Research map / Large language models / Reasoning and chain of thought

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

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Stepwise Confidence Attribution (SCA), a black-box method that scores reasoning steps by their alignment with consensus structures among correct sampled traces, via a training-free NIBS and a learned graph-based GIBS.

Problem

Diagnosing which step in a multi-step LLM reasoning trace is erroneous, without step-level annotations or internal token probabilities (closed-source setting), where existing confidence estimation only covers final answers or needs white-box access.

Approach

Applies the Information Bottleneck principle to step confidence: steps consistent with consensus structure (semantic-similarity or maximum-common-subgraph alignment) across correct sampled traces get high confidence, deviations get low confidence; GIBS learns a differentiable subgraph mask with a sparsity (KL) term and a consensus cross-entropy term.

Key results

• GIBS achieves the highest AUROC in 7 of 9 model-dataset configurations; e.g., 0.7892 on GSM8K with Phi-4-reasoning vs. 0.6600 for the best NIBS variant (NLI-Max) on the same setting (Phi4 column).
• Stepwise feedback with step-level confidence improves self-correction success rate by up to 13.5% over answer-level feedback on MoreHopQA.
• GIBS trained on MoreHopQA and tested on MATH without retraining still outperforms NIBS and white-box baselines in AUROC under domain shift.

Surprising findings

• Building consensus from all sampled traces (including incorrect ones) drops AUROC to near-random (~0.5), so correct-only anchors are necessary.
• Removing the edge encoder from GIBS sharply lowers AUROC (e.g., 0.5186 vs 0.7892 on GSM8K), more than removing the graph encoder.

Limitations

• Requires final-answer correctness labels to build correct-only consensus; label-free self-consistency variant performs below the oracle and depends on pseudo-label quality (overlap ~28% for Llama3.1-8B).
• Absolute ECE values remain high for several configurations (e.g., GIBS ECE 0.2293 to 0.3354 on GSM8K), so calibration is not uniformly strong.
• Relies on GPT-4o as judge for step-level labels on QA and on a rule-based graph parser; the pipeline is compute-heavy with quadratic NLI calls per edge pair.

Experimental setup

Datasets: GSM8K MATH MoreHopQA PRM800K

Benchmarks: GSM8K MATH MoreHopQA PRM800K

Metrics: AUROC AUCPR ACC@80% ECE self-correction success rate

Baselines: SL(norm) Token Entropy P(true) LeCo Cos-Max Cos-Mean NLI-Max NLI-Mean

Builds on: Information Bottleneck Variational Information Bottleneck Chain-of-Thought Graph-of-Thought

Closest papers by meaning

• Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.90)
• A Theoretical Study on Bridging Internal Probability and Self-Consistency for LLM Reasoning (NeurIPS 2025 · similarity 0.90)
• Learning to Route LLMs with Confidence Tokens (ICML 2025 · similarity 0.90)
• Inference Time Optimization with Confidence Dynamics (ICML 2026 · similarity 0.89)
• Argument Graph Uncertainty: Quantifying Uncertainty from the Logical Structure of Reasoning Chains (NeurIPS 2026 · similarity 0.89)
• CITE: Anytime Valid Statistical Inference in LLM Self-Consistency (NeurIPS 2026 · similarity 0.89)