{"id":"2ceb22d2-1ba9-42c2-86a8-f67625635be1","arxiv_id":"2608.08219","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A combined rule-based plus LLM-judge step reward, optimized with GRPO, improves multi-step visual tool orchestration for video anomaly detection on the new VAD-Tool benchmark.","lead":"An agent trained with process-supervised reinforcement learning learns to chain 12 vision tools (tracking, fall detection, violence detection, and more) to explain video anomalies. On a new benchmark, VAD-Tool, this raises multi-step tool-scheduling accuracy on an 8B model to about 95 percent, up from about 90 percent for supervised fine-tuning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"VTO's headline gain may measure agreement with a single Qwen2.5-VL-generated trajectory rather than improved anomaly detection, because both the reward (Eq. 3) and the Whole metric (Eq. 7) enforce exact match to that generator's path.","rationale":"The reader's weakest assumption identifies ground-truth validity and judge independence. I agree with the direction and concretize it: the threat is not just that annotations may be imperfect, but that the evaluation metric itself is a single-path exact match, so even perfect annotations that are merely non-unique make the benchmark reward one out of many correct behaviors. Because the same generator supplies training annotations, judge preference, and test trajectories, VTO's optimization target and the measurement are aligned with that generator's style. The ablation study (Table 3) shows large sensitivity to reward components, but no control for total training compute or parameters (SFT is LoRA for 1 epoch, while the RL stage is full-parameter for 5 epochs); including an outcome-based RL or continued-SFT control would separate process-supervision effects from additional training. None of this makes the numbers fraudulent or the engineering uninteresting; the reported 95.14% is likely reproducible on this benchmark. The concern is about what the number means. Since the public claim is benchmark-specific tool scheduling accuracy, a conditional acceptance requiring multi-trajectory or independent-judge validation is appropriate; hence my verdict remains unchanged from the reader's CONDITIONAL.","tokens_in":15074,"tokens_out":4972,"duration_ms":48804,"concrete_test":"Sample roughly 200 queries from the 1,841-test set and have two independent annotators (not the authors and not Qwen2.5-VL) enumerate all acceptable tool trajectories and final answers per query. Re-score VTO and SFT Whole accuracy as correct if the response matches any acceptable trajectory. If VTO's 5.4-point advantage shrinks or vanishes under this multi-trajectory scoring, the headline gain is an artifact of exact matching to one generated path. As a secondary check in the same pass, also score final-answer anomaly-detection correctness (true/false) per query, blinded to method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is the 10.2-point Whole accuracy gain on Interrelated Tools (89.74% SFT vs 95.14% VTO, Table 2). That metric (Eq. 7) requires exact equality of every predicted decision, tool name, and parameter string with one annotated trajectory. Those annotations are initially produced by Qwen2.5-VL and only human-reviewed for logical correctness (Sec. 3.3), so they represent one valid path, not the set of all valid paths. VTO's process reward (Eq. 3) also uses exact-match tool/parameter rewards, and the LLM judge is Qwen2.5-VL-72B, the same model family that generated the trajectories (Sec. 4.3.1). The test set is built by the same pipeline. Consequently, VTO is directly optimized, and evaluated, for reproducing Qwen2.5-VL's preferred trajectory style. If alternative tool sequences are correct, the benchmark penalizes them, and the reported accuracy conflates 'matches the annotator's path' with 'correctly detects the anomaly.' This threatens the external-validity part of the claim, not the internal reproducibility of Table 2. A further gap is that no outcome-based RL baseline is reported, so the paper's 'premature termination' narrative is not directly tested, although that is secondary to the headline number.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VTO, a process-supervised reinforcement learning framework for visual tool orchestration in video anomaly detection (VAD). The method first supervised fine-tunes a vision-language model on ReAct-format trajectories, then applies GRPO with a dual reward composed of rule-based exact-match rewards and an LLM-as-a-judge reward. The authors also introduce VAD-Tool, a benchmark with 12 vision tools and over 120k instruction-trajectory pairs, and report that VTO improves whole-response accuracy over SFT, e.g., from 89.74% to 95.14% on interrelated-tool responses with a Qwen3-VL-8B backbone (Table 2).","tokens_in":15370,"tokens_out":4396,"duration_ms":43949,"significance":"If validated, VTO would be a useful contribution to agentic VAD: it provides a deliberately constructed multi-tool benchmark, a concrete two-stage training recipe, and evidence that process-level rewards can improve multi-step tool scheduling. The paper also ships code and data, which is a practical strength. However, the headline result is currently difficult to interpret because the reward and evaluation metrics are anchored to exactly one Qwen2.5-VL-generated trajectory per query, and the LLM judge is from the same model family, so the reported