The main experiment compared Qwen3-4B-Base with a counterpart trained using reinforcement learning with verifiable rewards.
Sparse RL-Guided Token Choices Recover Most Math Reasoning Gains
Replacing about 4% of a base model’s generated tokens with an RL-trained model’s choices recovered 90.5% of the reinforcement-learning improvement on AIME.
Top University
Siyu Zhang · Yuquan Lu · Yoshua Y. Li · Yifan Ding · Kun Zeng · Yu Huo · +4 more
University of California, San Diego · Meituan, LongCat Team · The Chinese University of Hong Kong, Shenzhen
Research Digest··3 min read
The authors use controlled decoding interventions to identify which behavioral changes account for an RL-trained language model’s stronger mathematical reasoning.
§