Large language models (LLMs) increasingly exhibit sycophancy—the tendency to conform to user beliefs rather than provide factually accurate information—posing significant risks in healthcare applications where reliability is paramount. We evaluate sycophantic behavior in ten LLMs from OpenAI, Google, and Anthropic across multi-turn medical conversations using an escalatory pushback framework. To enable fine-grained analysis, we introduce Resistance, a metric that measures nonconformity to user stances at each conversational turn, providing insights beyond existing flip-based metrics. Evaluating on MedCaseReasoning (open-ended diagnostic questions) and PubMedQA (clear-answer biomedical questions), we find that Gemini models exhibit the highest Resistance, followed by OpenAI and Claude models. We further observe that response patterns (“Yes, but...” vs. “Yes, and...”) may be more predictive of sycophancy than specific phrases. Notably, all models are more easily persuaded to change their answers on clear multiple-choice questions than on ambiguous diagnostic cases. Our findings highlight critical vulnerabilities in deploying LLMs for clinical decision support and suggest that training toward contradiction-maintaining response patterns may serve as a potential mitigation strategy.

LLM agents are predominantly benchmarked on short, single-task trajectories, yet real deployments run for hours under contention, surfacing a different class of failures. We use the Clinical Environment Simulator (CES), in which an agent manages an entire emergency-department shift under continuous time and resource pressure, as a testbed: long-horizon execution failures manifest measurably in a single rollout under structured, multi-dimensional grading. On CES, current agents reach the correct diagnosis in most cases yet fail to deliver complete and timely critical actions, revealing an execution gap. We attribute this gap to three long-horizon failure modes, each operationalized as a per-trace counter: instruction-adherence drift, treatment incompleteness, and a severity-equity gap in timeliness. We then introduce Asclepius, an adaptive agent scaffolding with a self-evolving harness that rewrites the operating manual between shifts from trace-level feedback, an externalized clinical skills library for high-stakes regimen knowledge, and three isolated subagents that partition per-turn decisions across the patient queue. On held-out batches never observed during harness evolution, Asclepius improves critical-action correctness by 22% (p = 0.024) over a strong baseline agent framework while preserving diagnostic accuracy, with consistent gains across five LLM judges from three model families; on the full ten-batch set, improvements reach 25% on critical actions and 13% on timeliness. The three failure modes form a coupled bottleneck: decisive reductions appear only when all three components act together.

Real-world clinical practice demands multi-image comparative reasoning, yet current medical benchmarks remain limited to single-frame interpretation. We present MedFrameQA, a video-derived, transcript-grounded benchmark for multi-image reasoning VQA, built from educationally-validated diagnostic sequences. To construct this dataset, we develop a scalable pipeline that leverages narrative transcripts from medical education videos to align visual frames with textual concepts, automatically producing 2,851 high-quality multi-image VQA pairs with explicit, transcript-grounded reasoning chains. Our evaluation of 13 advanced MLLMs (including reasoning models) exposes severe deficiencies in multi-image synthesis, where accuracies mostly fall below 50% and exhibit instability across varying image counts. Error analysis demonstrates that models often treat images as isolated instances, failing to track pathological progression or cross-reference anatomical shifts. MedFrameQA provides a rigorous standard for evaluating the next generation of MLLMs in handling complex, temporally grounded medical narratives.