Research map / Large language models / Reasoning and chain of thought

CLR-voyance : Reinforcing Open-Ended Reasoning for Inpatient Clinical Decision Support with Outcome-Aware Rubrics

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: RL with verifiable rewards for reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces CLR-voyance, which reformulates inpatient clinical reasoning as a POMDP built from MIMIC-IV and trains Qwen3-8B and MedGemma-4B with GRPO on per-case outcome-grounded rubrics, followed by model merging, with clinician validation and hospital deployment.

Problem

Inpatient clinical decision support is a sequential decision under partial observability, but existing clinical LLM evaluations and RL rewards either use closed-form retrieval, leak the full patient journey, or rely on unanchored LLM-as-judge scoring.

Approach

Split each MIMIC-IV admission into a policy-visible past and an oracle-only future; an oracle LLM (Grounded Judge) writes a past-answerable query, a future-evidenced reference answer, and a per-case adaptive rubric that is graded to give a dense RL reward, with the policy never seeing the future or rubric.

Key results

• CLR-voyance-8B reaches 84.91% aggregate on the CLR-POMDP test split, versus 77.83% for GPT-5 and 66.66% for MedGemma-27B.
• On DDXPlus, CLR-voyance-8B reaches 48.5%, the only 8B-class model in the evaluation to outperform GPT-5 and HuatuoGPT-o1 there.
• Four physicians preferred CLR-voyance-8B over GPT-5 in the spine cohort at 85.6% win rate (CI 78-91); in the obesity cohort it was 48.0% vs GPT-5.

Surprising findings

• Off-the-shelf GPT-5 was the worst-aligned judge against clinician consensus in both cohorts, with 100% self-preference on the GPT-5 vs CLR-voyance pair.
• SFT on the oracle reference answers regressed Qwen3-8B from 75.32% to 52.36%, while GRPO on rubric rewards improved it.

Limitations

• Model weights, generated rubrics, and patient timelines cannot be released due to NDAs and privacy contracts.
• Clinician validation covers only two specialty cohorts (spine/orthopaedic and obesity/general medicine) with four physicians.
• The 4B merge reduced communication quality and instruction-following scores relative to the GRPO model.

Experimental setup

Datasets: MIMIC-IV MIMIC-IV-ED MIMIC-IV-Note MedMCQA DDXPlus MedCalc-Bench Mimic-Instr HealthBench

Benchmarks: CLR-POMDP MedMCQA DDXPlus MedCalc-Bench Mimic-Instr HealthBench

Metrics: rubric score (aggregate, per-axis) accuracy win rate Cohen's kappa Wilson 95% CI Wilcoxon signed-rank p-value

Baselines: GPT-5 GPT-4.1 GPT-4.1-mini MedGemma-27B HuatuoGPT-o1-7B DeepSeek-R1-Distill-Qwen-7B DeepSeek-R1-0528-Qwen3-8B Qwen3-8B base

Builds on: GRPO HealthBench adaptive rubrics DELLA Breadcrumbs

Closest papers by meaning

• MedS³: Towards Medical Slow Thinking with Self-Evolved Soft Dual-sided Process Supervision (AAAI 2026 · similarity 0.87)
• ER-Reason: A Benchmark Dataset for LLM Clinical Reasoning in the Emergency Room (NeurIPS 2026 · similarity 0.86)
• Simulating Viva Voce Examinations to Evaluate Clinical Reasoning in Large Language Models (NeurIPS 2025 · similarity 0.86)
• Large Language Models Are Clinical Reasoners: Reasoning-Aware Diagnosis Framework with Prompt-Generated Rationales (AAAI 2024 · similarity 0.86)
• CARE: Towards Clinical Accountability in Multi-Modal Medical Reasoning with an Evidence-Grounded Agentic Framework (ICLR 2026 · similarity 0.85)
• MTBBench: A Multimodal Sequential Clinical Decision-Making Benchmark in Oncology (NeurIPS 2025 · similarity 0.85)