Gated Steering Reduces LLM Sycophancy, Hallucination in Medicine

Large Language Models (LLMs) frequently exhibit sycophancy (agreeing with user pressure) and hallucination (generating unsupported information), which are particularly problematic in critical domains like clinical question answering. Existing solutions often address these issues separately or apply broad interventions that can degrade otherwise correct responses. This study proposes a novel framework to tackle both problems simultaneously. The researchers developed Gated Activation Steering, an Inference Time Intervention (ITI) method. This approach learns distinct steering directions for hallucination and sycophancy from contrastive clinical data and applies them to specific, causally verified attention heads within the LLM. During runtime, behavior-specific gates dynamically determine when an intervention is necessary, allowing for targeted mitigation: the hallucination component addresses unsupported claims, while the sycophancy component prevents answer shifts under user pressure. Evaluated on clinical questions grounded in Electronic Health Record (EHR) data, the framework demonstrated significant improvements. For a 4-billion-parameter model, gated steering helped the model resist user pressure in a vast majority of cases, achieving robustness levels comparable to much larger models. This indicates that targeted, inference-time steering can substantially enhance LLM reliability in medical contexts without the need for continuous, broad interventions.