Found first: a primary source the press has not covered yet.
Researchers from Renmin University of China, Tsinghua University, the University of Electronic Science and Technology of China, and Shanghai Jiao Tong University have trained a 4B process reward model that scores 85.2 F1 on ProcessBench using only 3,200 outcome-labeled trajectories, surpassing both GPT-5.4-Instruct and Claude-4.7-Opus. The paper, "Inducing Process Supervision from Outcome-Only Reinforcement Learning," is at arxiv.org/abs/2609.36641.
What the source says
The method, TIPS (Thinking-Induced Process Supervision), trains a generative PRM via reinforcement learning where the reward depends solely on whether the model's predicted outcome label matches ground truth. No step-level annotations are used. TIPS-Qwen3-4B-Thinking-2507 scores 85.2 F1 on ProcessBench, beating all evaluated trained PRMs including Qwen2.5-Math-PRM-72B (78.3) and GenPRM-32B (77.6), and both prompt-only frontier judges GPT-5.4-Instruct (79.4) and Claude-4.7-Opus (79.3). The model still trails o1-mini, which scores 87.9 on the same benchmark. Authors are Shengda Fan, Xin Cong, Zhong Zhang, Haotian Chen, and Yankai Lin; code and data are at github.com/RUCBM/TIPS.
Why it matters
Training process reward models has required step-level labels, which means human annotators evaluating every intermediate reasoning step across thousands of trajectories. TIPS shows that outcome-only RL training causes the model to learn step verification as an emergent behavior, eliminating that annotation requirement. A 4B model trained on 3,200 outcome labels outperforming 72B-class trained PRMs and frontier prompt judges suggests annotation cost has been the binding constraint on process supervision quality. The 2.7 F1 gap to o1-mini marks where the current ceiling sits.