{"id":"392a0294-3400-462f-b042-339401b6ad04","arxiv_id":"2505.21457","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A GRPO-based reinforcement learning framework teaches an MLLM to propose zoom-in regions, improving small-object detection and interactive segmentation under a fixed sensing budget.","lead":"ACTIVE-o3 trains a multimodal language model with reinforcement learning to choose which image regions to zoom into before detecting small objects or refining masks. The method beats simple prompting baselines on several benchmarks, but the experimental setup shares weights between the region selector and the detector, so the gains are not cleanly attributable to better active perception.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Shared-weight GRPO confound leaves the central claim of learned active perception unsupported.","rationale":"The reader's conditional verdict is appropriate. The paper releases code, provides detailed prompts and reward definitions, and includes a decoupling variant with GDINO, which are useful concrete artifacts. However, the central claim that GRPO learns a sensing policy for active perception requires that improvements be attributable to region selection. The shared-weight design undermines this attribution: the same model generates regions and performs detection, and the GRPO update can improve the detection component directly. The ACTIVE-O3+GDINO variant transfers the learned MO to a different MA, but without a GDINO+CoT control it does not show that the RL-trained sensing policy is better than zero-shot prompting. The coverage-based heuristic reward in Appendix B.4 also introduces ground-truth region information, so the 'no explicit region-selection supervision' claim needs an ablation that removes the coverage term. These are empirical gaps rather than logical impossibilities, so a conditional acceptance requiring the proposed control is the right level of scrutiny. I therefore keep the reader's verdict unchanged.","tokens_in":18413,"tokens_out":10858,"duration_ms":129008,"concrete_test":"Evaluate the released ACTIVE-O3 model as sensing policy MO with the original Qwen2.5-VL-7B-Instruct (frozen) as task model MA on LVISsmall, and compare with Qwen2.5-VL-CoT region proposals under the same frozen MA. If ACTIVE-O3's advantage largely disappears, the reported gains are explained by updating the shared task model rather than by learning where to look; if the advantage persists, the sensing-policy attribution is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The formal objective in Eq. (1) assumes MA is fixed and optimizes only MO. In the implementation, MA and MO are the same Qwen2.5-VL-7B-Instruct model, updated by the same GRPO objective (Appendix E.2, Remark D.2). The reported LVIS/SODA gains could therefore come from improving detection on arbitrary crops rather than from learning where to look. The decoupling experiment ACTIVE-O3+GDINO (Table 1) is not compared against GDINO with CoT-proposed regions, so it does not isolate the sensing policy. Additionally, the heuristic reward includes coverage against ground-truth boxes/masks (Appendix B.4), which is itself region-level supervision and conflicts with the claim of learning 'without explicit region-selection supervision.' The central attribution of gains to active perception is untested by the current experiments.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper formalizes MLLM-based active perception as a two-module policy—a sensing model M_O that proposes regions and a task model M_A that executes on crops—and proposes ACTIVE-O3, a GRPO-based reinforcement learning framework with a dual-form reward (heuristic plus task-aware). Experiments are reported on LVIS small/dense grounding, SODA-A/SODA-D small-object detection, and ThinObjects interactive segmentation, with qualitative zero-shot examples on the V* benchmark. The central claim is that ACTIVE-O3 learns efficient, stable region-selection strategies without explicit region-selection supervision, substantially improving active perception over Qwen2.5-VL-CoT under a fixed sensing budget.","tokens_in":18594,"tokens_out":3538,"duration_ms":36540,"significance":"If the attribution of gains to the learned sensing policy were established, ACTIVE-O3 would be a practical contribution to MLLM-based active vision, leveraging pure RL without separate critic or SFT. The paper provides a clean task definition, a modular formulation, a released codebase, and a multi-domain benchmark, which are strengths. However, the current experiments do not exclude an alternative explanation—that improvements come from fine-tuning the shared task/sensing backbone on detection and segmentation rewards—and the claimed absence of explicit region-selection supervision is contradicted by a ground-truth-based coverage reward. The significance of the specific active-perception claim is therefore not yet supported.","major_comments":[{"comment":"The formal objective in Eq. (1) assumes M_A is fixed and optimizes only M_O, but the implementation instantiates both M_A