Improving Reasoning Performance in Large Language Models via Representation Engineering

2025-04-28

AI Summary

This work addresses the challenge of precisely controlling large language models’ (LLMs) reasoning capabilities without fine-tuning. We propose a representation-level intervention method that identifies task-relevant activations within the residual stream, constructs task-specific control vectors from them, and directly modifies representations during inference to enhance inductive, deductive, and mathematical reasoning. To our knowledge, this is the first systematic application of representation engineering to LLM reasoning control—relying solely on forward-pass activation extraction and residual-stream intervention, thereby revealing the intrinsic decomposability of reasoning abilities. Evaluated on Mistral-7B-Instruct and Pythia models, our approach consistently improves accuracy across diverse reasoning benchmarks, stabilizes logit distributions, and its mechanistic validity is confirmed via KL divergence and entropy analysis. All code and analytical tools are publicly released to support reproducible representation intervention research.

Technology Category

• Knowledge Representation and Reasoning: Argumentation
• Machine Learning: Large Multimodal Models (LMMs)
• Reasoning under Uncertainty: Uncertainty Representations

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

Recent advancements in large language models (LLMs) have resulted in increasingly anthropomorphic language concerning the ability of LLMs to reason. Whether reasoning in LLMs should be understood to be inherently different is, however, widely debated. We propose utilizing a representation engineering approach wherein model activations are read from the residual stream of an LLM when processing a reasoning task. The activations are used to derive a control vector that is applied to the model as an inference-time intervention, modulating the representational space of the model, to improve performance on the specified task. We publish the code for deriving control vectors and analyzing model representations. The method allows us to improve performance on reasoning benchmarks and assess how control vectors influence the final logit distribution of a model via metrics such as KL divergence and entropy. We apply control vectors to Mistral-7B-Instruct and a range of Pythia models on an inductive, a deductive and mathematical reasoning task. We show that an LLM can, to a certain degree, be controlled to improve its perceived reasoning ability by modulating activations. The intervention is dependent upon the ability to reliably extract the model's typical state when correctly solving a task. Our results suggest that reasoning performance can be modulated in the same manner as other information-processing tasks performed by LLMs and demonstrate that we are capable of improving performance on specific tasks via a simple intervention on the residual stream with no additional training.

Problem

• Enhancing reasoning in LLMs via activation modulation
• Deriving control vectors to improve task performance
• Assessing reasoning modulation via KL divergence metrics

Innovation

• Representation engineering modulates model activations
• Control vectors improve reasoning without training
• Intervention on residual stream enhances performance