{"id":"ff152faf-848c-4bd3-84d4-39b9169c94a2","arxiv_id":"2605.10397","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"AnomalyClaw turns single-step VLM anomaly judgments into a multi-round tool-grounded refutation process, delivering consistent macro-AUROC gains of 3.5-7.9 percentage points over direct inference across 12 cross-domain datasets.","lead":"AnomalyClaw is a training-free agent that improves visual anomaly detection by having vision-language models engage in multi-round refutation of candidate anomalies using a library of 13 tools for verification against normal samples. This approach boosts cross-domain performance of existing VLMs on anomaly tasks without domain-specific training or fine-tuning.","discovery_kind":"unclear","skeptic_critique":{"model":"grok-4.3","headline":"Gains may arise from more VLM calls or context rather than refutation structure","rationale":"The reader's weakest assumption already flags the exact isolation problem. Because the performance numbers constitute the sole quantitative support for the central claim, failure of the proposed control would directly undermine the attribution to 'agentic refutation' rather than generic multi-step prompting. The optional self-evolution extension is downstream and does not address this core uncertainty.","tokens_in":1890,"tokens_out":371,"duration_ms":11872,"concrete_test":"Construct a matched-budget baseline that performs exactly the same average number of VLM calls per sample as AnomalyClaw (measured from the paper's reported round statistics) but replaces the refutation loop and tool invocations with repeated direct or chain-of-thought prompts on the same normal references; recompute macro-AUROC on CrossDomainVAD-12. If the delta versus single-step falls below 2 pp, the refutation mechanism is not the primary driver.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline result claims that the multi-round tool-grounded refutation (13-tool library for verification, reference parsing, and expert probing) produces the reported macro-AUROC lifts (+6.23 pp GPT-5.5, +7.93 pp Seed2.0-lite, +3.52 pp Qwen3.5-VL-27B) over single-step inference by forcing grounding in normal-sample references. No ablation is described that holds total VLM inference count, total context tokens, or number of reasoning steps fixed while removing the refutation loop and tool calls. Without that control, the observed deltas could be explained by longer effective chains or repeated sampling alone, leaving the specific causal role of the refutation process untested.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper introduces AnomalyClaw, a training-free VLM-based agent for cross-domain visual anomaly detection that reformulates judgment as a multi-round refutation process grounded in a 13-tool library for verification, reference parsing, and expert probing. It reports consistent macro-AUROC gains over single-step direct inference on the CrossDomainVAD-12 benchmark (+6.23 pp on GPT-5.5, +7.93 pp on Seed2.0-lite, +3.52 pp on Qwen3.5-VL-27B) and shows that an optional verbalized self-evolution extension yields further gains comparable to a supervised K=10 baseline.","tokens_in":2010,"tokens_out":571,"duration_ms":20892,"significance":"If the gains are shown to arise specifically from the refutation structure rather than increased inference budget, the work would provide evidence that agentic tool-grounded reasoning can improve VLM reliability for cross-domain VAD without any training. The training-free design and the self-evolution mechanism that builds an online rulebook from internal disagreements are clear strengths.","major_comments":[{"comment":"The headline results in the CrossDomainVAD-12 experiments compare AnomalyClaw only against single-step direct inference. No ablation is reported that holds fixed the total number of VLM calls, total context tokens, or number of reasoning steps while removing the refutation loop and tool calls; without this control the observed deltas could be explained by longer chains alone.","section":"Experiments (CrossDomainVAD-12 benchmark results)"},{"comment":"The claim that the 13-tool library enables grounding in normal-sample references and fine-grained evidence is central to the causal story, yet no ablation or sensitivity analysis is provided that isolates the contribution of individual tool categories (visual verification vs. reference parsing vs. expert probing).","section":"Method (tool library description)"}],"minor_comments":[{"comment":"The abstract states that the self-evolution extension delivers +2.09 pp on Qwen3.5-VL-27B, comparable to the K=10 oracle baseline (+1.99 pp), but does not specify whether the comparison uses identical base models, prompt templates, or evaluation protocols.","section":"Abstract"},{"comment":"Notation for the multi-round refutation process (e.g., how candidate anomalies are proposed and refuted across rounds) would benefit from a concise algorithm box or diagram.","section":"Method"}],"recommendation":"major_revision","confidential_remarks":"The absence of any compute-matched baseline is the primary load-bearing gap; once addressed the paper would be a stronger fit for the journal."