SPREG: Structured Plan Repair with Entropy-Guided Test-Time Intervention for Large Language Model Reasoning

TL;DR

SPREG is a real-time framework that detects and corrects logical errors in large language models during reasoning by monitoring entropy and dynamically repairing the reasoning process.

Contribution

It introduces an adaptive entropy-guided intervention mechanism that improves reasoning accuracy and stability without degrading language fluency.

Findings

20.0% accuracy improvement on AIME25

Effective suppression of entropy drift in complex tasks

Dynamic repair mechanism enhances reasoning stability

Abstract

Large Language Models (LLMs) are prone to logical hallucinations and stochastic drifts during long-chain reasoning. While Classifier-Free Guidance (CFG) can improve instruction adherence, standard static implementations often cause semantic dilution and linguistic degradation. We propose SPREG (Structured Plan-guided Real-time Entropy Gating), a lightweight inference-time framework for surgical error rectification. SPREG employs an adaptive dual-threshold mechanism to monitor real-time entropy, identifying sudden ``entropy spikes'' as reliable indicators of logical failure. Upon detection, it triggers a dynamic repair by replacing uninformative null-priors with reference distributions synthesized from historical high-confidence states. By modulating guidance intensity according to structured reasoning stages (e.g., Action, Observation), SPREG steers the model back to a stable manifold without compromising fluency. Our experiments demonstrate significant gains, notably a 20.0% absolute accuracy improvement on AIME25, while effectively suppressing uncontrolled entropy drift in complex tasks.