ICU-REACT paper: teaching LLMs how ICU physicians reason beats pure scaling
SoloGen · x · 2026-10-24
A new arXiv preprint (2608.22622) from Parisa Rashidi's team introduces ICU-REACT, a reasoning dataset built with 19 clinicians via a clinician-in-the-loop framework that teaches LLMs how ICU physicians select and reason over decision-relevant evidence, using OMOP-aligned retrieval rather than surface-level recall.
- The pipeline covers seed data creation, augmentation, and reasoning-refinement SFT on open backbones (Llama, Baichuan, Gemma), producing Clin-REACT models.
- Evaluation spans an internal ICU-REACT test plus external benchmarks (VivaBench, SCT Bench, MedRBench, ER Reason) with LLM-judge and automated metrics.
- Code, dataset, and models are released; the authors argue the approach improves clinical reasoning across domains beyond critical care.
More from AGI Musings
- The 'Bugpocalypse': LLMs Are Unearthing Long-Hidden Software Vulnerabilities at Furious Pace — esrtweet · 2026-10-24
- Dan Jeffries: AI Finding Vulnerabilities Speedruns Safer Critical Systems — Dan_Jeffries1 · 2026-10-24
- Mathematician: AI is revealing the space of mathematical ideas as hyperbolic — CatAstro_Piyush · 2026-10-24
- Altman: models find hardware-level bugs 'outside the normal threat model', patching won't scale — soldierofcinema · 2026-10-24
- Reading《菌落万象》: intelligence is just an anthropocentric benchmark AI is about to top — oran_ge · 2026-10-24
- CS grad: AI is killing my motivation to learn anything new — kubika7 · 2026-10-24