{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.04570", "kind": "paper", "url": "https://arxiv.org/abs/2608.04570"}], "authors": ["Yushi Sun", "Yanjie Zhang", "Rui Sheng"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.04570", "source_url": "https://arxiv.org/abs/2608.04570"}, "categories": ["General AI", "Language & Knowledge", "cs.CL"], "description": {"en": "Evaluates LLM personalization faithfulness: 150 personas, 6 tasks, four-way faithfulness taxonomy, 12 models across 7 families, 143,616 judged claims."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.04570", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.04570", "title": "MirageBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Hong Kong University of Science and Technology"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2608.04570", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2608.04570", "source_url": "https://openalex.org/W7196941275"}, "key": "claire-radar:2608.04570", "modality": null, "name": "MirageBench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_miragebench_adb719c2", "origin_source": "arxiv", "recognition_confidence": "0.65", "record_sha256": "b408388452ad187396507a925546ee61e637d73ff528a4b60883c00e0ef1967c", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-09-08T08:57:27.463024Z", "source_url": "https://arxiv.org/abs/2608.04570"}, "publisher": null, "released": "2026-08-05", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.04570", "source_url": "https://arxiv.org/abs/2608.04570"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-04570", "source": "claire_radar", "source_benchmark_id": "2608.04570", "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": "2026-08-06T00:00:00.000Z", "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 40, "hfPaperUrl": "https://huggingface.co/papers/2608.04570", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}}, "status": "fresh"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "A counterintuitive self-monitoring finding in a hot personalization safety topic should attract moderate attention despite a narrow evaluation scope.", "score": 45}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-08T08:57:27.463024Z", "inputHash": "8c25c8fdb425715377355128c353f554e858be863a0cdf7a06bf6150ce522390", "model": "deepseek-v4-pro"}, "curation": {"canonicalNameEvidence": "We introduce MirageBench", "canonicalNameSource": "abstract", "decisionReason": "Explicit benchmark with released personas, tasks, taxonomy, and leaderboard; validated judge provides stable scoring and reusable path.", "model": "deepseek-v4-pro", "reviewedAt": "2026-09-08T08:57:27.463024Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates LLM personalization faithfulness: 150 personas, 6 tasks, four-way faithfulness taxonomy, 12 models across 7 families, 143,616 judged claims.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "evaluation protocol evidence"], "snippet": "We introduce MirageBench, comprising 150 personas balanced across stereotypical, counter-stereotypical, and neutral profiles, 6 personalization tasks spanning an ``imagination gradient'', a four-way faithfulness taxonomy operationalized by an independent judge (validated against a blind human annotator on 400 claims: Cohen's kappa = 0.863 four-class, kappa = 0.900 binary), and a leaderboard of 12 models across 7 families on 143616 judged claims."}, "familyId": "bmf_991ed4a9392b", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_miragebench_adb719c2", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2608.04570", "pdf": "https://arxiv.org/pdf/2608.04570", "project": null, "report": "https://arxiv.org/abs/2608.04570"}, "metrics": [], "motivation": "Personalized LLMs with persistent memory are increasingly deployed, yet the faithfulness of their user models remains unexamined.", "name": "MirageBench", "oneLine": "Evaluates LLM personalization faithfulness: 150 personas, 6 tasks, four-way faithfulness taxonomy, 12 models across 7 families, 143,616 judged claims.", "paperTitle": "The Personalization Mirage: How LLMs Fabricate User Profiles, and Why Self-Monitoring Misleads", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.04570", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "ranking": {"90d": {"confidence": "Low", "coverage": 0.15, "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": 0.6561681661200994, "value": 40}}, "confidence": "Low", "coverage": 0.15, "rank": null, "score": 52}, "method": "attention-ranking-v13", "rank": null, "score": 52}}, "readiness": "Paper only", "recognitionConfidence": 0.65, "relation": "introduces", "releasedAt": "2026-08-05", "source": {"authors": ["Yushi Sun", "Yanjie Zhang", "Rui Sheng"], "categories": ["cs.CL"], "id": "2608.04570", "title": "The Personalization Mirage: How LLMs Fabricate User Profiles, and Why Self-Monitoring Misleads", "type": "arxiv", "url": "https://arxiv.org/abs/2608.04570"}, "sourceUpdatedAt": "2026-08-05T08:00:54Z", "topics": ["cs.CL"], "venueAttempts": [], "whyItMatters": "Reveals pervasive over-inference in personalized LLMs and shows self-reported confidence misleads model comparison, supporting external verification."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
