CURE-Med: Curriculum-Informed Reinforcement Learning for Multilingual Medical Reasoning

TL;DR

The paper tackles unreliable multilingual medical reasoning by introducing CureMed-Bench, a MedlinePlus-grounded, open-ended multilingual QA dataset across 13 languages, and Cure-Med, a two-stage training framework that combines code-switching-aware supervised fine-tuning with curriculum-guided reinforcement learning (GRPO). CureMed-Bench enables independent evaluation of logical correctness and language fidelity, while Cure-Med jointly optimizes reasoning and linguistic adherence, attaining up to $70.04\%$ logical correctness and $94.96\%$ language consistency at 32B and $54.35\%$ and $85.21\%$ at 7B, demonstrating scalable multilingual medical reasoning and improved out-of-distribution generalization. The methodology comprises (i) cold-start initialization with code-switched long CoT trajectories, (ii) a composite reward combining clinical accuracy, language fidelity, and format, and (iii) a language-resource-aware curriculum RL that progressively includes lower-resource languages. Results show Cure-Med outperforms baselines across parameter scales, with ablations confirming the necessity of both code-switched SFT and curriculum RL for robust multilingual reasoning. The work provides a valuable benchmark and open-source pipeline to support reliable and equitable multilingual medical guidance in LLMs.

Abstract

While large language models (LLMs) have shown to perform well on monolingual mathematical and commonsense reasoning, they remain unreliable for multilingual medical reasoning applications, hindering their deployment in multilingual healthcare settings. We address this by first introducing CUREMED-BENCH, a high-quality multilingual medical reasoning dataset with open-ended reasoning queries with a single verifiable answer, spanning thirteen languages, including underrepresented languages such as Amharic, Yoruba, and Swahili. Building on this dataset, we propose CURE-MED, a curriculum-informed reinforcement learning framework that integrates code-switching-aware supervised fine-tuning and Group Relative Policy Optimization to jointly improve logical correctness and language stability. Across thirteen languages, our approach consistently outperforms strong baselines and scales effectively, achieving 85.21% language consistency and 54.35% logical correctness at 7B parameters, and 94.96% language consistency and 70.04% logical correctness at 32B parameters. These results support reliable and equitable multilingual medical reasoning in LLMs. The code and dataset are available at https://cure-med.github.io/

Figures (11)

• Figure 1: The Cure-Med pipeline for multilingual medical reasoning. The framework progresses through three stages: (A) curation of clinically validated multilingual data from sources like MedlinePlus to enable cross-lingual reasoning; (B) supervised fine-tuning of the Qwen2.5-Instruct backbone on code-switched reasoning traces; and (C) GRPO-guided curriculum reinforcement learning, progressively training from high- to mid- and low-resource languages to enhance logical correctness and language consistency.
• Figure 2: An example from the cold-start multilingual dataset showing CoT reasoning in French. The reasoning combines English-based clinical terms and local-language expressions, reflecting code-switching in medical contexts.
• Figure 3: Qualitative Spanish medical-reasoning example comparing a baseline Qwen2.5-7B-Instruct model and Cure-Med-7B. The baseline model produces fluent but clinically flawed reasoning (red) and an incorrect diagnosis, whereas Cure-Med generates a structured, code-switched CoT (blue) and arrives at the correct diagnosis (green).
• Figure 4: Trade-off performance between logical of multilingual medical reasoning models, where each point represents a model instance with bubble size reflecting model scale. Baseline and Cure-Med models are shown as and $\bigstar$, respectively. Cure-Med shifts performance toward the upper-right, indicating consistent gains in language consistency and logical accuracy.
• Figure 5: Scaling performance of Cure-Med vs. base across Qwen2.5-Instruct variants on language consistency (left) and logical accuracy (right). Our method (solid red line) consistently outperforms the base model (dashed blue line), with performance gaps widening at larger model scales, highlighting the effectiveness of Cure-Med for multilingual medical reasoning.