Diagnosing Multi-step Reasoning Failures in Black-box LLMs via Stepwise Confidence Attribution

2026-05-18

AI Summary

This work addresses the challenge of identifying erroneous steps in multi-step reasoning by black-box large language models, where existing confidence estimation methods either apply only to final answers or require access to internal model states. The authors propose Stepwise Confidence Attribution (SCA), a novel framework that enables step-level confidence assessment without any internal model access. SCA leverages the information bottleneck principle through two complementary approaches: a non-parametric consistency measure (NIBS) and differentiable graph masking learning (GIBS), which jointly capture logical consistency and structural variability to identify high-confidence steps aligned with consensus correct solutions. Evaluated on mathematical reasoning and multi-hop question answering tasks, SCA effectively flags low-confidence steps strongly correlated with errors and guides self-correction, improving correction success rates by up to 13.5% over answer-level feedback.

Abstract

Large Language Models have achieved strong performance on reasoning tasks with objective answers by generating step-by-step solutions, but diagnosing where a multi-step reasoning trace might fail remains difficult. Confidence estimation offers a diagnostic signal, yet existing methods are restricted to final answers or require internal model access. In this paper, we introduce Stepwise Confidence Attribution (SCA), a framework for closed-source LLMs that assigns step-level confidence based only on generated reasoning traces. SCA applies the Information Bottleneck principle: steps aligning with consensus structures across correct solutions receive high confidence, while deviations are flagged as potentially erroneous. We propose two complementary methods: (1) NIBS, a non-parametric IB approach measuring consistency without graph structures, and (2) GIBS, a graph-based IB model that learns subgraphs through a differentiable mask to capture logical variability. Extensive experiments on mathematical reasoning and multi-hop question answering show that SCA reliably identifies low-confidence steps strongly correlated with reasoning errors. Moreover, using step-level confidence to guide self-correction improves the correction success rate by up to 13.5\% over answer-level feedback.

Problem

• multi-step reasoning
• black-box LLMs
• reasoning failure diagnosis
• confidence attribution
• step-level error detection

Innovation

• Stepwise Confidence Attribution
• Information Bottleneck
• Black-box LLMs
• Multi-step Reasoning Diagnosis
• Self-correction