Source-linked AI summary
CURE-Med: Curriculum-Informed Reinforcement Learning for Multilingual Medical Reasoning
Eric Onyame, Akash Ghosh, Subhadip Baidya, Sriparna Saha, Xiuying Chen, Chirag Agarwal
TL;DR
LLMs remain unreliable for multilingual medical reasoning, while existing resources underrepresent diverse languages and provide limited tests of reasoning quality. The paper introduces CUREMED-BENCH and CURE-MED, combining code-switching-aware supervised fine-tuning with curriculum-informed reinforcement learning. Across 13 languages, CURE-MED outperforms strong baselines and reaches 54.35% logical correctness with 85.21% language consistency at 7B parameters, and 70.04% logical correctness with 94.96% consistency at 32B.
Problem
Multilingual medical reasoning remains unreliable because reasoning-focused resources and evaluations underrepresent mid- and low-resource languages and language fidelity.
Method
CURE-MED combines code-switching-aware supervised fine-tuning with curriculum-informed GRPO reinforcement learning across high-, medium-, and low-resource language tiers.
Results
CURE-MED improves logical correctness and target-language consistency over baselines across model scales and languages, including 54.35% logical correctness and 85.21% consistency at 7–9B.
Takeaways & Limitations
The results support multilingual medical reasoning that is more reliable across languages, including improved robustness in low-resource settings and generalization to unseen questions and languages.
Takeaways & Limitations
CUREMED-BENCH has uneven language coverage and difficulty, excludes longitudinal, multi-visit, and multimodal clinical evidence, and partly relies on costly API-based models.
Abstract
from arXiv · showhide
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/
1 Introduction
Medical reasoning remains difficult for LLMs, especially in multilingual settings where benchmarks and resources underrepresent mid- and low-resource languages. CURE-MED addresses this gap with CUREMED-BENCH and a two-stage training framework combining code-switched supervision with curriculum-informed reinforcement learning.
- Medical reasoning requires domain knowledge, contextual interpretation, and inspectable reasoning beyond strong performance on mathematics and code generation.
- Multilingual medical reasoning is limited by reasoning-centric data and evaluations, with models producing fluent outputs that may lack dependable reasoning and language fidelity.
- CUREMED-BENCH evaluates open-ended multilingual medical reasoning across 13 high-, mid-, and low-resource languages, using queries with single verifiable answers.
- CURE-MED combines code-switching-aware supervised fine-tuning with curriculum-informed reinforcement learning to jointly optimize logical correctness and language fidelity.
- CURE-MED achieves state-of-the-art performance and improved out-of-distribution generalization, including stronger robustness in low-resource languages and on unseen questions and languages.
2 Related Work
Prior medical LLM work improves domain capability but does not consistently establish reliable reasoning, while multilingual evaluations remain concentrated on general-domain tasks and higher-resource languages. CURE-MED responds by jointly optimizing reasoning correctness and language fidelity across diverse resource settings.
- Large Medical Reasoning Models: Domain-specific pretraining and instruction tuning can improve factuality, but benchmark gains do not consistently translate into reliable medical reasoning.
- Large Medical Reasoning Models: Medical models may produce fluent but clinically unsound explanations, motivating evaluations that test reasoning rather than answer plausibility.
- Multilingual Reasoning and Language Fidelity: Cross-lingual reasoning evaluations mostly target general-domain mathematical or symbolic tasks and favor high-resource languages.
- Multilingual Reasoning and Language Fidelity: In medical settings, multilingual models exhibit degraded accuracy, language drift, and weak cross-lingual generalization.
- Multilingual Reasoning and Language Fidelity: CURE-MED jointly optimizes language fidelity and reasoning correctness across high-, mid-, and low-resource languages using code-switching-aware supervision and curriculum-informed reinforcement learning.
3 Methodology
CURE-MED combines a multilingual, open-ended medical reasoning benchmark with code-switching-aware initialization and curriculum-guided reinforcement learning. Its training and reward design jointly target clinical correctness, target-language fidelity, and structured outputs across resource tiers.
- Dataset Construction: Each benchmark instance has a clinically grounded solution with an explicit reasoning trace and a single verifiable answer.The construction removes multiple-choice cues while retaining supervised reasoning and answer correctness targets.
