Found first: a primary source the press has not covered yet.
A new paper from KAUST-affiliated researchers finds that RL training of hybrid reasoning models in NO-THINK mode does not eliminate reasoning. Internal activations shift toward THINK mode even when the visible reasoning tokens are absent, a phenomenon the authors call "thinking leakage." The paper argues the performance gains most labs attribute to NO-THINK RL are largely an artifact of this shift.
What the source says
Wenxuan Zhu, Shuming Liu, Bang An, Aznaur Aliev, Bing Li, Yibo Yang, and Bernard Ghanem trained Qwen3 models at 1.7B, 4B, and 8B parameters using GRPO in NO-THINK mode. They identified a low-dimensional THINK-related activation subspace and ran two probes: replacing that subspace in the RL-trained model with values from the original model removed most of the RL gain, while injecting the subspace into the original model transferred a considerable fraction of the gain back. The shift operates through attention heads processing the empty think block. Scaling RL training further approaches THINK-mode accuracy but does not exceed it.
Why it matters
Many labs train hybrid reasoning models in NO-THINK mode expecting genuine efficiency gains: shorter outputs, lower inference costs. This study suggests the model is instead reasoning implicitly through its hidden states. The performance ceiling at THINK-mode accuracy implies RL is recovering existing reasoning capability rather than building new patterns. Teams measuring NO-THINK accuracy improvements may be producing hidden reasoners, with no efficiency benefit to show for the compute spent training them.