Probabilistic Soundness Guarantees in LLM Reasoning Chains

2025-07-17

AI Summary

Initial errors in LLM reasoning chains tend to propagate, undermining conclusion reliability; existing detection methods are limited because they neglect the error propagation mechanism. To address this, we propose ARES, a framework that incrementally evaluates each reasoning step based on verified premises and introduces an inductive probabilistic assessment mechanism—providing statistically grounded, fine-grained correctness certification per step, thereby overcoming the fragility of binary correctness labels. ARES models autoregressive reasoning entailment stability and performs stepwise independent evaluation, effectively isolating error propagation paths. On four standard benchmarks, ARES achieves a Macro-F1 score of 72.1% (+8.2 points over prior work); for detecting propagated errors in long reasoning chains, it attains 90.3% F1 (+27.6 points), significantly outperforming state-of-the-art approaches.

Technology Category

• Knowledge Representation and Reasoning: Computational Complexity of Reasoning
• Reasoning under Uncertainty: Other Foundations of Reasoning under Uncertainty
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

In reasoning chains generated by large language models (LLMs), initial errors often propagate and undermine the reliability of the final conclusion. Current LLM-based error detection methods often fail to detect propagated errors because they do not properly account for how earlier errors might corrupt judgments of downstream reasoning. To better detect such propagated errors, we introduce Autoregressive Reasoning Entailment Stability (ARES), a novel probabilistic framework that prevents error propagation by judging each claim based only on previously-assessed sound premises. This inductive method yields a nuanced score for each step and provides certified statistical guarantees of its soundness, rather than a brittle binary label. ARES achieves state-of-the-art performance across four benchmarks (72.1% Macro-F1, +8.2 points) and demonstrates superior robustness on very long synthetic reasoning chains, where it excels at detecting propagated errors (90.3% F1, +27.6 points).

Problem

• Detecting propagated errors in LLM reasoning chains
• Improving reliability of final conclusions in LLM reasoning
• Providing probabilistic soundness guarantees for reasoning steps

Innovation

• Autoregressive probabilistic framework for error detection
• Inductive scoring with certified soundness guarantees
• State-of-the-art performance on multiple benchmarks