- Cold-Start SFT: CURE-MED initializes multilingual reasoning with cold-start SFT on code-switched long CoT trajectories, keeping the final answer in the target language.Intermediate reasoning may use controlled language mixing, while the final response remains in the query language.
- Reward Design: The weighted reward promotes clinical correctness, language fidelity, and compliance with a structured output format.Verification scores semantic and clinical equivalence, language consistency checks the target language, and a parser checks thinking, numbered-step, and answer tags.
- Curriculum-Guided GRPO: Curriculum-guided GRPO progresses from high- to medium- to low-resource languages while retaining prior-phase data to reduce catastrophic forgetting.The curriculum uses language resource availability as difficulty and mixes retained data with the current tier using α=0.85.
4 Experiments
The experiments evaluate CURE-MED against 28 baselines across 13 languages using held-out data and separate measures of logical accuracy and language consistency. The setup tests baseline improvement, trade-offs, scaling, and out-of-distribution generalization.
- Research Questions: The experiments address whether CURE-MED improves multilingual medical reasoning, balances language fidelity with accuracy, scales with curriculum learning, and generalizes out of distribution.These questions correspond to RQ1 through RQ4.
- Experimental Setup: The dataset uses an 80% train and 20% held-out test split, with the training portion divided between supervised and reinforcement fine-tuning.The train set allocates 80% to supervised fine-tuning and 20% to reinforcement fine-tuning.
- Experimental Setup: CURE-MED is benchmarked against 28 baseline models built from general-purpose instruction-tuned backbones across multiple parameter scales.The listed Qwen, LLaMA, and Gemma backbones include models from 1.5B through 32B parameters.
- Results Reporting: Table 2 reports mean results across 13 languages and compares CURE-MED with baselines at each parameter scale.The table distinguishes language consistency from logical accuracy and marks best overall and best-baseline results.
- Evaluation: Evaluation uses an LLM-as-a-judge framework on held-out data, measuring logical accuracy and target-language consistency.Logical accuracy concerns the clinical accuracy of the final answer, while language consistency concerns whether it is produced in the question’s target language.
5 Results
CURE-MED improves both logical accuracy and target-language consistency over baselines across model scales. It also transfers to held-out multilingual medical benchmarks, with especially clear gains for smaller models.
- 70.04% logical correctness and 94.96% consistency are achieved at ≥14B, while 54.35% correctness and 85.21% consistency are reached at 7–9B.
- CURE-MED shifts models toward higher language consistency and logical accuracy than baselines.
- CURE-MED improves accuracy over Qwen2.5 backbones in most language–scale settings, with the clearest gains for smaller models.
- At 1.5B, representative transfer gains include 6.00→24.00 and 20.00→57.50 on MMedBench, 1.40→44.80 on MedExpQA, and 21.00→59.50 on MedQA.
6 Ablation Study
Ablations attribute CURE-MED’s gains to structured code-switched SFT and curriculum- and language-resource-aware reinforcement learning. Naïve multilingual SFT and single-stage reinforcement produce smaller or uneven improvements.
- Code-switched supervised fine-tuning: Naïve multilingual SFT increases 3B language consistency from 8.39%→13.07% but decreases logic accuracy from 10.83%→9.50%.
- Code-switched supervised fine-tuning: Code-switched SFT raises 1.5B language consistency from 3.84%→53.67% and logic accuracy from 6.20%→22.97%.At 32B, it reaches 90.29% language consistency and 66.34% logic accuracy.
- Curriculum reinforcement learning: Curriculum- and language-resource-aware RL improves language consistency and logical accuracy across all model sizes beyond naïve single-stage GRPO-based RFT.
- Medical benchmark comparison: CURE-MED-32B achieves 70.04% on CUREMED-BENCH and 79.57% on MMed-Bench, while remaining competitive on MedQA and MedExpQA.
7 Conclusion
The paper contributes a multilingual benchmark and a two-stage training framework for improving medical reasoning while preserving target-language fidelity. Across languages, datasets, and scales, the approach improves both logical correctness and language consistency.
- CUREMED-BENCH contains open-ended multilingual medical questions with explicit reasoning traces and a single verifiable answer across 13 languages.
