Improving LLM Reasoning through Interpretable Role-Playing Steering

Abstract: Role-playing has emerged as an effective technique for enhancing the reasoning capabilities of 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.

• The paper introduces the SRPS framework that employs sparse autoencoders to extract interpretable role-playing features, significantly boosting LLM reasoning.
• The methodology uses latent role-related feature extraction and top-k activation selection to construct a controllable steering vector for improved model stability.
• Experiments across benchmarks like GSM8K, SVAMP, and CSQA demonstrate notable accuracy enhancements over traditional prompt engineering approaches.

Improving LLM Reasoning through Interpretable Role-Playing Steering

The paper introduces a novel framework called Sparse Autoencoder Role-Playing Steering (SRPS) aimed at enhancing the reasoning capabilities of LLMs through interpretable role-playing steering. This approach addresses limitations in existing prompt engineering methods by manipulating internal model features associated with role-playing behavior rather than relying solely on inputs examined at the surface level.

SRPS Framework Overview

Role-playing, wherein models adopt specific characters or personas to shape their reasoning, is often enhanced by adding role-related prompts before the task question. SRPS, however, introduces a more stable and interpretable approach by emphasizing internal model features. The framework uses a Sparse Autoencoder (SAE) to extract latent representations from role-play prompts, identify top-k features based on activation patterns, and construct a steering vector injected into the model’s residual stream at a controllable intensity.

Figure 1: Overview of the SRPS Framework. The LLM takes two types of inputs: one with a role-play prompt and one without.

Extracting and Selecting Role-Relevant Features

The SRPS framework extracts activations related to role-playing by averaging token activations, excluding non-semantic tokens. Selection is based on sensitivity scores combining mean activation differences and frequency differences, computed across sample pairs. The sensitivity score integrates both activation strength and frequency, aiding in selecting the most impactful features for steering.

Steering Vector Construction

Steering vectors are formed by combining the top-k selected features, weighted by their average activation. This composite steering vector is introduced to the model’s residual stream with a scaling factor allowing for adjustable steering strength. This method yields stable model behavior compared to prompt-based role-playing, often inconsistent due to LLM sensitivity to prompt phrasing.

Figure 2: Comparison of accuracy gains over the original prompting under zero-shot CoT setting when steering with different top-$k$ features.

Experimental Evaluation

The SRPS framework was evaluated across multiple reasoning benchmarks, including GSM8K, SVAMP, and CSQA, using models like Llama3.1‑8B and Gemma2 variants. The steering method consistently improved accuracy over standard prompting, demonstrating particularly significant performance gains in one-shot and few-shot CoT settings.

Figure 3: Comparison of model outputs before and after steering, using an example from the SVAMP dataset.

Interpretability and Impact on Reasoning

Through the use of Neuronpedia, the paper analyzes SAE features showing semantic alignment with domain-related concepts such as mathematics and logical thinking. This highlights how steering manages to effectively modulate model reasoning by leveraging domain-specific knowledge encoded within the model.

Figure 4: Arithmetic Reasoning.

Conclusion

The SRPS framework offers a robust and interpretable method for enhancing reasoning in LLMs by manipulating internal activations rather than relying solely on input-level changes. Extensive experiments validate that the SRPS method consistently improves reasoning performance, offering better stability and interpretability than traditional role-playing prompting methods. Future work could explore the integration of SRPS with even larger and more diverse models, expanding the applicability across various domains of reasoning and problem-solving.

Overall, the paper provides a compelling argument for shifting focus from input-level adjustments to more nuanced, internal-level steering in the development of advanced LLM reasoning techniques.