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.
Large language models (LLMs) are increasingly used for health-related advice. Existing research measures their safety with static questions rather than pressured patient-facing conversations. We introduce MedPRESS, a multi-turn benchmark for measuring patient-pressure-induced sycophancy in LLMs. MedPRESS contains 600 medically grounded five-turn dialogues across three scenario families: medication and treatment demand, personal health self-care, and symptom triage and care resistance. Each dialogue begins with a health query and escalates through personal experience, social proof, external evidence claims, and direct adversarial challenge. We evaluate 20 LLMs across general, medical-domain, lightweight, large, open-weight, and proprietary families using structured judging and safety-focused metrics. Results show that models frequently shift toward unsafe agreement under repeated patient pressure, with substantial variation across model families, model scale, and prompt type. Anti-sycophancy prompting improves robustness for several models, but does not eliminate unsafe agreement. MedPRESS highlights a critical gap in medical LLM evaluation: safe medical knowledge is not enough unless models can maintain it under conversational pressure.
The sycophancy of large language models can increase the safety risk in intervention dialogue for autistic children. Supervised fine-tuning can somewhat reduce sycophancy, but relying solely on positive examples is often insufficient to identify and correct failure patterns. We observe that sycophancy behaviors can often be localized to a limited span within the model response. In this regime, sequence-level preference optimization can over-update preference-irrelevant tokens and degrade intervention ability. To address this, we propose the \textbf{M}inimal \textbf{E}dit \textbf{D}ata \textbf{A}ugmentation (MEDA) strategy to construct controlled, stable, minimal edit preference pairs and \textbf{T}oken-level \textbf{D}ifference \textbf{D}irect \textbf{P}reference \textbf{O}ptimization (TD-DPO), which upweights difference tokens between chosen and rejected responses while downweighting shared tokens to suppress background drift. Extensive experiments across multiple backbones and evaluators show that TD-DPO achieves a better trade-off between sycophancy mitigation and intervention ability retention in our offline settings, highlighting its potential as a practical alignment approach for autism intervention.
Background: Consumer-facing large language models are now a common source of health information, and they interpret and personalize responses rather than retrieve them. Whether their responses vary across users is a clinical, equity, and governance question, sharpened by evidence that sycophantic responses can alter judgment and increase trust. Objective: To evaluate response variation and sycophancy in consumer-facing health LLMs under conditions resembling ordinary patient use. Methods: We constructed simulated user profiles differing in geography, browsing context, expressed beliefs, and social determinants of health, drawing on literature linking social context to health attitudes. We adapted validated instruments, including the Vaccination Attitudes Examination scale and reproductive attitudes scales, into multi-turn prompts designed to elicit clinically meaningful variation across users. Results: The evaluation encountered five linked barriers. Factual prompts produced stable responses that masked sycophancy emerging over multi-turn conversation. Browser-based interfaces did not disclose which signals influence outputs and could not be reset to a clean baseline. Large-scale testing was restricted by terms of service, rate limits, and bot detection. Accuracy-based criteria could not capture tone, framing, or omission, and LLM-as-judge methods risked shared alignment bias. Models changed without traceable version identifiers, preventing reliable replication. Conclusions: No reliable independent evaluation framework yet exists for examining how consumer-facing health LLMs behave in ordinary use. Oversight requires disclosure of personalization signals, stable version identifiers, researcher safe harbor programs, and post-deployment monitoring of health-related outputs.