AI Insight
Researchers developed Infoxmed2.0-27B, a specialized medical large language model built on Qwen3.5-27B through a four-stage training pipeline including medical data synthesis, instruction fine-tuning, preference optimization, and reward model training. The model achieved 77.0% accuracy on the MedMCQA medical question-answering benchmark with progressive improvements across training stages, demonstrating enhanced performance in medical domain tasks compared to the base model.
Why it matters
This work advances the adaptation of general-purpose AI models for specialized medical applications, potentially improving clinical decision support systems and medical information retrieval. The multi-stage training approach and comprehensive evaluation framework provide a methodology for developing domain-specific LLMs in healthcare settings.
Understand the Science
⚠️ Preprint – Noch nicht peer-reviewed
Dieser Artikel wurde noch nicht von unabhängigen Experten begutachtet. Die Ergebnisse sind vorläufig und sollten mit Vorsicht interpretiert werden.
Abstract-Large language models (LLMs) [1], [2] have demon strated remarkable capabilities across general domains, yet their application in specialized medical contexts demands rigorous domain adaptation [3], [4]. We present Infoxmed2.0-27B, a medical foundation model built upon Qwen3.5-27B [5] through a comprehensive multi-stage post-training pipeline: (1) proprietary medical data synthesis from a MySQL database with MedicalCategoryTree organization, medical PhD team validation, Chinese RoBERTa [6] semantic deduplication, and API-assisted language refinement; (2) instruction supervised fine-tuning of Qwen3.5- 27B via LoRA [7] (r = 8, = 32) using MS-Swift [8], producing iterations Infoxmed2.0.0[->]2.0.2[->]2.0.4; (3) Direct Preference Optimization (DPO) [9] on 6,283 curated medical preference pairs [10] using DPO-RPO loss ({beta} = 0.3, RPO = 0.1) across eight progressive training iterations (v0-v7); and (4) parallel Group Relative Policy Optimization (GRPO) [11]-based medical reward model training on Qwen3.5 combining internal rule-based reward functions with external DeepSeek signals. Comprehensive evaluations under a uniform LLM-as-Judge [12] framework with GPT-5.4 demonstrate 77.0% accuracy (mean quality score +7.18) on MedMCQA [10] and +2.59 on HLE, with pipeline progression from +6.69 (base) to +7.06 (SFT) to +7.18 (final).