Large language models are increasingly used to turn complex study output into plain-English summaries. But how do we know which models are safest and most reliable for healthcare?
In this most recent community AI research paper reading, Arjun Mukerji, PhD – Staff Data Scientist at Atropos Health – walks us through RWESummary, a new benchmark designed to evaluate LLMs on summarizing real-world evidence from structured study output — an important but often under-tested scenario compared to the typical “summarize this PDF” task.
Build better agents with Arize
Trace, evaluate, and learn. Build agents that work with Arize AX and start tracing your runs today.
Prefer open source? Try Arize Phoenix for self-hosted, open source agent observability.
Watch
Listen
Highlights
Why medicine is a special (and high-stakes) domain
Mukerji: “The medical field is one of the highest-stakes applications of AI—especially in real-world evidence. It’s very important to get things right, and approaches that work elsewhere don’t always translate to medicine.”
The “evidence gap” & the role of RWE
Mukerji: “Only a fraction of daily medical decisions are backed by high-quality evidence—the ‘evidence gap.’ RWE complements clinical trials by covering larger, more varied cohorts that trials often exclude.”
What RWESummary actually tests
Mukerji: “Our task isn’t ‘dump a full paper into an LLM.’ We supply structured study inputs—question, PICOT, outcomes, effect sizes—and ask for a plain-English summary. Current benchmarks rarely test this scenario.”
The three mission-critical evals
Mukerji: “We empirically chose three evaluations that matter most in production: getting the direction of effect right, reporting numbers accurately, and completeness. Direction of effect is the single most important.”
Why direction of effect is weighted so heavily
Mukerji: “If you could fix only one class of errors, fix direction of effect. Saying Treatment X is better than Y when it’s actually the reverse is the worst possible mistake.”
Results, speed, and trade-offs
Mukerji: “No model wins across the board. Gemini 2.5 led overall on direction and numbers; Gemini 2.0 Flash was the clear winner on latency. In production, inference time matters.”
Robust evals essential
Mukerji: “Given the stakes, you need robust evals and a human-in-the-loop. AI can make serious mistakes—like reversing an effect—so health workflows must be designed for safety from the ground up.”
