LLMs fail sequential clinical triage despite strong retrospective performance
A new evaluation framework exposes a critical gap in LLM deployment for clinical triage. While models perform adequately on retrospective case reviews, they degrade significantly when forced to make sequential decisions from incomplete information, mirroring real emergency department workflows. Six tested models dropped from moderate-to-substantial agreement to fair-to-moderate accuracy as new clinical data arrived incrementally. This finding challenges the assumption that benchmark performance on complete records translates to real-world clinical utility, raising questions about readiness for deployment in high-stakes medical settings where decisions compound over time.
Modelwire context
ExplainerThe paper doesn't just show LLMs fail on incomplete data; it demonstrates that models trained and tested on complete medical records show no degradation signal until forced into incremental decision sequences. This means current benchmarks actively hide the failure mode rather than simply underestimating it.
This connects directly to the faithfulness work from mid-September, which showed that explanations can sound coherent without reflecting actual reasoning. Here, the models likely anchor on the initial chief complaint and rationalize subsequent decisions post-hoc rather than updating their reasoning as new evidence arrives. The sequential triage failure also echoes the coverage-aware retrieval problem: LLMs default to pattern-matching the most salient input (chief complaint, dominant case features) while failing to weight rare but critical signals that arrive later. Both expose how LLM cognition privileges early, high-confidence signals over systematic integration of marginal information.
If the same six models are tested on a prospective dataset from a real ED where clinicians record decisions at each triage step (not retrospectively), and performance tracks the arXiv results, that confirms the gap is reproducible outside the lab. If performance recovers when models are explicitly prompted to revise prior assessments with each new data point, that points to an inference-time fix rather than a fundamental training limitation.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsEmergency Severity Index · LLMs
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “LLMs Anchor on Chief Complaint and Fail to Integrate Evidence in Sequential Clinical Triage”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.