{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2609.12446", "kind": "paper", "url": "https://arxiv.org/abs/2609.12446"}, {"kind": "website", "url": "https://rlg.iis.sinica.edu.tw/papers/werewolf-accusation-benchmark"}], "authors": ["Yu-Yu Yang", "Ti-Rong Wu", "Hung Guei", "Hsing-Yu Chen", "I-Chen Wu"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2609.12446", "source_url": "https://arxiv.org/abs/2609.12446"}, "categories": ["General AI", "Language & Knowledge", "cs.AI"], "description": {"en": "A belief-shift evaluation in Werewolf measures whether models update beliefs in response to accusers or accusations, using belief updating as a signal of communication skills."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.12446", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.12446", "title": "Do LLMs Trust the Accuser or the Accusation? Measuring Belief Shifts in Werewolf"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2609.12446", "modality": null, "name": "Do LLMs Trust the Accuser or the Accusation? Measuring Belief Shifts in Werewolf", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "false", "origin_record_id": "bm_do-llms-trust-the-accuser-or-the-accusatio_020c1e79", "origin_source": "arxiv", "recognition_confidence": "0.55", "record_sha256": "9ff5b44274fd8d2091998dfcea558c21672898eaf643ac8409f0eba43492037a", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-name-audit-deferred", "reviewed_at": "2026-09-14T10:05:20.149847Z", "source_url": "https://arxiv.org/abs/2609.12446"}, "publisher": null, "released": "2026-09-11", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.12446", "source_url": "https://arxiv.org/abs/2609.12446"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-12446", "source": "claire_radar", "source_benchmark_id": "2609.12446", "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/2609.12446", "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"}, "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "A Werewolf belief-shift probe may interest multi-agent communication researchers, though it lacks a standalone comparison path and any supplied artifacts.", "score": 16}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-15T09:37:10.086794Z", "inputHash": "d4cbe9ea57a237ccddcb18bba85b937ca2e79953961bd6f4caac82853500515c", "model": "deepseek-v4-flash"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "Do LLMs Trust the Accuser or the Accusation? Measuring Belief Shifts in Werewolf", "canonicalNameSource": "paper_title", "decisionReason": "The abstract only states a proposed benchmark without evidence of released tasks, scoring rules, or a public path for other teams to run or inspect it.", "evaluationEvidence": {"reuse": "Benchmark and code are available at https://rlg.iis.sinica.edu.tw/papers/werewolf-accusation-benchmark.", "scoring": "Belief shifts in village-side model after each message; evaluated on 1,224 annotated messages from 40 open-weight LLM configurations.", "task": "Social-deduction game Werewolf communication: LLM-played games with annotated suspicion and accusation messages, measuring belief shifts of an observing model."}, "model": "deepseek-v4-pro", "reviewedAt": "2026-09-14T10:05:20.149847Z", "state": "ai-name-audit-deferred"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "A belief-shift evaluation in Werewolf measures whether models update beliefs in response to accusers or accusations, using belief updating as a signal of communication skills.", "displayEligible": false, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "public artifact URL"], "snippet": "We propose a belief-shift evaluation benchmark in Werewolf for analyzing communication skills through belief updating."}, "familyId": "bmf_d6dfa655372e", "firstSeenAt": "2026-09-14", "heat": null, "id": "bm_do-llms-trust-the-accuser-or-the-accusatio_020c1e79", "indexedAt": "2026-09-14T00:39:56.575498Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2609.12446", "pdf": "https://arxiv.org/pdf/2609.12446", "project": "https://rlg.iis.sinica.edu.tw/papers/werewolf-accusation-benchmark", "report": "https://arxiv.org/abs/2609.12446"}, "metrics": [], "motivation": "Social-deduction games such as Werewolf are increasingly used to evaluate LLM agents, but existing evaluations often rely on final game outcomes.", "name": "Do LLMs Trust the Accuser or the Accusation? Measuring Belief Shifts in Werewolf", "oneLine": "A belief-shift evaluation in Werewolf measures whether models update beliefs in response to accusers or accusations, using belief updating as a signal of communication skills.", "paperTitle": "Do LLMs Trust the Accuser or the Accusation? Measuring Belief Shifts in Werewolf", "primaryDomain": "General AI", "publication": {"evidence": "Accepted by the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026 Main Conference)", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2609.12446", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026 Main Conference)", "verifiedAt": "2026-09-14T00:39:56.575498Z"}, "publications": [], "ranking": {"30d": {"confidence": "Low", "coverage": 0.0, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": 30, "historyDate": "2026-09-05", "rank": null, "score": null, "windowDays": 30}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": null, "value": null}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v13", "rank": null, "score": null}, "90d": {"confidence": "Low", "coverage": 0.0, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 90}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": null, "value": null}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v13", "rank": null, "score": null}}, "readiness": "Paper only", "recognitionConfidence": 0.55, "relation": "introduces", "releasedAt": "2026-09-11", "source": {"authors": ["Yu-Yu Yang", "Ti-Rong Wu", "Hung Guei", "Hsing-Yu Chen", "I-Chen Wu"], "categories": ["cs.AI", "cs.CL"], "id": "2609.12446", "title": "Do LLMs Trust the Accuser or the Accusation? Measuring Belief Shifts in Werewolf", "type": "arxiv", "url": "https://arxiv.org/abs/2609.12446"}, "sourceUpdatedAt": "2026-09-11T05:10:00Z", "topics": ["cs.AI"], "venueAttempts": [{"decisionRaw": "Accepted by the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026 Main Conference)", "evidence": [{"level": "author-claim", "observedAt": "2026-09-14T00:39:56.575498Z", "rawValue": "Accepted by the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026 Main Conference)", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2609.12446"}], "reviewStatus": "accepted", "venueName": "2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026 Main Conference)"}], "whyItMatters": "Multi-agent social-deduction settings stress belief updating under persuasive communication; a probe of accuser-versus-accusation trust could inform evaluation of interactive language agents."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