and M_O with the same Qwen2.5-VL-7B-Instruct model and updates it with the GRPO objective in Eq. (2). As a result, the gains in Tables 1 and 2 could come from improving the shared backbone on detection rewards rather than from learning where to look. The decoupling experiment ACTIVE-O3+GDINO in Table 1 replaces M_A with Grounding DINO at test time, but it is not compared against GDINO using CoT-proposed regions, so it does not isolate the sensing policy. Please add a control that trains with fixed or random crops under the same reward, or freezes M_A during RL training.","section":"Section 4.2, Eq. (2), Appendix E.2, Remark D.2"},{"comment":"The abstract promises that ACTIVE-O3 'can also serve as a proxy task for leveraging perception data, further improving performance on benchmarks such as RealWorldQA and MME,' but the manuscript contains no evaluation on RealWorldQA or MME. Either add the missing results or remove this claim from the abstract, since it is presently unsupported.","section":"Abstract and Section 1"},{"comment":"The abstract and Section 1 claim that ACTIVE-O3 learns region selection 'without explicit region-selection supervision,' but the coverage-based reward R_coverage in Appendix B.4 directly uses ground-truth masks and bounding boxes with thresholds θ and δ. This is explicit supervision of the selected regions, and the ablation in Table 4 shows that the combined reward includes this component. Please revise the claim or provide an ablation without R_coverage to show that the method works without it.","section":"Appendix B.4 and Section 1"},{"comment":"The SODA evaluation lowers the IoU threshold to 0.1 for detection metrics, which is non-standard and likely inflates the AP/AR values reported in Table 2. The paper should also report results under the standard COCO IoU range (0.5:0.95) or provide a systematic justification for the threshold and a sensitivity analysis, so that the advertised gains are comparable with prior work.","section":"Appendix E.3 and Table 2"}],"minor_comments":[{"comment":"The word 'signicade' appears to be a typo; the intended word is likely 'sign' or 'barricade.'","section":"Figure 6"},{"comment":"The notation M_A and M_O is used before the mapping to 'task model' and 'sensing model' is stated; a sentence clarifying the correspondence immediately after Eq. (1) would improve readability.","section":"Section 4"},{"comment":"The paper repeatedly refers to 'GPT-o3' without a citation; if this is a real model, a reference is needed, and if it is a hypothetical or renamed model, the text should say so.","section":"Section 1 and Appendix D.4"},{"comment":"The claim of zero-shot transfer to the V* benchmark is supported only by qualitative examples in Figures 1, 10–12; a quantitative table on V* would make the claim verifiable.","section":"Appendix G.1"},{"comment":"The row label 'LVIS + SODA-A + D' is ambiguous; it should be written as 'LVIS + SODA-A + SODA-D' for clarity.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":"The shared-weight confound is the main technical obstacle: the paper's central attribution of gains to active perception is not yet tested, and the missing RealWorldQA/MME results create a credibility gap between the abstract and the body. The SODA IoU-threshold choice and the coverage-reward contradiction further weaken the claims. These issues are fixable with additional control experiments and text revisions, so major revision seems appropriate rather than rejection. I would ask the authors to provide the missing comparisons and to tone down or support the 'without explicit region-selection supervision' statement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The paper gives a clean formalization of active perception for MLLMs and a reusable benchmark, with code. That part is solid and worth knowing. But the central claim—that GRPO trains the model to learn where to look—is not yet supported, because the sensing policy and the task model are the same network updated by the same GRPO pass, and the reward includes a coverage term computed from ground-truth boxes and masks. Gains could come from improving the detector on arbitrary crops, not from better region selection.\n\nWhat is genuinely new: the two-module sensing/action split, the GRPO training for region proposals, the dual-form reward, and the LVIS/SODA/ThinObjects protocol. The paper is the first of the cited GRPO-MLLM works to tackle active perception, and the released artifacts make reimplementation feasible. The improvements over CoT prompting are real and consistent.\n\nThe soft spots are the following. First, Eq. (1) fixes the task model MA, but the implementation trains a shared Qwen2.5-VL-7B for both MA and MO under a single GRPO objective. There is no control with random crops or fixed MA, so you cannot attribute the gains to 'where to look.' The ACTIVE-O3+GDINO variant is a good idea but lacks the comparison to GDINO with CoT-proposed crops, which would isolate the sensing policy. Second, the 'without explicit region-selection supervision' claim is contradicted by the coverage reward in Appendix B.4, which uses GT masks and boxes. That is region-level supervision, even if called heuristic. Third, the abstract promises RealWorldQA and MME results that do not appear in the manuscript. Fourth, the SODA evaluation lowers the IoU threshold to 0.1, trains on the test domains, and the paper reports no error bars. These choices are permissive and need justification or tightening.