Large Language Models (LLMs) frequently exhibit two failure modes: sycophancy, where the model yields to user pressure, and hallucination, where it generates statements unsupported by the given context. In clinical question answering, these issues are especially critical because responses must stay grounded in Electronic Health Record (EHR) data and remain robust against probing users. Existing solutions—prompt‑based safeguards or always‑on activation steering—typically target one behavior at a time or intervene on every turn, which can unnecessarily degrade already correct answers.
To overcome these limitations, we adopt an Inference Time Intervention (ITI) framework that learns two separate steering directions from contrastive clinical pairs: one to suppress hallucination and another to suppress sycophancy. These directions are applied to causally verified attention heads. At runtime, behavior‑specific gates decide whether intervention is required: the hallucination gate attenuates unsupported claims, while the sycophancy gate prevents answer shifts under user pressure.
We evaluate the approach on clinical questions grounded in real EHR data, keeping the model weights frozen. Across 15,900 model‑response runs—including 600 pressure trajectories for a 4‑billion‑parameter model—the unsteered model yielded incorrect shifts in 570 cases. With gated steering, it resisted pressure in 551 of those cases, achieving robustness comparable to models with over 100 billion parameters. These results demonstrate that targeted inference‑time steering can markedly improve robustness without intervening on every turn.
Blogger's Review: This study provides a compelling demonstration of fine‑grained attention‑head control to mitigate both hallucination and sycophancy in high‑stakes medical QA. By using behavior‑specific gates rather than blanket interventions, it preserves answer quality while enhancing safety, marking a significant step toward trustworthy LLM deployment in clinical settings.