Improving LLM Reasoning through Interpretable Role-Playing Steering

TL;DR

This paper introduces SRPS, a novel method for improving LLM reasoning by controlling role-playing behavior through internal feature manipulation, leading to better interpretability and performance gains.

Contribution

The paper presents SRPS, a new framework that extracts and manipulates internal model features for stable, interpretable role-playing control in LLMs, surpassing prompt engineering methods.

Findings

Improved zero-shot CoT accuracy on CSQA from 31.86% to 39.80%.

Enhanced SVAMP accuracy from 37.50% to 45.10%.

Demonstrated better interpretability and stability over traditional prompt-based methods.

Abstract

Role-playing has emerged as an effective technique for enhancing the reasoning capabilities of large language models (LLMs). However, existing methods primarily rely on prompt engineering, which often lacks stability and interpretability. In this paper, we introduce Sparse Autoencoder Role-Playing Steering (SRPS), a novel framework that identifies and manipulates internal model features associated with role-playing behavior. Our approach extracts latent representations from role-play prompts, selects the most relevant features based on activation patterns, and constructs a steering vector that can be injected into the model's residual stream with controllable intensity. Our method enables fine-grained control over role-specific behavior and offers insights into how role information influences internal model activations. Extensive experiments across various reasoning benchmarks and model sizes demonstrate consistent performance gains. Notably, in the zero-shot chain-of-thought (CoT) setting, the accuracy of Llama3.1-8B on CSQA improves from 31.86% to 39.80%, while Gemma2-9B on SVAMP increases from 37.50% to 45.10%. These results highlight the potential of SRPS to enhance reasoning ability in LLMs, providing better interpretability and stability compared to traditional prompt-based role-playing.