Qwen3.8-27B trained with binary-gated GRPO boosts GDPval pass@1

s46486.pcdn.co9 min read

Researchers trained the Qwen3.8-27B model on professional tasks using expert-built training data and binary-gated GRPO, raising its pass@1 rate on OpenAI's GDPval gold set from 31.9% to 40.7%. The training method addresses the problem of sparse learning signals in difficult tasks by falling back on a detailed training score when all group rollouts fail. The trained model also outperformed human experts in 44.0% of LLM-judged head-to-head comparisons. Open article →