๐ค AI Summary
Existing implicit chain-of-thought (CoT) reasoning methods avoid discrete token-level error propagation but lack inter-step consistency guarantees, leading to divergent reasoning paths and unstable outputs. To address this, we propose EBM-CoT: the first energy-based model (EBM) framework for calibrating implicit CoT reasoning. It defines a consistency-aware energy function over a continuous latent space and employs gradient-based optimization to dynamically steer reasoning trajectories toward high-consistency regionsโwithout modifying the base language model. This work pioneers the application of EBMs to consistency modeling in implicit multi-step reasoning. Experiments demonstrate substantial improvements in accuracy and stability across mathematical, commonsense, and symbolic reasoning tasks. EBM-CoT effectively suppresses spurious reasoning paths, enhances logical coherence, and improves inference efficiency.
๐ Abstract
Large Language Models (LLMs) have demonstrated strong reasoning capabilities through emph{Chain-of-Thought} (CoT) prompting, which enables step-by-step intermediate reasoning. However, explicit CoT methods rely on discrete token-level reasoning processes that are prone to error propagation and limited by vocabulary expressiveness, often resulting in rigid and inconsistent reasoning trajectories. Recent research has explored implicit or continuous reasoning in latent spaces, allowing models to perform internal reasoning before generating explicit output. Although such approaches alleviate some limitations of discrete CoT, they generally lack explicit mechanisms to enforce consistency among reasoning steps, leading to divergent reasoning paths and unstable outcomes. To address this issue, we propose EBM-CoT, an Energy-Based Chain-of-Thought Calibration framework that refines latent thought representations through an energy-based model (EBM). Our method dynamically adjusts latent reasoning trajectories toward lower-energy, high-consistency regions in the embedding space, improving both reasoning accuracy and consistency without modifying the base language model. Extensive experiments across mathematical, commonsense, and symbolic reasoning benchmarks demonstrate that the proposed framework significantly enhances the consistency and efficiency of multi-step reasoning in LLMs.