{"record": {"aliases": ["Measuring LLM Sycophancy under Sustained Multi-Turn Pressure"], "artifacts": [{"id": "arxiv:2609.09090", "kind": "paper", "url": "https://arxiv.org/abs/2609.09090"}, {"kind": "website", "url": "https://anonymous.4open.science/r/SPINE"}], "authors": ["Leyuan Tang", "Kangda Wei", "Tianyu Jiang", "Ruihong Huang"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2609.09090", "source_url": "https://arxiv.org/abs/2609.09090"}, "categories": ["General AI", "Language & Knowledge", "cs.CL"], "description": {"en": "SPINE evaluates LLM sycophancy by having an adaptive LLM proxy persistently challenge a target model across up to 25 turns on false-presupposition and unethical-query items, measuring collapse rates."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.09090", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.09090", "title": "SPINE"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Texas A&M University", "University of Cincinnati"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2609.09090", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2609.09090", "source_url": "https://openalex.org/W7212049589"}, "key": "claire-radar:2609.09090", "modality": null, "name": "SPINE", "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": "true", "origin_record_id": "bm_measuring-llm-sycophancy-under-sustained-m_c0976161", "origin_source": "arxiv", "recognition_confidence": "0.55", "record_sha256": "f80e973cabe615635381fbaff8b37f1863e96be9cb616a1b6b046dbe17522148", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-09-10T00:52:28.558291Z", "source_url": "https://arxiv.org/abs/2609.09090"}, "publisher": null, "released": "2026-09-08", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.09090", "source_url": "https://arxiv.org/abs/2609.09090"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-09090", "source": "claire_radar", "source_benchmark_id": "2609.09090", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Measuring LLM Sycophancy under Sustained Multi-Turn Pressure"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attention": {"asOf": "2026-09-09", "attemptedAt": "2026-09-09T12:31:07.558041Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2609.09090", "observedAt": "2026-09-08T01:38:09.842216Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-09-09T12:31:07.558041Z", "state": "unavailable"}}, "status": "unavailable"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The benchmark addresses LLM sycophancy under sustained pressure, a specific but increasingly relevant issue, with modest artifact visibility.", "score": 42}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-10T00:52:28.558291Z", "inputHash": "3a9c45703e273b70015c4459a1867634d21b40aafb2adc8659058c0987eb8b3d", "model": "deepseek-v4-pro"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "We introduce SPINE, a benchmark in which an LLM proxy plays a persistent but mistaken user and adaptively challenges a target model for up to 25 turns.", "canonicalNameSource": "abstract", "decisionReason": "SPINE is a formally named benchmark with released code and data, a defined task, and measurable scoring metrics.", "evaluationEvidence": {"reuse": "The code and data are released at https://anonymous.4open.science/r/SPINE", "scoring": "Our experimental results show that collapse rates increase with conversation length for every model, short-horizon protocols underestimate sycophancy and resistance under sustained pressure remains unreliable across current models.", "task": "We introduce SPINE, a benchmark in which an LLM proxy plays a persistent but mistaken user and adaptively challenges a target model for up to 25 turns. We evaluate four production systems and three Olmo3-7b variants on 100 false-presupposition and 100 unethical-query items."}, "model": "deepseek-v4-pro", "reviewedAt": "2026-09-10T00:52:28.558291Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "SPINE evaluates LLM sycophancy by having an adaptive LLM proxy persistently challenge a target model across up to 25 turns on false-presupposition and unethical-query items, measuring collapse rates.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "public artifact URL"], "snippet": "We introduce SPINE, a benchmark in which an LLM proxy plays a persistent but mistaken user and adaptively challenges a target model for up to 25 turns."}, "familyId": "bmf_0c050c4062c9", "firstSeenAt": "2026-09-09", "heat": null, "id": "bm_measuring-llm-sycophancy-under-sustained-m_c0976161", "indexedAt": "2026-09-09T06:18:29.794093Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2609.09090", "pdf": "https://arxiv.org/pdf/2609.09090", "project": "https://anonymous.4open.science/r/SPINE", "report": "https://arxiv.org/abs/2609.09090"}, "metrics": [], "motivation": "Large language models (LLMs) may abandon correct positions when users push back, exhibiting a failure mode known as sycophancy.", "name": "SPINE", "oneLine": "SPINE evaluates LLM sycophancy by having an adaptive LLM proxy persistently challenge a target model across up to 25 turns on false-presupposition and unethical-query items, measuring collapse rates.", "paperTitle": "Measuring LLM Sycophancy under Sustained Multi-Turn Pressure", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2609.09090", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-09T06:18:29.794093Z"}, "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": null, "historyDate": null, "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}, "today": {"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": 1, "historyDate": "2026-09-08", "rank": null, "score": null, "windowDays": 1}, "level": {"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}, "llmAttentionForecast": {"maxContribution": 0.25, "percentile": 0.45, "role": "bonus", "value": 45}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": 11}, "method": "attention-ranking-v13", "rank": null, "score": 11}}, "readiness": "Paper only", "recognitionConfidence": 0.55, "relation": "introduces", "releasedAt": "2026-09-08", "source": {"authors": ["Leyuan Tang", "Kangda Wei", "Tianyu Jiang", "Ruihong Huang"], "categories": ["cs.CL", "cs.AI"], "id": "2609.09090", "title": "Measuring LLM Sycophancy under Sustained Multi-Turn Pressure", "type": "arxiv", "url": "https://arxiv.org/abs/2609.09090"}, "sourceUpdatedAt": "2026-09-08T17:35:04Z", "topics": ["cs.CL"], "venueAttempts": [], "whyItMatters": "It exposes sycophantic collapse that short-horizon benchmarks miss and shows that resistance under sustained pressure remains unreliable, informing deployment in adversarial conversational settings."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
