{"paper":{"title":"Confidence Estimation for LLMs in Multi-turn Interactions","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"A new logit probe called P(Sufficient) tracks how LLMs accumulate evidence across conversation turns while staying calibrated.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Caiqi Zhang, Chengzu Li, Deqing Yang, Nigel Collier, Ruihan Yang, Tiancheng Hu, Xiaochen Zhu, Yijiang River Dong","submitted_at":"2026-01-05T14:58:04Z","abstract_excerpt":"While confidence estimation is a promising direction for mitigating hallucinations in Large Language Models (LLMs), current research overwhelmingly focuses on single-turn settings. The dynamics of model confidence in multi-turn conversations, where context accumulates and ambiguity is progressively resolved, remain largely unexplored. This work presents the first systematic study of confidence estimation in multi-turn interactions, establishing a formal evaluation framework grounded in two key desiderata: per-turn calibration and monotonicity of confidence as more information becomes available"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"a novel logit-based probe we introduce, P(Sufficient), proves comparatively more effective, robustly tracking evidence accumulation and distinguishing it from conversational filler.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the two key desiderata (per-turn calibration and monotonicity of confidence as information accumulates) are sufficient to evaluate and improve confidence estimation in multi-turn settings, and that the Hinter-Guesser paradigm produces datasets representative of real ambiguity resolution.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"The work establishes a framework for multi-turn LLM confidence estimation using per-turn calibration and monotonicity, with a new P(Sufficient) probe outperforming standard methods on controlled datasets.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A new logit probe called P(Sufficient) tracks how LLMs accumulate evidence across conversation turns while staying calibrated.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"e30fabd74fe3255cd9b37fe7a907b770a75e4114ce5aacb73b141061540ca0c7"},"source":{"id":"2601.02179","kind":"arxiv","version":2},"verdict":{"id":"09fc81be-d47e-4888-83a4-8019281825c4","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T17:53:50.551350Z","strongest_claim":"a novel logit-based probe we introduce, P(Sufficient), proves comparatively more effective, robustly tracking evidence accumulation and distinguishing it from conversational filler.","one_line_summary":"The work establishes a framework for multi-turn LLM confidence estimation using per-turn calibration and monotonicity, with a new P(Sufficient) probe outperforming standard methods on controlled datasets.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the two key desiderata (per-turn calibration and monotonicity of confidence as information accumulates) are sufficient to evaluate and improve confidence estimation in multi-turn settings, and that the Hinter-Guesser paradigm produces datasets representative of real ambiguity resolution.","pith_extraction_headline":"A new logit probe called P(Sufficient) tracks how LLMs accumulate evidence across conversation turns while staying calibrated."},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"da6d5be116616ed3b18dec04e7b2a6308eacf9eb272485ea30fd061e8d9179de"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}