Think Consistently, Reason Efficiently: Energy-Based Calibration for Implicit Chain-of-Thought

TL;DR

This paper introduces EBM-CoT, an energy-based calibration method that refines latent reasoning trajectories in LLMs, significantly improving reasoning consistency and accuracy across various benchmarks without altering the base model.

Contribution

The paper proposes a novel energy-based calibration framework for implicit Chain-of-Thought reasoning, enhancing consistency and accuracy in LLMs without modifying their core architecture.

Findings

Improves reasoning consistency across multiple benchmarks.

Enhances reasoning accuracy without changing the base model.

Demonstrates effectiveness in mathematical, commonsense, and symbolic reasoning tasks.

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.