Uncovering Latent Chain of Thought Vectors in Language Models

TL;DR

This paper demonstrates that targeted perturbations in language model activation spaces can encode and induce complex reasoning patterns, enabling Chain-of-Thought reasoning without natural language prompts, and achieving competitive results.

Contribution

It introduces a novel method of injecting steering vectors into activation space to induce reasoning in LMs, bypassing traditional prompt-based approaches.

Findings

Activation-space interventions outperform traditional CoT prompting on multiple benchmarks.

Neural activations can encode complex reasoning patterns.

Method works on Llama3 8B and Mistral 7B models.

Abstract

In this work, we examine how targeted perturbations in the activation space of Language Models (LMs) can encode complex reasoning patterns. We inject steering vectors, derived from LM activations, into LMs during inference time and study whether these vectors can induce Chain-of-Thought (CoT) reasoning in LMs without the need for natural language prompting. We demonstrate this approach on Llama3 8B Instruct and Mistral 7B v0.2 Instruct and show that activation-space interventions achieve competitive, if not superior, performance compared to traditional CoT prompting across multiple reasoning benchmarks, including GSM8k, MMLU, AGI Eval, and ARC AI2. These findings suggest that neural network activations can encode reasoning patterns, offering a new application of activation space manipulation as a tool for tuning model behavior.