Incidental information contaminates patient notes and disrupts clinical reasoning in large language models
Abstract
Large language models (LLMs) are increasingly relied upon to support ambient documentation and clinical reasoning. Here we examine the impact of a failure mode shared between these two applications by assessing their sensitivity to information incidental to the patient encounter. In 576 patient-clinician dialogues, we found that frontier models inserted small-talk exchanges into 35% of notes, while mean quality scores changed by at most 0.20 points on five-point scales. In 3.7% of frontier notes, models misattributed the asides or used them clinically. In 57 mock recorded consultations, background speech from a separate patient encounter at -10 dB leaked into 48.2% of transcripts, with contamination detected in 5.3% of downstream notes generated by four open-weight models. We propose a dual encoding hypothesis of clinical reasoning and distraction in LLMs, with preliminary evidence that LLM components associated with disruption by incidental information also support clinical reasoning. These findings support evaluating resistance to incidental information before clinical use, with safeguards that prevent contamination while preserving clinical reasoning.
Community
Large language models (LLMs) are increasingly relied upon to support ambient documentation and clinical reasoning. Here we examine the impact of a failure mode shared between these two applications by assessing their sensitivity to information incidental to the patient encounter. In 576 patient-clinician dialogues, we found that frontier models inserted small-talk exchanges into 35% of notes, while mean quality scores changed by at most 0.20 points on five-point scales. In 3.7% of frontier notes, models misattributed the asides or used them clinically. In 57 mock recorded consultations, background speech from a separate patient encounter at -10 dB leaked into 48.2% of transcripts, with contamination detected in 5.3% of downstream notes generated by four open-weight models. We propose a dual encoding hypothesis of clinical reasoning and distraction in LLMs, with preliminary evidence that LLM components associated with disruption by incidental information also support clinical reasoning. These findings support evaluating resistance to incidental information before clinical use, with safeguards that prevent contamination while preserving clinical reasoning.
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- Knowledge Graph-Augmented Ambient AI for Clinical Note Generation (2026)
- K-Bench: a clinically calibrated benchmark for evaluating large language models in high-risk mental health conversations (2026)
- Symphony for Text Generation: Benchmarking Clinical Note Generation (2026)
- Performance of Clinical AI System and Physicians and Frontier Language Models in primary care diagnostics (2026)
- KlinikeBench: Evaluating Language Models Beyond Diagnostic Accuracy (2026)
- Sense and Sensitivity: Benchmarking LLM Clinical Triage Recommendations with Physician Experts (2026)
- EviGen: Predictive Evidence Scaffolding for Verifiable Clinical Rationale Generation (2026)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend
A 0.2-point mean shift is the whole story, and that's precisely why it's easy to wave off — contamination is a tail event, and mean rubric scores are structurally blind to tails. A handful of notes where the model folds an incidental detail into the assessment barely budges an average while being the only thing that matters clinically. The sub-4% misattribution rate is what I'd put on a dashboard, not the mean. I'd also stop reaching for a better rubric; rubrics grade the output after the damage is done. What I want is a deterministic traceability check — every clinical claim in the note maps to a specific clinician utterance, and anything unmapped gets flagged before a human reads it. Same failure mode I keep hitting in agent transcripts: the eval says fine, the trace says the model invented a tool result, and only the trace was right.
Get this paper in your agent:
hf papers read 2610.08585 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 2
Spaces citing this paper 0
No Space linking this paper