\n\nNone of these are fatal to the framework's potential. The paper is a useful starting point and deserves a serious referee, but the authors should be asked to add the missing controls and fix the overclaims.","headline":"A useful framework and benchmark for MLLM active perception, but the central claim that RL learns 'where to look' is not yet supported by the experiments.","tokens_in":19165,"tokens_out":4402,"would_cite":true,"duration_ms":46482,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ACTIVE-O3 uses GRPO reinforcement learning to teach a multimodal language model where to zoom in an image, improving small-object detection and interactive segmentation under a fixed sensing budget.","keywords":["active perception","active vision","multimodal large language models","GRPO","reinforcement learning","region proposal","small object detection","interactive segmentation"],"falsifier":"Train ACTIVE-O3 under the same rewards and data but with a fixed or random crop policy instead of the learned sensing policy, and compare downstream $AP$/$AR$ and $mIoU$; if that control matches or beats ACTIVE-O3, the claim that region selection is learned would be refuted.","tokens_in":18213,"feed_emoji":"🔍","tokens_out":7546,"duration_ms":74550,"temperature":0.7,"pith_summary":"ACTIVE-O3 asks whether a multimodal large language model can learn active perception—deciding where to zoom in an image—using only reinforcement learning, without any labels for good regions. The paper argues yes: with Group Relative Policy Optimization (GRPO), a lightweight policy-gradient update that needs no separate critic model, and a reward combining task success with simple heuristics about format, overlap, and region size, a single model learns to propose informative crops. These crops are then fed to a task model, and the whole loop improves small-object grounding on LVIS, small-object detection on SODA-A and SODA-D, and interactive segmentation on ThinObjects under a fixed sensing budget. The authors also show the trained sensing policy transfers zero-shot to fine-grained visual question answering such as the V* benchmark, where it outperforms a chain-of-thought baseline.","feed_headline":"Pure RL teaches multimodal LLMs where to zoom in","feed_subtitle":"Trained only on task rewards, it lifts small-object detection and interactive segmentation under a fixed zoom budget.","key_machinery":"The load-bearing mechanism is the modular sensing–action decomposition combined with Group Relative Policy Optimization (GRPO), a policy-gradient update that estimates advantages from the mean and standard deviation of rewards within a group of sampled responses and therefore needs no critic model. The sensing policy $M_O$ is the same multimodal language model prompted to output up to three bounding boxes; each crop is resized and passed to the task model $M_A$, and the reward is a weighted sum of task-aware terms ($AP+AR$ for detection, $mIoU$ after SAM-based interactive segmentation) and heuristic terms (JSON validity, non-overlap threshold $\\tau=0.3$, area ratio between 0.01 and 0.5, and coverage of ground-truth masks or boxes). Under the fixed-budget objective $\\max_{M_O} \\mathbb{E}\\left[\\sum_{k=1}^K R(M_A(o_k), \\mathcal{I})\\right]$, the reward signal is what transfers downstream utility back to region selection.","core_discovery":"The central claim is that MLLM-based active perception can be formalized as a two-stage policy—a sensing policy $M_O$ proposing up to $K$ rectangular regions from a low-resolution global view, and a task model $M_A$ operating on the crops—and that GRPO, with a dual-form reward, can learn the sensing policy directly from downstream task feedback. The paper demonstrates the claim by training Qwen2.5-VL-7B-Instruct this way and reporting consistent improvements: on LVIS small and dense grounding, $AP_s$ and $AR_s$ rise by +1.0/+2.8 and +2.7/+3.5 over Qwen2.5-VL-CoT; on SODA-A and SODA-D, $AP_s$ improves by +8.5 and +13.0; and on ThinObjects interactive segmentation, $mIoU$ at zoom budget 3 rises from 0.796 to 0.863 while the chain-of-thought baseline falls to 0.561. The paper further claims the learned sensing policy is reusable: paired with Grounding DINO it improves that detector, and it transfers zero-shot to fine-grained VQA without being trained on reasoning or question-answering data.","pith_inferences":["A testable extension would be a control in which crops come from a fixed or random policy under identical rewards; if that control matched ACTIVE-O3, the gains would come from the shared backbone improving on detection and segmentation rather than from learned region selection.","The modular