Steering Risk Preferences in Large Language Models by Aligning Behavioral and Neural Representations

2025-05-16

AI Summary

Controlling large language models’ (LLMs) risk preferences without fine-tuning or retraining remains challenging due to the lack of interpretable, systematic mechanisms for preference modulation. Method: We propose a behavior–neuro dual-path representational alignment framework: (1) LLM-driven MCMC sampling constructs interpretable behavioral representations of risk decisions; (2) concurrent analysis of residual stream activations in Transformer layers extracts neural activation patterns; and (3) alignment between behavioral and neural representations enables identification and construction of “steering vectors” across risk dimensions (e.g., conservative vs. risk-seeking tendencies). Contribution/Results: Our method achieves an average 32.7% accuracy improvement across diverse risk-related tasks, demonstrates strong robustness to input perturbations, and generalizes across distinct LLM architectures. It establishes the first systematic, representation-driven paradigm for controllable and interpretable preference editing—moving beyond opaque, ad-hoc interventions toward principled, neuro-behaviorally grounded control.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Preferences
• Natural Language Processing: Safety and Robustness

Application Category

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

Abstract

Changing the behavior of large language models (LLMs) can be as straightforward as editing the Transformer's residual streams using appropriately constructed"steering vectors."These modifications to internal neural activations, a form of representation engineering, offer an effective and targeted means of influencing model behavior without retraining or fine-tuning the model. But how can such steering vectors be systematically identified? We propose a principled approach for uncovering steering vectors by aligning latent representations elicited through behavioral methods (specifically, Markov chain Monte Carlo with LLMs) with their neural counterparts. To evaluate this approach, we focus on extracting latent risk preferences from LLMs and steering their risk-related outputs using the aligned representations as steering vectors. We show that the resulting steering vectors successfully and reliably modulate LLM outputs in line with the targeted behavior.

Problem

• Systematically identify steering vectors for LLMs
• Align behavioral and neural representations for risk preferences
• Modulate LLM outputs using aligned steering vectors

Innovation

• Editing Transformer residual streams with steering vectors
• Aligning behavioral and neural latent representations
• Modulating LLM outputs using aligned steering vectors