gains may reflect alignment with one annotator path rather than improved anomaly detection. The absence of an outcome-based RL baseline also leaves the paper's core motivation untested.","major_comments":[{"comment":"The evaluation and training are circular relative to the benchmark's ground truth. The ReAct trajectories are generated by Qwen2.5-VL (Section 3.3), the rule-based rewards in Eq. (3) reward exact matches to those same trajectories, the LLM judge in Eq. (4) is Qwen2.5-VL-72B (Section 4.3.1), and the Whole metric in Eq. (7) requires exact equality of every decision, tool name, and parameter string with the same Qwen2.5-VL-generated annotation. The human review described in Section 3.3 checks logical correctness but does not enumerate alternative valid tool sequences. Consequently, the reported 5.4-point Interrelated Whole improvement in Table 2 may measure how faithfully VTO reproduces one specific annotator path rather than how well it detects anomalies. I ask the authors to add an outcome-based evaluation (e.g., human-judged correctness of the final anomaly verdict, independent of the tool path) or a multi-annotation study showing that alternative valid trajectories are not penalized.","section":"Section 3.3, Section 4.3.1, Eq. (3), Eq. (7)"},{"comment":"The paper's central motivation is that standard outcome-based RL causes premature termination, but no outcome-based RL baseline is reported. Without a GRPO baseline using only a final outcome reward (or a coarse completion reward) under the same compute and rollout settings, the experiments cannot establish that process supervision rather than RL in general is responsible for the gains. The ablation in Table 3 removes individual VTO components but never compares against pure outcome-based RL. Please add such a baseline and report it in Table 2 or Table 3.","section":"Section 4.1, Section 5.2, Table 2"},{"comment":"The reward model is underspecified. The paper does not give the functional forms or numerical scales of Rcomp, Rtool, Rparam, Reff, Rlogic, Rrel, and Rcomplete, nor the combination weights in Eq. (2), nor the discount factor gamma used in the cumulative return. These are free parameters that directly determine the optimization landscape, and the ablation results in Table 3 cannot be reproduced or interpreted without them. Please provide the exact reward definitions, coefficient values, and gamma, or point to released code that contains them.","section":"Section 4.3.1, Eq. (2)-(4)"},{"comment":"No variance or significance information is reported for the main accuracy comparisons. The headline numbers (e.g., 89.74% vs. 95.14% for Interrelated Whole) come from a single table with no confidence intervals or multiple-seed results, so it is unclear whether the differences are stable. I request at least three seeds with mean and standard deviation for the main VTO and SFT comparisons, or an explicit statement of why a single run is sufficient.","section":"Table 2"}],"minor_comments":[{"comment":"The abstract's '10.2% absolute accuracy improvement' appears to refer to the All Responses Whole column (86.19% to 96.39%), while the interrelated-tool improvement is 5.4 points; please state explicitly which number is being reported to avoid confusion.","section":"Abstract and Table 2"},{"comment":"The term 'human-in-the-loop' is used for a pipeline where Qwen2.5-VL generates the trajectories and human experts review them; the level of human intervention (e.g., fraction of trajectories edited or rejected) is not quantified, which would help assess the independence of the benchmark from the generator.","section":"Section 3.3"},{"comment":"Figure 4 labels the reward phase as 'Step 3' while Section 4.3 describes Rollout, Reward Calculation, and Policy Optimization; the figure and text should use consistent step numbering.","section":"Figure 4 and Section 4.3"},{"comment":"The 'Prompt' column uses 'Zero-shot' for VTO rows even though VTO is a trained model; consider renaming the column to 'Setting' or 'Paradigm' to avoid implying VTO is inference-only.","section":"Table 2"},{"comment":"The benchmark name is inconsistently hyphenated as 'VAD-Tool' and 'V AD-Tool'; please use a single spelling throughout.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The self-referential setup—ground truth generated by Qwen2.5-VL, rewards and evaluation anchored to those exact trajectories, and an LLM judge from the same model family—is likely to be the main point of contention for other reviewers. I would require an independent outcome-based or human evaluation before considering acceptance. The paper is positioned primarily as an agent-training and benchmark paper, which is within scope for ACM MM, but the VAD-specific claims need the proposed additional experiments to be convincing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on VTO. It's a solid, incremental engineering result: process-supervised RL with a mix of rule-based exact-match rewards and an LLM judge gives a real gain on the authors' new VAD-Tool benchmark, and the benchmark itself is a useful resource (12 vision tools, 120k instruction-trajectory pairs, human review). The comparison against SFT and a 72B model is honest, and the ablations show each reward term contributes. The 5.4-point Whole improvement on interrelated queries is believable as an internal result.