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive and insightful comments. We address each major point below and commit to revisions that strengthen the empirical support for our claims.","responses":[{"response":"We agree that an explicit control for inference budget is necessary to isolate the contribution of the refutation structure. The current experiments focus on the end-to-end comparison to direct inference, but we will add a new ablation in the revised manuscript. This ablation will match the total number of VLM calls and approximate token budget of AnomalyClaw while replacing the structured refutation loop and tool calls with unstructured repeated direct queries or generic chain-of-thought prompting. The results will clarify whether the observed gains arise from the tool-grounded refutation mechanism rather than increased computation alone.","revision_made":"yes","referee_comment":"The headline results in the CrossDomainVAD-12 experiments compare AnomalyClaw only against single-step direct inference. No ablation is reported that holds fixed the total number of VLM calls, total context tokens, or number of reasoning steps while removing the refutation loop and tool calls; without this control the observed deltas could be explained by longer chains alone."},{"response":"We concur that category-level ablations would provide stronger evidence for the role of each tool group. The manuscript presents the 13-tool library as an integrated system supporting the refutation process. In the revision we will add sensitivity analyses that remove or disable entire categories (visual verification tools, reference parsing tools, and expert probing tools) one at a time and report the resulting macro-AUROC changes across the CrossDomainVAD-12 benchmark. These results will quantify the marginal contribution of each category to the grounding effect.","revision_made":"yes","referee_comment":"The claim that the 13-tool library enables grounding in normal-sample references and fine-grained evidence is central to the causal story, yet no ablation or sensitivity analysis is provided that isolates the contribution of individual tool categories (visual verification vs. reference parsing vs. expert probing)."}],"tokens_in":1532,"tokens_out":443,"duration_ms":18663,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper's main point is that single-pass VLM judgments on anomalies are unreliable because they lean on priors instead of normal-sample references. AnomalyClaw fixes this by turning the task into an iterative refutation loop where the agent proposes candidates and checks them using a library of 13 tools for visual verification, reference parsing, and expert probing. On the CrossDomainVAD-12 benchmark it shows AUROC lifts of 3.5 to 7.9 points over direct inference across three different VLMs, plus a self-evolution extension that builds rules from internal disagreements and nearly matches a supervised baseline without labels. The training-free, cross-domain angle is the practical strength here, and the agent framing is distinct from standard one-shot prompting or classic anomaly pipelines. The benchmark breadth helps too. The soft spot is the missing control. Nothing described holds total VLM calls, context length, or reasoning steps fixed while removing the refutation loop and tool calls, so the deltas could come from longer chains or repeated sampling alone. The abstract also skips details on baseline implementations and statistical tests, which leaves the causal claim only moderately supported. This is for researchers working on VLM applications in industrial inspection, medical imaging, or monitoring where retraining per domain is expensive. A reader interested in agentic methods would find the structure and numbers worth examining. It deserves peer review because the core idea is new enough and the evaluation covers enough ground, even though tighter ablations would strengthen the case.","headline":"AnomalyClaw frames VLM anomaly detection as a multi-round tool-grounded refutation agent and reports AUROC gains on 12 datasets, but the gains may stem from extra calls rather than the refutation structure itself.","tokens_in":2515,"tokens_out":387,"would_cite":false,"duration_ms":49416,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[{"relation":"unclear","rs_module":"IndisputableMonolith/Cost/FunctionalEquation.lean","rs_theorem":"washburn_uniqueness_aczel","paper_passage":"multi-round refutation process... 13-tool library for visual verification, reference parsing, and frozen expert probing"},{"relation":"unclear","rs_module":"IndisputableMonolith/Foundation/ArithmeticFromLogic.lean","rs_theorem":"LogicNat.induction","paper_passage":"verbalized self-evolution extension... builds an online rulebook from internal-branch disagreement"}],"headline":"Applied CV/agentic VLM refutation loop with no overlap to RS cost or forcing structures","alignment":"orthogonal","rationale":"Paper centers on a training-free multi-turn refutation agent using a 13-tool catalog and branch-disagreement self-evolution for cross-domain visual anomaly detection. No J-cost, ratio symmetry, golden-ratio ladder, 8-tick periodicity, or parameter-free constant derivation appears; the architecture is standard ReAct-style tool orchestration plus internal disagreement feedback. RS theorems (e.g., Jcost uniqueness, phi-ladder, 8-tick dimension forcing) are untouched.","tokens_in":62575,"confidence":"high","tokens_out":287,"duration_ms":10661,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"AnomalyClaw