Information-seeking failures of large language models in agentic clinical reasoning

arXiv cs.AI Papers

Summary

The paper develops an agentic evaluation framework for clinical reasoning in hematologic oncology, finding that LLMs primarily fail due to systematic information-seeking deficits rather than insufficient knowledge, with error patterns resembling cognitive biases in novice clinicians.

arXiv:2607.10275v1 Announce Type: new Abstract: Large language models achieve high scores on medical knowledge assessments, yet clinical reasoning requires actively deciding what to investigate under uncertainty. We developed an agentic evaluation framework in hematologic oncology in which models must proactively request clinical data across three sequential rounds before committing to a diagnosis and treatment plan. Across 32 frontier models, the best achieved only 68% overall accuracy. Information utilization, the fraction of available data actually requested, was the strongest predictor of diagnostic accuracy (R = 0.69, P < 0.001), yet utilization collapsed from 57% to 26% in the final round, leaving molecular and cytogenetic data critical for treatment selection unexamined. Reasoning traces scored high on a clinical reasoning rubric (91% above threshold) but decorrelated from accuracy, revealing a gap between locally coherent rationales and globally correct conclusions. Error analysis identified search satisficing, anchoring and premature closure as the dominant failure modes, the same cognitive biases that characterize novice clinicians under dual-process models of diagnostic reasoning. These findings demonstrate that the primary limitation of current models in clinical oncology is not insufficient medical knowledge but a systematic failure of information-seeking under uncertainty.
Original Article
View Cached Full Text

Cached at: 07/14/26, 04:19 AM

# Information-seeking failures of large language models in agentic clinical reasoning
Source: [https://arxiv.org/abs/2607.10275](https://arxiv.org/abs/2607.10275)
Authors:[Krischan Braitsch](https://arxiv.org/search/cs?searchtype=author&query=Braitsch,+K),[Laura K\. Schmalbrock](https://arxiv.org/search/cs?searchtype=author&query=Schmalbrock,+L+K),[Theresa Weltermann](https://arxiv.org/search/cs?searchtype=author&query=Weltermann,+T),[Andrew F\. Berdel](https://arxiv.org/search/cs?searchtype=author&query=Berdel,+A+F),[Isabella Miller](https://arxiv.org/search/cs?searchtype=author&query=Miller,+I),[Kai Tran](https://arxiv.org/search/cs?searchtype=author&query=Tran,+K),[Michael Heider](https://arxiv.org/search/cs?searchtype=author&query=Heider,+M),[Sabrina Kraus](https://arxiv.org/search/cs?searchtype=author&query=Kraus,+S),[Florian Bassermann](https://arxiv.org/search/cs?searchtype=author&query=Bassermann,+F),[Jacqueline Lammert](https://arxiv.org/search/cs?searchtype=author&query=Lammert,+J),[Sebastian Ziegelmayer](https://arxiv.org/search/cs?searchtype=author&query=Ziegelmayer,+S),[Marcus Makowski](https://arxiv.org/search/cs?searchtype=author&query=Makowski,+M),[Lisa C\. Adams](https://arxiv.org/search/cs?searchtype=author&query=Adams,+L+C),[Keno K\. Bressem](https://arxiv.org/search/cs?searchtype=author&query=Bressem,+K+K)

[View PDF](https://arxiv.org/pdf/2607.10275)

> Abstract:Large language models achieve high scores on medical knowledge assessments, yet clinical reasoning requires actively deciding what to investigate under uncertainty\. We developed an agentic evaluation framework in hematologic oncology in which models must proactively request clinical data across three sequential rounds before committing to a diagnosis and treatment plan\. Across 32 frontier models, the best achieved only 68% overall accuracy\. Information utilization, the fraction of available data actually requested, was the strongest predictor of diagnostic accuracy \(R = 0\.69, P < 0\.001\), yet utilization collapsed from 57% to 26% in the final round, leaving molecular and cytogenetic data critical for treatment selection unexamined\. Reasoning traces scored high on a clinical reasoning rubric \(91% above threshold\) but decorrelated from accuracy, revealing a gap between locally coherent rationales and globally correct conclusions\. Error analysis identified search satisficing, anchoring and premature closure as the dominant failure modes, the same cognitive biases that characterize novice clinicians under dual\-process models of diagnostic reasoning\. These findings demonstrate that the primary limitation of current models in clinical oncology is not insufficient medical knowledge but a systematic failure of information\-seeking under uncertainty\.

## Submission history

From: Keno Bressem \[[view email](https://arxiv.org/show-email/0c212e4b/2607.10275)\] **\[v1\]**Sat, 11 Jul 2026 12:10:33 UTC \(2,783 KB\)

Similar Articles