Releases·Model Release
FreedomIntelligence releases HuatuoGPT-3-9B, a medical LLM trained with one-stage RL (OnePO) on Qwen3.5-9B
FreedomIntelligence released HuatuoGPT-3-9B, a medical LLM built on Qwen3.5-9B and trained with One-stage Policy Optimization (OnePO), a reinforcement-learning method that adapts models to medicine without prior domain-specific supervised fine-tuning. The release includes training code, a 20K medical RL dataset, and an 8B rubric grader, giving practitioners a reproducible medical-RL recipe.
24 engagement1 sourceThu, Sep 17, 2026, 04:30 PM
AI Overview
- FreedomIntelligence's HuatuoGPT-3-9B adapts Qwen3.5-9B to medicine via OnePO, single-stage reinforcement learning without domain-specific SFT
- OnePO uses teacher responses as temporary guidance, retired as the model improves
- FreedomIntelligence released training code, a 20K-sample medical RL dataset (OnePO-Medical-20K), and an 8B rubric grader
Related
Releases·Model Release
FreedomIntelligence releases HuatuoGPT-3-27B, a medical LLM trained with one-stage RL (OnePO) on Qwen3.8-27B
FreedomIntelligence released HuatuoGPT-3-27B, a medical LLM built on Qwen3.8-27B and trained with One-stage Policy Optimization (OnePO), a reinforcement-learning method that adapts models to medicine without prior domain-specific supervised fine-tuning. The release includes training code, a medical RL dataset, and an 8B rubric grader, following a 9B variant released the prior week.
FreedomIntelligence released HuatuoGPT-3-27B, a medical LLM built on Qwen3.8-27B and trained with One-stage Policy Optimization (OnePO), a reinforcement-learning method that adapts models to medicine without prior domain-specific supervised fine-tuning. The release includes training code, a medical RL dataset, and an 8B rubric grader, following a 9B variant released the prior week.
2 engagement1 sourceredditThu, Sep 24, 2026, 07:20 PM
Developments·Paper
HuatuoGPT-3: RL-Only Domain Adaptation from Base Models
A paper proposes OnePO, a one-stage RL-only domain adaptation method that treats teacher outputs as transient guidance, addressing Gradient Starvation and Teacher-Distribution Anchoring. Using it, the team built HuatuoGPT-3, an open-source medical LLM series whose 27B variant surpasses frontier models on HealthBench. Models and code are public on GitHub.
A paper proposes OnePO, a one-stage RL-only domain adaptation method that treats teacher outputs as transient guidance, addressing Gradient Starvation and Teacher-Distribution Anchoring. Using it, the team built HuatuoGPT-3, an open-source medical LLM series whose 27B variant surpasses frontier models on HealthBench. Models and code are public on GitHub.
0 engagement1 sourcearxivMon, Oct 5, 2026, 08:16 AM
Releases·Model Release
EpistemeAI releases Reasoning-Medical-27B, a Qwen3.6-27B finetune for medical reasoning
EpistemeAI released Reasoning-Medical-27B, a fine-tuned version of Qwen3.6-27B designed for advanced medical reasoning. The model was trained on 370,000 QA examples with Chain-of-Thought reasoning using GRPO and Unsloth optimization.
EpistemeAI released Reasoning-Medical-27B, a fine-tuned version of Qwen3.6-27B designed for advanced medical reasoning. The model was trained on 370,000 QA examples with Chain-of-Thought reasoning using GRPO and Unsloth optimization.
87 engagement1 sourceredditTue, Jul 28, 2026, 07:19 AM
Developments·Paper
PIHF paper proposes policy iteration with human feedback for in-context learning
A new arXiv paper introduces Policy Iteration with Human Feedback (PIHF), a method that applies post-training reinforcement learning to in-context learning. PIHF uses a pretrained language model as its execution substrate and maintains a versioned natural-language policy and tool set, with a language-model critic and clinical expert reviewing complete-panel reasoning.
A new arXiv paper introduces Policy Iteration with Human Feedback (PIHF), a method that applies post-training reinforcement learning to in-context learning. PIHF uses a pretrained language model as its execution substrate and maintains a versioned natural-language policy and tool set, with a language-model critic and clinical expert reviewing complete-panel reasoning.
0 engagement1 sourcearxivMon, Aug 17, 2026, 05:16 PM
Developments·Paper
ConRub-Med: RL with Consensus Rubrics for Open-Ended Medical QA
Researchers propose ConRub-Med, a reinforcement learning method that uses consensus rubrics to provide verifiable rewards for open-ended medical question answering. The approach addresses the lack of cheap outcome verifiers in medical domains by using physician-validated or model-generated rubrics, making supervision scalable while preserving clinical grounding.
Researchers propose ConRub-Med, a reinforcement learning method that uses consensus rubrics to provide verifiable rewards for open-ended medical question answering. The approach addresses the lack of cheap outcome verifiers in medical domains by using physician-validated or model-generated rubrics, making supervision scalable while preserving clinical grounding.
0 engagement1 sourcearxivTue, Aug 11, 2026, 02:48 PM