formulation suggests a direct path to embodied active perception: replace the crop action with camera-viewpoint control and keep the same reward structure, which the static-image experiments here do not test.","Measuring accuracy against the number of crops used, rather than only against a fixed budget, would connect this work to compute-adaptive inference, where the model spends resolution only where the task demands."],"forward_implications":["A model trained only on object-level rewards learns to emit spatial proposals, suggesting region-selection skill can emerge without region annotations.","With a fixed sensing budget of three regions, ACTIVE-O3 consistently beats the chain-of-thought prompting baseline on $AP_s$/$AR_s$ for small and dense objects and on $mIoU$ for interactive segmentation.","The same sensing policy can be paired at test time with a different, stronger task model such as Grounding DINO, improving its results and supporting the modularity of the design.","The policy transfers zero-shot to fine-grained VQA such as V*, so detection-oriented active perception training appears to generalize to other tasks that require zooming."],"supporting_citations":[{"why":"Supplies GRPO, the reinforcement-learning optimizer that updates the sensing policy without a critic model.","marker":"[17]"},{"why":"Qwen2.5-VL is the base model used for both the sensing policy and the task model.","marker":"[10]"},{"why":"V* is the benchmark used to demonstrate zero-shot transfer to fine-grained visual question answering.","marker":"[18]"},{"why":"Grounding DINO serves as a non-MLLM baseline and as the alternative task model in ACTIVE-O3+GDINO.","marker":"[39]"},{"why":"LVIS provides the small-object and dense-object grounding benchmark and training images.","marker":"[41]"},{"why":"SODA supplies the domain-specific small-object detection benchmarks for autonomous driving and aerial imagery.","marker":"[42]"},{"why":"SAM provides the interactive segmentation API used to compute the task-aware segmentation reward.","marker":"[43]"},{"why":"ThinObjects provides the fine-grained masks used to evaluate zoom-in interactive segmentation.","marker":"[44]"}],"fun_headline_variants":["RL alone teaches MLLMs to choose zoom regions","No labels for zoom: pure RL trains active perception","MLLMs learn to zoom in via rewards, not supervision","Task rewards alone teach MLLMs where to look","Zero-shot zoom selection from pure reinforcement learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the task model can be treated as fixed while only the sensing policy is learned; in the implementation both roles are played by the same Qwen2.5-VL-7B model updated by GRPO, so gains could in principle come from improving the shared model on detection and segmentation rather than from learning where to look.","fun_headline_variants_meta":{"raw":{"variants":["RL alone teaches MLLMs to choose zoom regions","No labels for zoom: pure RL trains active perception","MLLMs learn to zoom in via rewards, not supervision","Task rewards alone teach MLLMs where to look","Zero-shot zoom selection from pure reinforcement learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000183,"raw_usage":{"total_tokens":1373,"prompt_tokens":1064,"completion_tokens":309,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":680,"completion_tokens_details":{"reasoning_tokens":233}},"tokens_in":680,"tokens_out":309,"duration_ms":4057,"temperature":1.0,"reasoning_tokens":233,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:27:32.603541+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train ACTIVE-O3 under the same rewards and data but with a fixed or random crop policy instead of the learned sensing policy, and compare downstream $AP$/$AR$ and $mIoU$; if that control matches or beats ACTIVE-O3, the claim that region selection is learned would be refuted.","supporting_citations":[{"cited_title":"V?: Guided visual search as a core mechanism in multimodal llms","cited_arxiv_id":null,"evidence_quote":"V* is the benchmark used to demonstrate zero-shot transfer to fine-grained visual question answering."},{"cited_title":"Grounding dino: Marrying dino with grounded pre-training for open-set object detection","cited_arxiv_id":null,"evidence_quote":"Grounding DINO serves as a non-MLLM baseline and as the alternative task model in ACTIVE-O3+GDINO."},{"cited_title":"Lvis: A dataset for large vocabulary instance segmentation","cited_arxiv_id":null,"evidence_quote":"LVIS provides the small-object and dense-object grounding benchmark and training images."},{"cited_title":"Towards large-scale small object detection: Survey and benchmarks.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):13467–13488, 2023","cited_arxiv_id":null,"evidence_quote":"SODA supplies the domain-specific small-object detection benchmarks for autonomous driving and aerial imagery."},{"cited_title":"Deep interactive thin object selection","cited_arxiv_id":null,"evidence_quote":"ThinObjects provides the fine-grained masks used to evaluate zoom-in interactive segmentation."}],"review_version":1}