The authors followed open-model training histories, including OLMo-3 checkpoints, released RL trajectories, and models fine-tuned on reasoning traces.
Reasoning models usually inherit “wait” from training data, not reinforcement learning
Across open-model training histories, the authors find that reinforcement learning can improve mathematics accuracy without introducing “wait”, while different training recipes amplify or erase an inherited marker.
Chinese Tech
Zhizhuo Kou · Sizai Hou · Long Chen · Jian Yang · Junyu Luo · Jingshu Peng · +3 more
The Hong Kong University of Science and Technology · Huawei Noahah’s Ark Lab · Peking University
Research Digest··2 min read
Kou et al.
§