Interpretable Risk Mitigation in LLM Agent Systems

TL;DR

This paper proposes an interpretable method for mitigating risks in LLM-based autonomous agents by steering their behavior through features extracted from autoencoder latent spaces, improving safety in a game-theoretic environment.

Contribution

It introduces a novel, interpretable steering technique for LLM agents that reduces unsafe behaviors without relying on task-specific prompts or strategies.

Findings

Steering with good-faith negotiation features reduces defection by 28 percentage points.

Feasible steering ranges identified for multiple open-source LLMs.

Game-theoretic evaluation combined with representation steering can generalize to real-world applications.

Abstract

Autonomous agents powered by large language models (LLMs) enable novel use cases in domains where responsible action is increasingly important. Yet the inherent unpredictability of LLMs raises safety concerns about agent reliability. In this work, we explore agent behaviour in a toy, game-theoretic environment based on a variation of the Iterated Prisoner's Dilemma. We introduce a strategy-modification method-independent of both the game and the prompt-by steering the residual stream with interpretable features extracted from a sparse autoencoder latent space. Steering with the good-faith negotiation feature lowers the average defection probability by 28 percentage points. We also identify feasible steering ranges for several open-source LLM agents. Finally, we hypothesise that game-theoretic evaluation of LLM agents, combined with representation-steering alignment, can generalise to real-world applications on end-user devices and embodied platforms.