Research shows that many performance gains in 'NO-THINK' mode for hybrid reasoning models like Qwen3 are due to 'thinking leakage.' RL training in NO-THINK mode partially reactivates the model's original THINK mode activations rather than teaching new, enhanced reasoning behaviors.
HOW THIS AFFECTS YOU
●
researcherYou should be cautious when interpreting RL performance gains in reasoning models, as they may stem from internal mode reactivation.