVPT-7B outperform their instruct baselines by 9.03% and 9.44% on CV-Bench, and beat SFT and text-only RL baselines across BLINK, BLINK-Hard, MMVP, and MMStar.","RL changes tool-selection behavior: after training, the model calls tools less often but with a higher share of correct answers, indicating it is learning when tools help.","Object detection and depth estimation are the load-bearing tools: removing object detection costs 5 points on BLINK relation and 12.33 points on MMVP.","Cold-start data composition is decisive: combining synthesized tool-use traces with general TACO questions preserves general capabilities while adding tool-use skill.","Perception-specialized training brings a small general-capability tradeoff on MMMU and MMBench even as perception benchmarks improve substantially."],"supporting_citations":[{"why":"Supplies the GRPO algorithm that ReVPT adapts from language reasoning to visual tool use.","marker":"(Shao et al., 2024b)"},{"why":"Provides the SAT dataset used for error-filtered RL training and as a benchmark.","marker":"(Ray et al., 2024)"},{"why":"Provides the TACO CoTA dataset and the supervised tool-use baseline ReVPT must outperform.","marker":"(Ma et al., 2024)"},{"why":"Defines the Qwen2.5-VL base models that ReVPT fine-tunes.","marker":"(Wang et al., 2024a)"},{"why":"Supplies the Depth Anything v2 tool for depth estimation.","marker":"(Yang et al., 2024)"},{"why":"Supplies the open-vocabulary object detection tool.","marker":"(Fu et al., 2025)"},{"why":"The text-only R1-Zero GRPO baseline and the evidence that binary rule-based rewards induce reasoning.","marker":"(Zhou et al., 2025)"}],"fun_headline_variants":["RL tool selection lifts 3B/7B models 9% on CV-Bench","Reward-trained tool use beats instruct models by 9% on CV-Bench","Small LLMs gaining 9% by letting RL pick vision tools","RL with four vision tools beats instruct by 9% on CV-Bench"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the 20k questions the base model answers incorrectly are representative of visual perception failures in general, so training on them transfers to held-out benchmarks; if that filtered set merely captures the base model's idiosyncratic blind spots, the reported gains could be artifacts of data selection rather than a general tool-use ability.","fun_headline_variants_meta":{"raw":{"variants":["RL tool selection lifts 3B/7B models 9% on CV-Bench","Reward-trained tool use beats instruct models by 9% on CV-Bench","Small LLMs gaining 9% by letting RL pick vision tools","RL with four vision tools beats instruct by 9% on CV-Bench"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001222,"raw_usage":{"total_tokens":4872,"prompt_tokens":767,"completion_tokens":4105,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":4020}},"tokens_in":511,"tokens_out":4105,"duration_ms":34134,"temperature":1.0,"reasoning_tokens":4020,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:19:00.040312+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ReVPT's pipeline unchanged but replace the error-filtered RL set with a random sample of SAT and TACO questions the base model answers correctly, keeping everything else fixed. If CV-Bench and BLINK gains disappear or reverse, the claim that RL on hard, error-filtered perception questions produces general tool-using visual reasoning is falsified; the gains would instead trace to the specific filtered distribution.","supporting_citations":[],"review_version":1}