TrAC: A New Framework for Efficient Uncertainty Quantification in LLMs
A new paper on arXiv (2608.00422) introduces Trace-Conditioned Answer Consistency (TrAC), a framework for uncertainty quantification in large language models (LLMs). The method combines active and passive signals anchored to a single completed reasoning trace, addressing a gap in existing approaches. Current methods fall into three categories: passive single-trace methods using token-level confidence, sampling-based methods comparing multiple traces at higher cost, and active prefix-based methods probing partial traces. None re-elicit an answer from a completed trace to measure consistency. TrAC's active component, Prefix-Conditioned Answer Consistency, is designed to improve efficiency and accuracy in detecting incorrect answers, aiding in abstention, human review, and adaptive compute allocation. The paper is categorized as 'new' and was announced on arXiv.
Key facts
- Paper arXiv:2608.00422 introduces TrAC (Trace-Conditioned Answer Consistency).
- TrAC is a correctness-supervised uncertainty quantification framework for LLMs.
- It combines active and passive signals anchored to one completed reasoning trace.
- Existing methods include passive single-trace, sampling-based, and active prefix-based approaches.
- TrAC's active component is called Prefix-Conditioned Answer Consistency.
- The framework aims to improve abstention, human review, and adaptive compute allocation.
- The paper was announced as a new type on arXiv.
- The method addresses a gap: no existing approach re-elicits an answer from a completed trace.
Entities
Institutions
- arXiv