Uncovering Latent Chain of Thought Vectors in Language Models

2024-09-21

arXiv.org

AI Summary

Large language models (LLMs) often lack controllable chain-of-thought (CoT) reasoning, undermining their trustworthiness in societally impactful applications. Method: We propose a prompt-free, fine-tuning-free latent-space intervention method that identifies and exploits linearly separable CoT directions within LLMs’ internal representations. By constructing task-specific steering vectors, we directly modulate hidden states during forward propagation. Contribution/Results: Evaluated on Llama3-8B and Mistral-7B-v0.2, our approach enables zero-shot, lightweight, and generalizable CoT control—matching standard CoT prompting performance on GSM8k, MMLU, AGIEval, and ARC, while achieving high reasoning consistency and significantly lower computational overhead than fine-tuning. Crucially, we uncover the intrinsic geometric structure of CoT representations and establish the first interpretable, intervention-based reasoning control paradigm grounded in latent-space directional manipulation.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

As language models grow more influential and trusted in our society, our ability to reliably steer them toward favorable behaviors becomes increasingly paramount. For this, we investigate the technique of steering vectors: biasing the forward pass of language models using a"steering vector"derived from a specific task. We apply them to steer language models toward performing Chain of Thought (CoT) Reasoning without the need to prompt through natural language. We demonstrate this approach on Llama3 8b and Mistral 7b v0.2, and obtain competitive results compared to CoT-prompted performances on a series of reasoning benchmarks (GSM8k, MMLU, AGI Eval, ARC AI2) and qualitative examples. We find this approach yields consistent steering towards CoT responses and takes less compute than traditional methods of fine-tuning models towards CoT.

Problem

• Language Models
• Chain-of-Thought Reasoning
• Social Applications

Innovation

• Guiding Vectors
• Chain-of-Thought Reasoning
• Computational Efficiency