Abstract: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. To facilitate this, we introduce novel metrics, including a length-normalized Expected Calibration Error (InfoECE), and a new "Hinter-Guesser" paradigm for generating controlled evaluation datasets. Our experiments reveal that widely-used confidence techniques struggle with calibration and monotonicity in multi-turn dialogues. In contrast, a novel logit-based probe we introduce, P(Sufficient), proves comparatively more effective, robustly tracking evidence accumulation and distinguishing it from conversational filler. Our work provides a foundational methodology for developing more reliable and trustworthy conversational agents.
| Comments: | ACL 2026 Findings |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2601.02179 [cs.CL] |
| (or arXiv:2601.02179v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2601.02179 arXiv-issued DOI via DataCite |
Submission history
From: Caiqi Zhang [view email]
[v1]
Mon, 5 Jan 2026 14:58:04 UTC (9,278 KB)
[v2]
Wed, 13 May 2026 21:26:22 UTC (9,278 KB)