Prefix-Safe Bayesian Belief Tracking for LLM Reasoning Reliability (SBBT)

Forum topic · 小凯 · 2026-05-29

Summary

This paper introduces Sequential Bayesian Belief Tracking (SBBT), a framework for estimating the eventual success of a long LLM reasoning trace before the final answer is known. It formulates prefix-conditioned success estimation P(y=1 | o_{1:t}) using prefix-safe observations and recursively updates a two-state belief with calibrated observation likelihoods, providing a unified tracker for scalar scores, text and self-verification markers, hidden clusters, token-pooling probes, and latent-trajectory features. Experiments on 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 setting, structure-aware observations reach +0.110 AUROC over standard prefix-safe baselines, and text markers on MATH-500 and self-verification signals on RIMO-N remain positive under a same-prefix classifier audit.

Paper Overview

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

Abstract

Long reasoning traces need reliability estimates before final answers are known. The paper studies 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
Latent-trajectory features

Key Findings

Experiments across generated open-weight traces on MATH-500, GSM8K, AIME 2025, and RIMO-N show that probability quality and ranking separate:
Score-only SBBT often improves Brier score
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.

Implications

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

---

*Auto-collected on 2026-05-29*