\n\nThe soft spot is the one the stress-test flagged: both the reward and the evaluation metric enforce exact match to a single annotated trajectory, and those trajectories were generated by Qwen2.5-VL and only lightly human-reviewed. The judge is Qwen2.5-VL-72B, same family. So the headline gain partly measures how well the policy reproduces Qwen2.5-VL's preferred tool-call style, not necessarily better anomaly detection. If there are several valid paths for a query, the benchmark penalizes correct alternatives. That's a real threat to external validity, though it doesn't undermine the internal comparison between VTO and SFT under the same metric.\n\nMissing baselines: no outcome-based RL baseline, which is odd given the paper's premature-termination narrative. Error bars and statistical tests are also absent. The 95.14% vs 89.74% difference could be from a single run for all we know. These are easy fixes.\n\nOverall, the central claim — process supervision helps tool orchestration on this benchmark — is defensible. The broader claim that this transfers to real VAD quality is not established. The paper is worth a serious referee; I'd ask for the outcome-RL baseline, multiple judges or human evaluation, and an analysis of how many test queries admit multiple valid tool sequences. I'd also want to see the code and data actually released before trusting the numbers.\n\nI'd bring this to a reading group as a case study in benchmark circularity. I'd probably cite the benchmark if I work on tool learning, but with a caveat.","headline":"A useful benchmark and a believable internal result, but the reward/evaluation loop is anchored to one model family's trajectories, so the headline gain partly measures alignment with Qwen2.5-VL rather than better anomaly detection.","tokens_in":15915,"tokens_out":2210,"would_cite":true,"duration_ms":19947,"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":"VTO claims that process-supervised reinforcement learning—rule-based rewards plus a vision-language judge—raises an 8B agent's whole-response accuracy on interrelated video-anomaly tool queries to 95.14%, a 10.2-point gain over SFT.","keywords":["video anomaly detection","tool learning","multimodal agents","process-supervised reinforcement learning","GRPO","LLM-as-a-judge","tool orchestration","VAD-Tool benchmark"],"falsifier":"Have several independent security-video experts annotate the same 200 interrelated test queries with every tool sequence they consider correct, then re-score VTO accepting any expert-valid sequence instead of requiring exact match. If experts frequently propose more than one valid path, the exact-match reward will have marked legitimate alternatives as errors, and the reported 95.14% whole accuracy and the 10.2-point improvement over SFT would shrink under any-of-valid scoring.","tokens_in":14860,"feed_emoji":"🎥","tokens_out":7819,"duration_ms":71152,"temperature":0.7,"pith_summary":"VTO reframes video anomaly detection as a multi-step tool-orchestration problem: an agent must call several specialized vision models in sequence to trace how an incident unfolds. The paper's claim is that supervising each reasoning step rather than only the final answer fixes two failures of prior approaches—supervised fine-tuned agents imitate a fixed path, and outcome-based reinforcement learning agents stop too early. To do this, VTO combines exact rule-based rewards (tool name, parameters, completion, efficiency) with judgments from a large vision-language model scoring logicality, relevance, and completeness, then optimizes the policy with GRPO. On the new VAD-Tool benchmark, the framework reports 95.14% whole-response accuracy for interrelated tool queries on an 8B backbone, an absolute gain of 10.2 points over SFT and higher than a 72B model. If correct, this gives a practical recipe for making small multimodal agents orchestrate complex visual analyses without waiting for bigger models.","feed_headline":"Process-supervised RL lifts video tool scheduling to 95.14%","feed_subtitle":"New step-by-step rewards let an 8B agent beat SFT and 72B models on multi-step video anomaly queries.","key_machinery":"The load-bearing mechanism is the dual-reward Process-Supervised Cognitive Alignment wrapped around Group Relative Policy Optimization (GRPO). At each step the total reward is $r_t = R_{\\text{rule}} + R_{\\text{LLM}}$: $R_{\\text{rule}}$ gives exact-match credit for task completion, tool name, and tool parameters, plus an efficiency penalty for extra steps, while $R_{\\text{LLM}}$ uses a 72B vision-language judge to score logicality, relevance, and completeness of the trajectory. Rewards are accumulated over the trajectory, normalized within a group of eight sampled rollouts to form advantages, and used in a clipped GRPO objective with a KL penalty to the SFT reference model. This mechanism is what explicitly penalizes premature termination: a trajectory that stops after the first anomaly detection receives a low completeness score at every subsequent step.","core_discovery":"The paper argues that the bottleneck in video anomaly detection with tool-using agents is not the tools or the backbone but the reward granularity. Its central discovery, on its own terms, is that process-supervised cognitive alignment—dense step-level rewards combining deterministic rule checks with a foundation-model judge—converts an SFT-trained 8B model from a path mimicker into a reliable multi-step orchestrator. On interrelated queries in VAD-Tool, VTO reports perfect decision and tool-selection accuracy, 95.60% input accuracy, and 