ThinkRouter introduces confidence-aware routing between latent and discrete spaces for efficient AI reasoning. It switches to discrete tokens during low-confidence steps to reduce noise from latent embeddings. Experiments show major accuracy gains on STEM and coding tasks while shortening outputs.
Key Points
- 1.19.70 point average Pass@1 improvement
- 2.Up to 15.55% shorter generation length
- 3.Outperforms CoT and latent reasoning baselines
Impact Analysis
Enhances large reasoning models' performance on complex tasks. Calibrates errors from traditional methods. Accelerates practical deployment by improving speed and accuracy.
Technical Details
Routes to discrete space on low confidence, latent otherwise. Addresses noise in soft embeddings from low-confidence alternatives. Analyzes confidence dynamics in reasoning trajectories.