Author

L. Klein

View on Semantic Scholar

2 papers indexed here

We haven’t gathered this author’s papers yet. Follow them and we’ll fetch their work.

Not the right person? Other researchers publish under this name.

Review Open access Sep 2026

Evaluating large language models as clinical decision support tools in primary healthcare settings: Protocol for a multi-country comparative validation study on expert-adjudicated hypothetical vignettes (hypMOOVE-PHC)

Introduction: Large language models (LLMs) have the potential to strengthen clinical decision-making in low-resource primary healthcare (PHC) settings. However, most LLMs are developed and benchmarked in high-resource settings and evidence on their safety and contextual appropriateness in Sub-Saharan Africa remains limited. The hypMOOVE-PHC study is the hypothetical vignette phase of the Massive Open Online Validation and Evaluation (MOOVE) initiative, implemented in Kenya, Malawi, and Tanzania. It aims to validate a pool of LLMs through clinical review of expert-generated vignettes. Methods and analysis: This is a fully crossed repeated-measures comparative evaluation study. In each country, experienced clinicians develop 200-250 hypothetical clinical vignettes reflecting realistic patient presentations and independently produce a human benchmark care plan for each. Vignettes are used to prompt a selection of six open-source and proprietary LLMs selected based on code availability, local hostability, and model size. During in-person workshops (valiDATAthons), independent clinical experts rate LLM- and human-generated responses in source-attribution masked side-by-side comparisons across five dimensions (clinical soundness, safety, contextual fit, clarity & completeness, and appropriate confidence). The primary endpoints are each LLM's overall performance profile and non-inferior safety profile, as compared to the human benchmark. At minimum, 358 evaluations per LLM (or 1,253 paired evaluations in total) are required per country. Ethics and dissemination: The study is approved by the EPFL Human Ethics Research Committee in Switzerland, Harvard T.H. Chan School of Public Health in the USA, KNH-UoN Ethics and Research Committee in Kenya, MUBAS Research Ethics Committee in Malawi, and MUHAS Research and Ethics Committee and National Institute for Medical Research in Tanzania. Findings will be reported according to the TRIPOD-LLM framework and shared with national ministries of health, disseminated at conferences and in peer-reviewed journals, and de-identified benchmark data will be released under FAIR principles.

P. Macharia, C. Kachimanga, M. Mahende et al. · medRxiv · 0 citations

Preprint Aug 2026

The Working Set of a Coding Agent: Coherence Debt in Repository-Scale Tasks

Repository-scale coding requires an agent to keep tests, imports, configuration, and migration rules consistent within a bounded context window. We model this as reconstructing a coupled-fact graph: at each edit, a required fact comes from recent context or parametric memory, and the facts covered by neither form coherence debt. We supply and withhold each channel and inject faults across seven models and five harnesses. As expected, no model completes a task on an unseen API with both channels empty, and putting the facts in the prompt restores success. When a rename defeats what models memorized about a real library, all seven fail in the same place, passing and missing the same tests. Availability decides the outcome and distance does not: withholding a fact costs exactly the work it supports, and a supplied fact works as well far from the edit as next to it. Harnesses pay unequal prices for it: configurations that all pass every test differ more than tenfold in tokens consumed because they rebuild the same content at different rates, and spending more recovers nothing when facts are withheld. A missing fact produces wrong work rather than absent work: an agent asked to act acts, fabricating the file or guessing the value, so instruments built on reads look for a hole already filled. How often it says it is blocked instead is a property of the model, from every trial to none. Availability does not settle every edit: where standard and code disagree, agents follow the standard even when it prescribes the worse code, so a stale convention file costs more than no file. Because parametric memory substitutes for reading, on SWE-bench, where models likely know the repositories, reads no longer predict success. Harnesses should keep the facts an edit depends on available when the agent writes, and check that availability against what the agent produces rather than what it reads.

Bardia Mohammadi, L. Klein, Aman Chadha et al. · 2 citations