TrAC: Trace-Conditioned Answer Consistency for Efficient Uncertainty Quantification in LLMs
Dahai Yu, Lin Jiang, Rongchao Xu, Guang Wang · 2026
Read TrAC: Trace-Conditioned Answer Consistency for Efficient Uncertainty Quantification in LLMs by Dahai Yu, Lin Jiang, Rongchao Xu, Guang Wang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2608-00422.txt.
Description
Yu and co-authors present TrAC, a way to gauge doubt in LLM answers from one reasoning trace. Fluent traces can still end in wrong answers. Response-level doubt scores guide abstention, human review, and compute choices. Past methods use token confidence, many sampled traces, or probes of partial traces. None re-asks for an answer from a finished trace.
TrAC re-elicits a short answer conditioned on the completed trace. That probe scores both consistency with the first answer and token support. A passive profile tracks how token doubt evolves across the trace. A light head merges both signals into a correctness score. On five math benchmarks, TrAC beat eight-sample self-consistency on AUROC and AURC. It used one full trace plus a short cached probe.
Catalog record
| CREATOR | Dahai Yu, Lin Jiang, Rongchao Xu, Guang Wang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 66 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2608.00422