Research · Friday 16 October 2026 · story 5
Paper says RL gains in Qwen3 NO-THINK mostly reactivate THINK mode
Hugging Face reports a paper finds RL gains in Qwen3 NO-THINK mode largely come from reactivating THINK-mode reasoning, not new capabilities. The authors say Qwen3-1.7B, 4B, and 8B trained with GRPO grew more THINK-like in responses, activations, and attention heads.
Why it matters. For agent builders, RL gains in NO-THINK mode may reuse existing reasoning in Qwen3 rather than add new reasoning behavior.
Read the original at huggingface.co