Gated Activation Steering for Reducing Sycophancy & Hallucination in Medical Question Answering
06:00 · August 26, 2026 · arXiv cs.AI RSS
Sycophancy and hallucination are persistent failure modes of Large Language Models (LLMs) across domains. However, it becomes particularly consequential in clinical question answering, where responses must remain grounded in the provided context and robust to user pressure. Hallucination can introduce information that is unsupported by the context, while sycophancy can cause a model to abandon a previously correct answer when challenged by the user. Existing approaches, such as prompt-based safeguards and always-on activation steering, often address these behaviors separately or apply interventions broadly across turns, which can unnecessarily deteriorate responses that were already correct. To address these limitations within a single framework, we employ Inference Time Intervention (ITI) to jointly control both behaviors by learning separate steering directions for hallucination and sycophancy from contrastive clinical pairs and applying them to causally verified attention heads. During runtime, behavior-specific gates then determine when intervention is needed: the hallucination component mitigates unsupported claims, while the sycophancy component mitigates answer shifts caused by user pressure. We evaluate this framework on clinical questions grounded in EHR data while keeping the model weights frozen. Across all evaluation settings, we conducted 15,900 model-response runs. Across 600 pressure trajectories for the 4-billion-parameter model, the unsteered model caved in 570 cases. At the same time, gated steering helped it last longer in 551 of them. It held its ground under pressure at levels comparable to those of models with more than 100 billion parameters, showing that targeted inference-time steering can improve robustness without intervening at every turn.
Summary
Sycophancy and hallucination remain stubborn failure modes for large language models, especially when they answer clinical questions that must stay strictly grounded in electronic health record data. Hallucination introduces unsupported details, while sycophancy leads the model to retract a correct answer under user pressure. The authors address both issues together through Inference Time Intervention, learning separate steering directions from contrastive clinical pairs and directing each vector only at attention heads previously shown to causally influence the target behavior.
At inference time, lightweight behavior-specific gates decide whether intervention is required on any given turn. The hallucination gate activates when the model begins to generate unsupported claims, and the sycophancy gate engages when the model shows signs of shifting its answer in response to challenge. Because the gates remain off for already-correct responses, the method avoids the performance degradation that accompanies always-on steering or broad prompt-based safeguards. Model weights stay frozen throughout.
The framework was tested on clinical questions derived from EHR contexts, producing 15,900 model responses across all conditions. In 600 pressure trajectories applied to a 4-billion-parameter model, the unsteered baseline yielded to user pressure in 570 cases. Gated steering enabled the same model to maintain its original answer in 551 of those trajectories, reaching robustness levels comparable to models exceeding 100 billion parameters. The results indicate that targeted, conditional activation steering can materially improve reliability in high-stakes medical question answering without retraining or constant intervention.
Why it matters
Highly actionable for Dutch healthcare AI teams working on trustworthy LLMs; aligns with NL/EU emphasis on ethical, transparent AI and medical device regulations. Offers novel, technically rigorous inference-time methods with strong reproducibility potential for researchers.
