AI

HuatuoGPT-3: RL-Only Domain Adaptation from Base Models

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 07/10/2026, 03:18:24 UTC.

Briefings

HuatuoGPT-3 uses One-stage Policy Optimization (OnePO) to adapt large language models to medical domains, achieving 67.2 on HealthBench with 20K samples, outperforming SFT+RL by 2.7 points and pure RL by 7.4 points. The 27B variant reaches 70.1 on HealthBench (Total) and 71.4 on HealthBench Professional, surpassing GPT-6 Astra. The method avoids cold-start issues by treating teacher outputs as transient guidance, combining Adaptive Objective Evolution and Teacher Retirement.

Key points

• OnePO achieves 67.2 on HealthBench (Total) using only 20K training samples, outperforming SFT+RL by 2.7 points and pure RL by 7.4 points.
• The 27B variant of HuatuoGPT-3 reaches 70.1 on HealthBench (Total) and 71.4 on HealthBench Professional, surpassing GPT-6 Astra.

Context

Domain adaptation converts general-purpose LLMs into domain experts, but SFT+RL pipelines often reduce exploration diversity and introduce multi-stage optimization complexity. OnePO addresses these by using teacher outputs as transient guidance, avoiding cold-start and stale output issues.

Limitations

Results are from a single snapshot on HealthBench; no additional background on methodology or limitations is supplied. Evidence is untrusted data, preprint status is unknown, and causal claims remain unverified.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 25/10/2026, 00:04:53 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeics2fhfqwsnhaxoqjjxpi6cp23sih6jh3ap2wfvaq3tjzxjoy3sfu