{"paper":{"title":"Can VLMs Unlock Semantic Anomaly Detection? A Framework for Structured Reasoning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Structured reasoning framework lets VLMs detect semantic anomalies in driving scenes with 18.5 percent higher recall.","cross_cats":["cs.AI","cs.RO"],"primary_cat":"cs.CV","authors_text":"David Pop, Johannes Betz, Mattia Piccinini, Roberto Brusnicki, Yuan Gao","submitted_at":"2025-10-20T19:14:29Z","abstract_excerpt":"Autonomous driving systems remain critically vulnerable to the long-tail of rare, out-of-distribution semantic anomalies. While VLMs have emerged as promising tools for perception, their application in anomaly detection remains largely restricted to prompting proprietary models - limiting reliability, reproducibility, and deployment feasibility. To address this gap, we introduce SAVANT (Semantic Anomaly Verification/Analysis Toolkit), a novel model-agnostic reasoning framework that reformulates anomaly detection as a layered semantic consistency verification. By applying SAVANT's two-phase pip"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Applying SAVANT improves VLM's absolute recall by approximately 18.5% compared to prompting baselines, and leveraging the best proprietary model within the framework enables automatic labeling of around 10,000 images to fine-tune a 7B open-source model achieving 90.8% recall and 93.8% accuracy.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The evaluation uses a 'balanced set of real-world driving scenarios' whose selection criteria and representativeness of long-tail anomalies are not specified, which is required to support that the reported recall gains are due to the structured reasoning pipeline rather than dataset construction.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"SAVANT boosts VLM recall for semantic anomaly detection in driving images by 18.5% via structured reasoning and enables fine-tuning a 7B open model to 90.8% recall and 93.8% accuracy.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Structured reasoning framework lets VLMs detect semantic anomalies in driving scenes with 18.5 percent higher recall.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"e065cd65ec00f8368f1fb3e237ce517ca0d19694fded47e8b3900325c085e044"},"source":{"id":"2510.18034","kind":"arxiv","version":3},"verdict":{"id":"c52c4db2-f984-44e3-963a-df5747e99981","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-18T05:43:04.751411Z","strongest_claim":"Applying SAVANT improves VLM's absolute recall by approximately 18.5% compared to prompting baselines, and leveraging the best proprietary model within the framework enables automatic labeling of around 10,000 images to fine-tune a 7B open-source model achieving 90.8% recall and 93.8% accuracy.","one_line_summary":"SAVANT boosts VLM recall for semantic anomaly detection in driving images by 18.5% via structured reasoning and enables fine-tuning a 7B open model to 90.8% recall and 93.8% accuracy.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The evaluation uses a 'balanced set of real-world driving scenarios' whose selection criteria and representativeness of long-tail anomalies are not specified, which is required to support that the reported recall gains are due to the structured reasoning pipeline rather than dataset construction.","pith_extraction_headline":"Structured reasoning framework lets VLMs detect semantic anomalies in driving scenes with 18.5 percent higher recall."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2510.18034/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"2fceb3457930879b5343fa37b2737a3e105095b36484eb43a163ae2953f3523f"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}