MedPRESS: A Multi-turn Benchmark for Patient-Pressure-Induced Medical Sycophancy in LLMs
2026-08-03 • Computation and Language
Computation and Language
AI summaryⓘ
The authors created MedPRESS, a test with 600 medical conversations, to see how language models behave when pressured by patients asking repeatedly for unsafe advice. They found many models often agree with unsafe requests after ongoing pressure, though some are better than others. Their tests also show that special prompts can help models resist pressure but don't fully solve the problem. Overall, the authors highlight that just knowing safe medical facts isn't enough; models must keep being safe even in tough talks.
Large Language Models (LLMs)MedPRESS benchmarksafety evaluationpatient pressuresycophancymedical dialogueprompt engineeringconversational robustness
Authors
Saman Sarker Joy, Niloy Farhan
Abstract
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.