95.14% whole-response accuracy, against 89.74% for SFT and 67.40% for a 72B model used zero-shot. The ablations attribute the gain to every reward component: without the task-completion reward accuracy collapses to 29.92%, without parameter-match supervision it falls to 77.66%, and without the LLM judge it drops to 93.70%.","pith_inferences":["The paper leaves implicit that the same rule-plus-judge process reward could transfer to other multi-step visual reasoning settings, such as medical imaging or autonomous driving, where stopping after the first finding is also dangerous.","An unexamined consequence is that exact-match rewards train the agent toward the annotator's single path; if alternative tool sequences are legitimate, a softer any-acceptable-path reward would be needed, and the 10.2-point gap would likely shrink.","A testable extension is to compare the LLM judge's completeness scores against human expert ratings on held-out trajectories; large disagreement would indicate the policy is optimizing judge-pleasing style rather than causal completeness."],"forward_implications":["SFT alone leaves a multi-step tool-orchestration gap; adding process-supervised RL closes most of it, raising interrelated whole-response accuracy from 89.74% to 95.14% on Qwen3-VL-8B.","An 8B parameter agent trained with process rewards can outperform a 72B model used zero-shot or few-shot on interrelated anomaly queries.","Penalizing incomplete causal chains at each step prevents the premature-termination failure of outcome-based RL in cascading-hazard scenarios.","A hierarchy of 12 visual tools, from person re-identification to fire detection, provides a usable action space for explainable video anomaly detection.","Exact tool-name and parameter rewards anchor the agent to real visual observations, reducing hallucinated frame or coordinate grounding."],"supporting_citations":[{"why":"Supplies the GRPO optimization objective that VTO uses for policy updates.","marker":"[30]"},{"why":"Defines the ReAct thought-action-observation format used for trajectories and SFT initialization.","marker":"[44]"},{"why":"Motivates the hierarchical taxonomy of vision tools and the scheduling view of video IoT analysis.","marker":"[55]"},{"why":"Is the agentic video anomaly detection system whose static tool-use limitation VTO targets as a baseline.","marker":"[43]"},{"why":"Provides retrospective evidence that SFT-based tool-augmented LLMs fail to adapt dynamically.","marker":"[2]"},{"why":"Motivates fine-grained per-tool-call rewards that VTO extends with an LLM judge.","marker":"[23]"},{"why":"Is the Qwen3-VL-8B backbone on which the main reported result is achieved.","marker":"[1]"},{"why":"Is the Llama-3-8B backbone used in the second VTO variant and comparison.","marker":"[33]"}],"fun_headline_variants":["Video anomaly detection: step-wise RL boosts tool orchestration","Process rewards beat outcome RL for video tool agents","VTO: fine-grained RL for video anomaly tool scheduling","Step-level cognitive rewards sharpen video anomaly reasoning","8B agent outdoes 72B with process-supervised RL on VAD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every benchmark query has exactly one correct tool-use trajectory and that the human-checked Qwen-generated annotations plus the 72B judge's scores capture what good reasoning is; if alternative valid sequences exist or the judge is biased, the accuracy numbers measure agreement with the annotator rather than anomaly-detection skill.","fun_headline_variants_meta":{"raw":{"variants":["Video anomaly detection: step-wise RL boosts tool orchestration","Process rewards beat outcome RL for video tool agents","VTO: fine-grained RL for video anomaly tool scheduling","Step-level cognitive rewards sharpen video anomaly reasoning","8B agent outdoes 72B with process-supervised RL on VAD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000669,"raw_usage":{"total_tokens":3077,"prompt_tokens":997,"completion_tokens":2080,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":1999}},"tokens_in":613,"tokens_out":2080,"duration_ms":15872,"temperature":1.0,"reasoning_tokens":1999,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:15:16.294626+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have several independent security-video experts annotate the same 200 interrelated test queries with every tool sequence they consider correct, then re-score VTO accepting any expert-valid sequence instead of requiring exact match. If experts frequently propose more than one valid path, the exact-match reward will have marked legitimate alternatives as errors, and the reported 95.14% whole accuracy and the 10.2-point improvement over SFT would shrink under any-of-valid scoring.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the ReAct thought-action-observation format used for trajectories and SFT initialization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the hierarchical taxonomy of vision tools and the scheduling view of video IoT analysis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the agentic video anomaly detection system whose static tool-use limitation VTO targets as a baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides retrospective evidence that SFT-based tool-augmented LLMs fail to adapt dynamically."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates fine-grained per-tool-call rewards that VTO extends with an LLM judge."}],"review_version":1}