converts single-step VLM anomaly judgments into a multi-round refutation process that checks candidate anomalies against normal-sample references using a 13-tool library.","keywords":["visual anomaly detection","vision-language models","cross-domain transfer","agentic reasoning","tool use","refutation","training-free method","multi-round verification"],"falsifier":"A controlled ablation that runs the same VLMs on the same twelve datasets once with the full multi-round refutation loop and once with an equivalent number of single-step inferences or random tool calls; if the AUROC gap disappears, the refutation mechanism is not the source of the gains.","tokens_in":2769,"feed_emoji":"🔍","tokens_out":808,"duration_ms":32737,"temperature":0.7,"pith_summary":"The paper argues that vision-language models produce unreliable anomaly detections because they lean on pre-trained priors instead of concrete evidence from normal samples or fine-grained features. It introduces a training-free agent that repeatedly proposes anomalies and refutes them tool by tool, drawing on normal references for verification and expert probing. This approach is tested on a benchmark spanning twelve datasets from industrial, medical, and remote-sensing domains. If the refutation mechanism works as claimed, it would let the same VLM handle anomaly detection across domains without retraining or domain-specific models.","feed_headline":"Refutation loop with 13 tools lifts VLM anomaly detection","feed_subtitle":"AnomalyClaw checks candidate anomalies against normal samples on 12 datasets, adding up to 8 AUROC points to GPT, Seed, and Qwen models.","key_machinery":"The multi-round refutation process that proposes candidate anomalies and refutes each one against normal-sample references with a 13-tool library for visual verification, reference parsing, and frozen expert probing.","core_discovery":"AnomalyClaw turns anomaly judgment into a multi-round refutation process in which candidate anomalies are proposed and then refuted against normal-sample references using a library of thirteen tools for visual verification, reference parsing, and frozen expert probing. On the CrossDomainVAD-12 benchmark the method raises macro-AUROC by 6.23 points with GPT-5.5, 7.93 points with Seed2.0-lite, and 3.52 points with Qwen3.5-VL-27B over direct single-step inference. An optional self-evolution extension builds an online rulebook from internal disagreements and adds a further 2.09 points on Qwen3.5-VL-27B without any oracle labels.","pith_inferences":["If the refutation loop succeeds mainly by forcing explicit comparison to normal samples, a smaller tool set focused only on reference comparison might achieve most of the benefit.","The same pattern of proposing then refuting could be tested on other VLM tasks that currently suffer from over-reliance on priors, such as medical diagnosis or safety-critical scene understanding.","The self-evolution rulebook could be made persistent across datasets, turning the agent into a growing knowledge base for cross-domain anomaly patterns."],"forward_implications":["The same agentic loop produces measurable AUROC gains on three different VLMs without any task-specific training.","An optional self-evolution step that derives rules from model disagreements adds further improvement comparable to a supervised baseline that uses ten labeled examples.","The gains are attributed to better anomaly understanding rather than simple aggregation of tool outputs.","The approach works across industrial, medical, infrastructure, and remote-sensing datasets despite differing anomaly definitions and modalities."],"fun_headline_variants":["13-tool refutation improves VLM anomaly detection","Multi-round refutation raises VLM AUROC on 12 datasets","Tool refutation advances cross-domain visual anomaly detection","AnomalyClaw refutes anomalies with 13 verification tools"],"cache_read_input_tokens":64,"weakest_assumption_plain":"The multi-round refutation process with the 13-tool library reliably improves anomaly judgments by grounding them in normal-sample references and fine-grained feature evidence rather than from longer context or tool artifacts alone.","fun_headline_variants_meta":{"raw":{"variants":["13-tool refutation improves VLM anomaly detection","Multi-round refutation raises VLM AUROC on 12 datasets","Tool refutation advances cross-domain visual anomaly detection","AnomalyClaw refutes anomalies with 13 verification tools"]},"model":"grok-4.3","cost_usd":0.009607,"raw_usage":{"total_tokens":4299,"prompt_tokens":859,"num_sources_used":0,"completion_tokens":58,"cost_in_usd_ticks":96065500,"prompt_tokens_details":{"text_tokens":859,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3382,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":859,"tokens_out":58,"duration_ms":24718,"temperature":1.0,"reasoning_tokens":3382,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-12T04:05:30.551461+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A controlled ablation that runs the same VLMs on the same twelve datasets once with the full multi-round refutation loop and once with an equivalent number of single-step inferences or random tool calls; if the AUROC gap disappears, the refutation mechanism is not the source of the gains.","supporting_citations":[],"review_version":1}