Researchers find Qwen3’s reasoning gains may stem from ‘thinking leakage’
A new paper on arXiv examines how reinforcement learning (RL) training affects hybrid reasoning models like Qwen3. The study suggests that much of the reported improvement in reasoning performance—when models switch between THINK and NO-THINK modes—actually comes from thinking leakage: the model partially reactivates its original THINK mode rather than learning new reasoning behavior.
Key points
- Qwen3’s reasoning gains in NO-THINK mode may reflect reactivation of original THINK mode, not new learning
- RL-trained models’ responses grow longer and more reflective, mirroring THINK mode behavior
- Swapping activation subspaces shows original model absorbs most RL gains when given THINK-related data
The authors trained Qwen3 variants (1.7B, 4B, and 8B parameters) using GRPO in NO-THINK mode. As accuracy rose, responses grew longer and more reflective, resembling the original model’s THINK mode. Weight and activation analysis showed that further RL scaling merely shifted behavior toward THINK mode without exceeding its baseline accuracy. Swapping activation subspaces between RL-trained and original models confirmed this: the original model absorbed most of the RL gain when given its THINK-related activations.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Asset learns to compress speech encoders · 1 src
- Researchers test AI agents improving LLM training strategies with up to 13.34% gains · 1 src
- Researchers model how AI and humans time new words in text · 1 src
- Researchers propose Next-Concept Modeling for faster text generation · 1 src
- Researchers test Typed-Decision models on Indonesian regional languages · 1 src
Comments
via GitHub Discussions