7B, 4B, 8B) with Group Relative Policy Optimization (GRPO) in no-think mode on the DAPO-Math-17k dataset.
RL gains in no-think mode stem from reactivating existing reasoning
Training Qwen3 models in no-think mode with GRPO primarily shifts behavior toward the model's own think mode, accounting for most accuracy improvements.
Top University
Wenxuan Zhu · Shuming Liu · Bang An · Aznaur Aliev · Bing Li · Yibo Yang · +1 more
King Abdullah University of Science and Technology · Shanghai Jiao Tong University
Research Digest··3 min read
7B, 4B, and 8B models with GRPO in no-think mode on the DAPO-Math-17k dataset.
Why this paper
From Shanghai Jiao Tong University and King Abdullah University of Science and Technology · Part of Reward Collapse in Reasoning RL, now 10 papers
In one line
Reinforcement learning gains in no-think mode of hybrid LLMs largely come from reactivating the existing think mode reasoning.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§