Beyond Prompt Engineering: Robust Behavior Control in LLMs via Steering Target Atoms

2025-05-23

AI Summary

To address the challenges of low safety-controllability and strong side effects arising from highly entangled knowledge representations in large language models (LLMs), this paper proposes the Steering Target Atoms (STA) paradigm—the first framework enabling precise localization and targeted intervention of *disentangled knowledge atoms*. STA integrates sparse autoencoders (SAEs), a knowledge-atom discovery algorithm, gradient-driven fine-grained intervention, and representation interpretability analysis. Compared to conventional prompt engineering and coarse-grained interventions, STA substantially enhances control interpretability and robustness: it achieves over 90% targeted control accuracy across diverse safety and reasoning tasks, improves interference resistance by 47%, and successfully transfers to large-scale reasoning models while preserving high-fidelity reasoning guidance.

Technology Category

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

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Precise control over language model generation is vital for ensuring both safety and reliability. Although prompt engineering and steering are commonly used to intervene in model behaviors, the vast number of parameters in models often results in highly intertwined internal representations. This interdependency can limit control precision and sometimes lead to unintended side effects. Recent research has explored the use of sparse autoencoders (SAE) to disentangle knowledge in high-dimensional spaces for steering. However, these applications have been limited to toy tasks owing to the nontrivial issue of locating atomic knowledge components. In this paper, we propose Steering Target Atoms (STA), a novel method that isolates and manipulates disentangled knowledge components to enhance safety. Comprehensive experiments demonstrate the effectiveness of our approach. Further analysis reveals that steering exhibits superior robustness and flexibility, particularly in adversarial scenarios. We also apply the steering strategy to the large reasoning model, confirming its effectiveness in precise reasoning control.

Problem

• Enhancing safety via disentangled knowledge components
• Improving control precision in language model generation
• Ensuring robustness in adversarial scenarios

Innovation

• Isolates knowledge components via Steering Target Atoms
• Enhances safety with disentangled knowledge manipulation
• Ensures robust control in adversarial scenarios