Towards agents that know when they don't know: uncertainty as a control signal for structured reasoning

Conference item

Synthetic chain-of-thought (CoT) traces are widely used to train large reasoning models (LRMs), improving generalization by providing step-level supervision. Yet most approaches require ground-truth labels to seed or filter these traces—an expensive bottleneck in domains like biology where wet-lab data are scarce. We propose a label-free alternative: uncertainty-based filtering, which uses a model’s own confidence—quantified through established uncertainty metrics like self-consistency and predictive perplexity—as a substitute for external labels. We sample multiple reasoning traces and retain only low-uncertainty subsets. Applied to biological perturbation prediction, a domain where wet-lab labels are especially costly, we show that the filtered subset has higher accuracy, and that supervised fine-tuning (SFT) on uncertainty-filtered data outperforms unfiltered synthetic data, narrows the gap to ground-truth training, and surpasses strong LRM baselines. Ablations show that per-class filtering corrects for class-specific uncertainty scales and that hybrid uncertainty metrics yield higher-quality datasets. Our results suggest that modelinternal confidence is a powerful signal for efficient reasoning dataset creation, enabling LRMs in domains where supervision is expensive.

Authors: Stoisser, JL; Phillips, L; Misra, A; Lamb, TA; Torr, P

• Publication status: Accepted
• Peer review status: Peer reviewed
• Publisher: Open Review
• Publication date: 2025-09-16
• Acceptance date: 2025-10-16
• Event title: 39th Conference on Neural Information Processing Systems (NeurIPS 2025)
• Event location: San Diego, California, USA
• Event website: https://neurips.cc/Conferences/2025
• Event start date: 2025-11-30
• Event end date: 2025-12-07
• Language: English