- CURE-MED combines cold-start code-switched initialization, structured supervised fine-tuning, and language-resource-aware curriculum-RL.
- CURE-MED improves logical correctness and language consistency over strong baselines across languages, datasets, and model scales.
- Ablations show that the supervised and reinforcement-learning stages provide complementary gains for stable multilingual reasoning.
8 Limitations
CUREMED-BENCH has uneven language coverage and difficulty because clinically reliable source material varies across languages. Its open-ended, single-answer design and API-dependent pipeline also limit scope and reproducibility.
- Clinically reliable source availability limits coverage and can create uneven difficulty between high- and low-resource languages.
- Open-ended questions with single verifiable answers do not capture longitudinal care, multi-visit decisions, or multimodal clinical evidence.
- API-based generation and verification can be costly and hinder reproducibility for some researchers.
- Future work will expand languages, clinical settings, and modalities while reducing reliance on proprietary APIs.
9 Ethical Considerations
The work emphasizes clinically and linguistically reviewed multilingual data, explicit reward verification, and reproducible training procedures. Its pipeline combines human review with automated correctness, language, and format checks.
- Native speakers and medical experts reviewed samples for clinical correctness, linguistic fidelity, and cultural appropriateness under IRB-approved procedures.
- GPT-4o compares model predictions with reference answers and returns binary judgments of logical correctness and linguistic validity.
- The composite reward combines clinical correctness, language fidelity, and output-format compliance, prioritizing medical correctness while penalizing drift and violations.
- Correctness and language rewards use deterministic gpt-4.1 scoring, while a rule-based check requires exactly one <thinking> block and one <answer> block.
- Experiments use three independent runs with random seed 42 on a single node containing 8 NVIDIA A100 GPUs.
D.2 Construction of the SFT Dataset
The SFT dataset is built from teacher-generated multilingual reasoning traces and target-language answers, then filtered through clinical and native-speaker verification. The resulting data is organized across thirteen languages and assessed for quality.
- GPT-4o generates stepwise reasoning traces and final answers entirely in the target language for curated question–language–answer triplets.
- Human verification retains samples whose answers are clinically consistent with gold references and satisfy formatting and language constraints.
- Medical professionals score clinical correctness, while native speakers score target-language quality and fidelity on a 1–5 scale.
- Verification scores are generally high, though Amharic, Thai, and Turkish receive slightly lower ratings.
E.1 Per-Language Results for Qwen2.5-7B
CURE-MED improves Qwen2.5-7B across languages in logical correctness and language consistency, with especially large gains in low-resource languages. The results also expose persistent baseline brittleness and a mismatch between reasoning accuracy and target-language control.
- CURE-MED substantially improves both logical accuracy and language consistency, especially in Amharic, Hausa, Swahili, and Yoruba.
- In higher-resource French, Japanese, and Spanish, improvements are more moderate but consistent without degrading performance.
- Inference-only frontier models remain brittle: strong target-language adherence does not ensure robust multilingual reasoning or consistent language control.
- OPENBIOLLM-LLAMA3-8B reaches 36.62% logical accuracy but only 1.47% language consistency, demonstrating a mismatch between correctness and language fidelity.
- Qwen2.5-7B records 0.00% language consistency and near-zero logical accuracy in Amharic, Hausa, Swahili, and Yoruba, versus 71.43% language consistency in French.
- At 32B, CURE-MED reaches 94.96 language consistency and 70.04 logical accuracy, the strongest overall reported results.
H Curriculum Ordering Ablation
The ablation compares the default high-resource-to-low-resource curriculum against the reverse ordering across 1.5B, 3B, and 7B models. Reversing the schedule consistently degrades language consistency and logical accuracy, although CURE-MED remains stronger than size-matched instruction-tuned baselines.
- The default curriculum trains from high-resource languages to medium-resource and then low-resource languages, while the counterfactual reverses this order.
- The ablation is conducted only for the 1.5B, 3B, and 7B settings because of resource constraints.
- Reversing the curriculum order consistently degrades both language consistency and logical accuracy across all evaluated model scales.
- Even with the reverse ordering, CURE-MED remains stronger than representative size-matched instruction-tuned baselines.
- Table 14 reports language consistency and logical accuracy as mean ± standard deviation over three runs under the same evaluation protocol.