'Thinking leakage': no-think RL gains in hybrid models just reuse think-mode skills
kaust-generative-ai · hf · 2026-10-15
KAUST and SJTU researchers show that RL gains reported for hybrid reasoning models' NO-THINK mode (e.g. GRPO on Qwen3-1.7B/4B/8B) largely stem from 'thinking leakage' — reactivating existing THINK-mode reasoning rather than learning new capabilities. Extended RL approaches but never exceeds THINK-mode accuracy, and swapping a low-dimensional THINK-related activation subspace removes most of the gain.
More from Models
- 220-question test: 70% of model web searches were unnecessary, cost 75x more tokens — demian_ai · 2026-10-15
- NVIDIA releases open Nemotron 3.5 Super VL: 120B total params, 12B active for vision agents — nvidia · 2026-10-15
- Gemini team's new 3D puzzle agent becomes only the 4th with a non-trivial MazeBench run — patience_cave · 2026-10-15
- MazeBench: Argon one-shots puzzles that stumped GPT-6 Astra, but walks away near the finish — patience_cave · 2026-10-15
- MazeBench: Argon explores more thoroughly while GPT-6 Astra is stronger but sporadic — patience_cave · 2026-10-15
- MazeBench heatmaps: Argon's maze coverage beats Opus 5.5, rivals GPT-6 Astra — patience_cave · 2026-10-15