Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Prototype-Based Dynamic Steering for Large Language Models

Ceyhun Efe Kayan, Li Zhang

TL;DR

Prototype-Based Dynamic Steering (PDS) introduces a lightweight, inference-time mechanism to amplify reasoning in large language models without fine-tuning or prompt redesign. By collecting activation differences between CoT and neutral prompts, clustering them into reasoning prototypes, and projecting input activations onto these prototypes to form a context-sensitive steering vector, PDS yields input-adaptive guidance injected into the residual stream. Across GSM8K, AQuA-RAT, and BIG-Bench with LLaMA-3-Instruct models, PDS improves accuracy and remains effective even under Anti-CoT prompts, indicating enhancement of latent reasoning rather than superficial behavior change. The work demonstrates that reasoning behaviors occupy a structured subspace in activation space and that dynamic, prototype-guided steering can serve as a practical, scalable alternative to training-time approaches for improving LLM reasoning.

Abstract

Despite impressive breadth, LLMs still rely on explicit reasoning instructions or static, one-fits-all steering methods, leaving a gap for adaptive, instruction-free reasoning amplification. We present Prototype-Based Dynamic Steering (PDS), a test-time method that amplifies large language model (LLM) reasoning without adding or altering instructions. We introduce "reasoning prototypes" by clustering activation differences between Chain-of-Thought (CoT) and neutral prompts. At inference, an input's hidden state is projected onto these prototypes to form an instance-specific steering vector. Evaluated on GSM8K, AQuA-RAT, and BIG-Bench tasks, PDS consistently improves accuracy without fine-tuning or prompt engineering. Notably, the gains persist even when CoT is explicitly suppressed to improve cost-efficiency, indicating that the intervention strengthens latent reasoning processes rather than inducing a superficial behavioral shift. These results position dynamic, prototype-guided steering as a lightweight alternative to training-time approaches for enhancing LLM reasoning.

Prototype-Based Dynamic Steering for Large Language Models

TL;DR

Prototype-Based Dynamic Steering (PDS) introduces a lightweight, inference-time mechanism to amplify reasoning in large language models without fine-tuning or prompt redesign. By collecting activation differences between CoT and neutral prompts, clustering them into reasoning prototypes, and projecting input activations onto these prototypes to form a context-sensitive steering vector, PDS yields input-adaptive guidance injected into the residual stream. Across GSM8K, AQuA-RAT, and BIG-Bench with LLaMA-3-Instruct models, PDS improves accuracy and remains effective even under Anti-CoT prompts, indicating enhancement of latent reasoning rather than superficial behavior change. The work demonstrates that reasoning behaviors occupy a structured subspace in activation space and that dynamic, prototype-guided steering can serve as a practical, scalable alternative to training-time approaches for improving LLM reasoning.

Abstract

Despite impressive breadth, LLMs still rely on explicit reasoning instructions or static, one-fits-all steering methods, leaving a gap for adaptive, instruction-free reasoning amplification. We present Prototype-Based Dynamic Steering (PDS), a test-time method that amplifies large language model (LLM) reasoning without adding or altering instructions. We introduce "reasoning prototypes" by clustering activation differences between Chain-of-Thought (CoT) and neutral prompts. At inference, an input's hidden state is projected onto these prototypes to form an instance-specific steering vector. Evaluated on GSM8K, AQuA-RAT, and BIG-Bench tasks, PDS consistently improves accuracy without fine-tuning or prompt engineering. Notably, the gains persist even when CoT is explicitly suppressed to improve cost-efficiency, indicating that the intervention strengthens latent reasoning processes rather than inducing a superficial behavioral shift. These results position dynamic, prototype-guided steering as a lightweight alternative to training-time approaches for enhancing LLM reasoning.

Paper Structure

This paper contains 23 sections, 5 equations, 3 figures, 12 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Foundations of Activation Steering
  4. Behavioral Steering via Difference-of-Means
  5. Alternative Steering Interventions
  6. Prototype-Based & Sparse Autoencoder Methods
  7. Positioning of Our Work
  8. Methods
  9. Activation Difference Collection
  10. Prototype Discovery via Clustering
  11. Theoretical Justification for Prototype-Based Steering
  12. Dynamic Steering via Projection
  13. Geometric Analysis of the Prototype Space
  14. Evaluation
  15. Results & Discussion
  16. ...and 8 more sections

Figures (3)

  • Figure 1: Overview of technical differences between Difference-of-Means and proposed Prototype-Based Dynamic Steering. Instead of using a single vector for each sample like DoM, PDS leverages the projections of input activations onto prototypes to compute the steering vector, which is then injected into the model's residual stream to enhance reasoning behavior.
  • Figure 2: Overview of prototype discovery in Prototype-Based Dynamic Steering. Activation differences are extracted for the contrast input pairs. Centroids obtained after K-Means clustering are treated as reasoning prototypes.
  • Figure 3: Overview of steering vector injection in Prototype-Based Dynamic Steering. During inference, projections of input activations onto prototypes are summed to compute the steering vector, which is then injected into the model's residual stream to enhance reasoning behavior.