Mitigating Content Effects on Reasoning in Language Models through Fine-Grained Activation Steering

Researchers have developed a new technique called activation steering to reduce reasoning biases in large language models, particularly the tendency to confuse content plausibility with logical validity. Their novel K-CAST method achieved up to 15% improvement in formal reasoning accuracy while maintaining robustness across different tasks and languages.

• →Large language models suffer from reasoning biases that conflate content plausibility with formal logical validity.
• →Activation steering is an inference-time technique that can modulate internal model activations to reduce content biases.
• →The new K-CAST conditional steering method improved formal reasoning accuracy by up to 15% in unresponsive models.
• →The technique is robust to prompt variations and maintains multilingual capabilities with minimal side effects.
• →Activation-level interventions offer a scalable approach for enhancing LLM reasoning without requiring model retraining.