Med-CoReasoner: Multilingual Medical Reasoning

• Med-CoReasoner is a language-informed co-reasoning framework that integrates parallel English and local-language reasoning to produce culturally grounded clinical outputs.
• It abstracts free-form reasoning into ordered medical concepts via position-aware fusion and concept alignment, ensuring logical consistency and clinical localization.
• Empirical results show significant performance gains, with improvements of up to 9 percentage points in low-resource languages and robust expert validation.

Med-CoReasoner is a language-informed co-reasoning framework designed to address the persistent gap in multilingual medical reasoning exhibited by LLMs. Although reasoning-augmented LLMs perform robustly in English, their ability substantially degrades in local languages, impeding equitable deployment of clinical AI worldwide. Med-CoReasoner elicits and fuses parallel English and local-language reasoning, abstracts these into structured medical concepts, and integrates region-specific clinical knowledge within a logical English scaffold, using concept-level alignment and targeted knowledge retrieval. This architecture produces reasoning traces that are both clinically sound and culturally grounded, as confirmed by expert evaluation, and demonstrably improves multilingual performance, particularly in low-resource languages (Gao et al., 13 Jan 2026).

1. Pivot-Anchored Co-Reasoning Framework

Med-CoReasoner implements a multi-stage pipeline that encompasses the following stages:

• Parallel elicitation of English and local-language reasoning chains from the same medical question.
• Abstraction of free-form reasoning into ordered concept chains.
• Concept-level alignment and fusion, combining local-language clinical nuances with the logical structure of English chains.
• Retrieval of pertinent documents from multilingual knowledge bases using the fused concept graph.
• Final synthesis to generate a culturally and clinically coherent answer in the local language.

Mathematically, for an LLM $\mathcal{M}$ and multilingual KB $\mathcal{K}$:

• $R_e = \mathcal{M}(Q_e)$ where $Q_e = \text{translate}(Q_l)$
• $R_l = \mathcal{M}(Q_l)$
• $C_e = \phi(R_e)$, $C_l = \phi(R_l)$ with concept abstraction $\phi$
• $C_f = \text{Fusion}(C_e, C_l)$
• $D = \text{Retrieve}(C_f;\mathcal{K})$
• $A_l = \mathcal{M}(Q_l, C_f, D)$

Alignment between concepts $c_e^i$ and $c_l^j$ is scored as $$s(c_e^i, c_l^j) = \text{cosine}(f(c_e^i), f(c_l^j))$$, using a multilingual embedding function $f(\cdot)$ (BGE-M3). Local concepts with maximum similarity above threshold $\tau=0.5$ are inserted into $C_f$ at the contextually appropriate position according to “Position-Aware Concept Fusion” [(Gao et al., 13 Jan 2026), Algorithm 1].

2. Parallel Reasoning Elicitation

Med-CoReasoner explicitly elicits two independent chains of reasoning:

• English chain-of-thought (CoT): Prompt—“Think step by step in English, then answer.”
• Local-language CoT: Prompt—“Think step by step in [local language], then answer.”

This parallel, non-overlapping design enforces unbiased, native-language reasoning and boosts diversity, surfacing clinically relevant cues unique to each language and cultural context. Temperature tuning and chain-of-thought prompting strategies ensure both comprehensive step coverage and diversity—including region-specific vocabulary and practices.

3. Structured Concept Abstraction, Alignment, and Fusion

Each chain-of-thought output $R$ is abstracted into an ordered chain of atomic medical concepts $$C = \phi(R) = [c^1 \rightarrow c^2 \rightarrow \ldots \rightarrow c^k]$$, typically using the LLM with a specialized “Concept Extraction” prompt yielding a structured list (e.g., JSON format).

Concept fusion proceeds via:

$$C_f = C_e \cup \{c_l \in C_l : \max_{c_e \in C_e} s(c_e, c_l) > \tau\}$$

Matched local concepts are positioned next to their English backbone alignments, as determined by the maximal alignment score and the bidirectional context scan (see Algorithm 1 in the Appendix of (Gao et al., 13 Jan 2026)). This approach allows the system to preserve the logical rigor of English reasoning while augmenting it with culturally or clinically specific elements that would be lost in simple translation pipelines.

4. Multilingual Benchmarking with MultiMed-X

Performance is quantified using MultiMed-X, a benchmark specifically constructed to test medical reasoning in seven non-English languages (Chinese, Japanese, Korean, Thai, Swahili, Yoruba, Zulu). For each language, MultiMed-X contains:

• 200 long-form question answering (LFQA) items from LiveQA, expert-revised.
• 150 medical natural language inference (NLI) items from BioNLI, expert-revised.

Each item is annotated by two clinicians per language (except Yoruba), focusing on medical correctness, cultural alignment, and parallelism between language pairs. The resulting dataset comprises 2,450 expert-reviewed instances requiring open-ended, multi-sentence reasoning and complex terminology, enabling robust cross-lingual evaluation of clinical AI (Gao et al., 13 Jan 2026).

5. Empirical Results and Performance Gains

Med-CoReasoner exhibits significant improvement across multiple medical reasoning benchmarks:

• Global-MMLU (Medical MCQA):
  • GPT-5.1 baseline average: 84.11%
  • Med-CoReasoner: 88.61% (+4.5 pp)
• MMLU-ProX (Health):
  • GPT-5.1: 71.39%
  • Med-CoReasoner: 77.42% (+6.0 pp)
• Low-resource performance:
  • Swahili: +8 pp improvement on both MCQA datasets
  • Yoruba (LFQA): pass rate increase of +9%
• MultiMed-X LFQA (GPT-5.1 backbone):
  • Quality (Likert 1–5): ~4.33 to 4.60
  • Completeness: ~4.54 to 4.85
  • Pass Rate (Quality & Safety ≥4): 0.89 to 0.94

Distillation experiments show further transferability: fine-tuning Qwen2.5-14B on Med-CoReasoner derived traces achieves an average +2.86 pp improvement over baselines, with similar gains for other mid-sized LLMs. In expert pairwise reviews, Med-CoReasoner outperforms the baseline for localization (60% win rate) and clarity (50% win, 26.7% tie), while maintaining high standards of soundness and safety (>80% win+tie) (Gao et al., 13 Jan 2026).

6. Clinical and Cultural Contributions

Expert physicians emphasize several advantages of Med-CoReasoner:

• Localization: Integration of regionally specific terminology, drug names, and guideline citations.
• Clarity: Step-by-step reasoning closely mirrors established clinical teaching and case review methods.
• Safety: Improved explicitness, reduced speculation, and identification of unknowns or uncertainties. Direct alignment and fusion of local-language reasoning enables retention of culturally salient details, such as regionally preferred diagnostic procedures or test availability, that are often omitted by English-centric LLM pipelines. This ensures that outputs are not only accurate but also contextually and culturally appropriate for diverse medical environments (Gao et al., 13 Jan 2026).

7. Scalability and Generalization

The core Med-CoReasoner pipeline, including position-aware concept fusion and retrieval augmentation, is amenable to model distillation: distilled student models retain much of the accuracy and cultural coverage of the original system, enabling practical deployment without prohibitive computation or latency (Gao et al., 13 Jan 2026). A plausible implication is the potential extension of the Med-CoReasoner framework beyond medicine to other multilingual, high-stakes reasoning domains—where parallel, concept-level co-reasoning can bridge representational gaps and enhance cultural robustness.