Activation Control for Efficiently Eliciting Long Chain-of-thought Ability of Language Models

TL;DR

This paper reveals that a small set of activations in large language models largely controls long chain-of-thought reasoning, and introduces a training-free activation control method to efficiently elicit and enhance this ability.

Contribution

It uncovers the internal activation mechanisms behind long CoT reasoning and proposes a novel, training-free activation control technique to improve reasoning without costly fine-tuning.

Findings

Amplifying key activations boosts long CoT reasoning.

Activation dynamics follow predictable trajectories.

Training-free activation control improves reasoning performance.

Abstract

Despite the remarkable reasoning performance, eliciting the long chain-of-thought (CoT) ability in large language models (LLMs) typically requires costly reinforcement learning or supervised fine-tuning on high-quality distilled data. We investigate the internal mechanisms behind this capability and show that a small set of high-impact activations in the last few layers largely governs long-form reasoning attributes, such as output length and self-reflection. By simply amplifying these activations and inserting "wait" tokens, we can invoke the long CoT ability without any training, resulting in significantly increased self-reflection rates and accuracy. Moreover, we find that the activation dynamics follow predictable trajectories, with a sharp rise after special tokens and a subsequent exponential decay. Building on these insights, we introduce a general training-free activation control technique. It leverages a few contrastive examples to identify key activations, and employs simple analytic functions to modulate their values at inference time to elicit long CoTs. Extensive experiments confirm the effectiveness of our method in efficiently eliciting long CoT reasoning in LLMs and improving their performance. Additionally, we propose a parameter-efficient fine-tuning method that trains only a last-layer activation amplification module and a few LoRA layers, outperforming full LoRA fine-tuning on reasoning benchmarks with significantly fewer parameters. Our code and data are publicly released.