{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.09890", "kind": "paper", "url": "https://arxiv.org/abs/2606.09890"}], "authors": ["Hainiu Xu", "Italo Luis da Silva", "Jiangnan Ye", "Yuhao Wang", "Wei Liu", "Linyi Yang", "Jonathan Richard Schwarz", "Nicola Paoletti", "Yulan He", "Hanqi Yan"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.09890", "source_url": "https://arxiv.org/abs/2606.09890"}, "categories": ["General AI", "Language & Knowledge", "cs.LG"], "description": {"en": "PreActBench is a benchmark for predictive monitoring in LLMs, consisting of 1,000 paired ethical and unethical action trajectories across five domains. It evaluates whether models can infer if a partial trajectory will culminate in unethical action, using the Prefix Foresight F1 metric."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.09890", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.09890", "title": "PreAct-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["King's College London", "National University of Singapore", "Southern University of Science and Technology", "Thomson Reuters Foundation", "Imperial College London", "The Alan Turing Institute"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.09890", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.09890", "source_url": "https://openalex.org/W7164181810"}, "key": "claire-radar:2606.09890", "modality": null, "name": "PreAct-Bench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "High", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "unknown", "origin_record_id": "bm_preact-bench_d5c8c56f", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "66a60082d0ae8466e5f9f8ef27aff9f342ebd92fd04906a3d0f480935ba4b9dd", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2606.09890"}, "publisher": null, "released": "2026-06-03", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.09890", "source_url": "https://arxiv.org/abs/2606.09890"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-09890", "source": "claire_radar", "source_benchmark_id": "2606.09890", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2606.09890", "observedAt": "2026-10-04T03:02:42.754940Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "state": "unavailable"}}, "status": "unavailable"}, "capabilities": [], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "476be280857a1820deeefed7ec9eea374221783707fbbf60ebdd8f2328a60cb3", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "PreActBench is a benchmark for predictive monitoring in LLMs, consisting of 1,000 paired ethical and unethical action trajectories across five domains. It evaluates whether models can infer if a partial trajectory will culminate in unethical action, using the Prefix Foresight F1 metric.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "To support this task, we present PreActBench, a benchmark of 1,000 paired ethical and unethical action trajectories spanning five domains."}, "familyId": "bmf_cb7720730271", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_preact-bench_d5c8c56f", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2606.09890", "paper": "https://arxiv.org/abs/2606.09890", "pdf": "https://arxiv.org/pdf/2606.09890", "project": null, "report": "https://arxiv.org/abs/2606.09890"}, "metrics": [], "motivation": "Large language models (LLMs) are increasingly deployed as autonomous agents capable of executing multi-step action trajectories toward a given objective.", "name": "PreAct-Bench", "oneLine": "PreActBench is a benchmark for predictive monitoring in LLMs, consisting of 1,000 paired ethical and unethical action trajectories across five domains. It evaluates whether models can infer if a partial trajectory will culminate in unethical action, using the Prefix Foresight F1 metric.", "paperTitle": "PreAct-Bench: Benchmarking Predictive Monitoring in LLMs", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.09890", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "arXiv", "organizationType": "benchmark-organization", "role": "benchmark-publisher", "sourceUrl": "https://arxiv.org/abs/2606.09890"}], "ranking": {}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-06-03", "source": {"authors": ["Hainiu Xu", "Italo Luis da Silva", "Jiangnan Ye", "Yuhao Wang", "Wei Liu", "Linyi Yang", "Jonathan Richard Schwarz", "Nicola Paoletti", "Yulan He", "Hanqi Yan"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "id": "2606.09890", "title": "PreAct-Bench: Benchmarking Predictive Monitoring in LLMs", "type": "arxiv", "url": "https://arxiv.org/abs/2606.09890"}, "sourceUpdatedAt": "2026-06-03T13:26:41Z", "topics": ["cs.LG"], "venueAttempts": [], "whyItMatters": "Safety research often detects unethical behavior only after it occurs. Predictive monitoring enables anticipation of harm before execution, and PreActBench measures this capability across models and guardrails."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
