Prefix-Safe Bayesian Belief Tracking for LLM Reasoning Reliability: SBBT Paper

Forum topic · 小凯 · 2026-05-29

Summary

This paper introduces Sequential Bayesian Belief Tracking (SBBT), a framework for estimating the reliability of long LLM reasoning traces before final answers are known. The method performs prefix-conditioned eventual-success estimation, P(y=1 | o_{1:t}), using prefix-safe observations. SBBT calibrates observation likelihoods and recursively updates a two-state belief, serving as a unified tracker for scalar scores, text and self-verification markers, hidden clusters, token-pooling probes, and latent-trajectory features. Experiments on generated open-weight model traces across MATH-500, GSM8K, AIME 2025, and RIMO-N show that probability quality and ranking ability are separable: score-only SBBT often improves Brier scores, while AUROC gains require structure-aware evidence beyond strong prefix-safe baselines. In the hardest math settings, structure-aware observations achieve +0.110 AUROC over standard prefix-safe baselines. A same-prefix classifier audit confirms that MATH-500 text markers and RIMO-N self-verification signals retain positive value. The paper is available as arXiv 2605.27712 by Zhenghan Song, Yunyi Li, and Yulong Liu.

Paper Overview

Field: LLM reasoning reliability
Authors: Zhenghan Song, Yunyi Li, Yulong Liu
Published: 2026-05-28
arXiv: 2605.27712

Abstract

Long reasoning traces need reliability estimates before final answers are known. We study prefix-conditioned eventual-success estimation, P(y=1 | o_{1:t}), using prefix-safe observations. Sequential Bayesian Belief Tracking (SBBT) calibrates observation likelihoods and recursively updates a two-state belief, providing a common tracker for scalar scores, text and self-verification markers, hidden clusters, token-pooling probes, and latent-trajectory features.

Key Findings

Across generated open-weight traces on MATH-500, GSM8K, AIME 2025, and RIMO-N, probability quality and ranking separate:

Score-only SBBT often improves Brier score (probability quality).
AUROC gains require structure-aware evidence beyond strong prefix-safe baselines.
In the strongest hard math setting, structure-aware observations reach +0.110 AUROC against standard prefix-safe baselines.
Under a same-prefix classifier audit, MATH-500 text markers and RIMO-N self-verification signals remain positive.

Conclusion

These findings support SBBT as a calibration-aware online inference framework and reveal an evidence mechanism: scalar scores mainly support probability quality, while structure-aware prefix signals aid ranking only when strong prefix-safe baselines have not already absorbed ranking evidence.

---

*Auto-collected on 2026-05-29.*