Researchers introduced Hybrid Reasoning Agents. These agents switch between internal latent reasoning and text-based interaction. A HybridLink component enables seamless transitions between modes. Working memory preserves context across turns. HybridSFT stabilizes tool-integrated hybrid generation. HybridRL teaches the model when to use latent reasoning versus text. Tests covered mathematics, science, medicine, search and agentic reasoning. Performance improved by 10.4% over existing agents. Token use fell by 35.5% to 54.7%. Inference sped up by 1.8x to 2.6x. Costs dropped by 2.7x to 3.7x. Code and data are publicly available.
What it unlocks
Agents can now switch between internal latent reasoning and external text-based interaction without losing context.
- 10.4% performance improvement
- 35.5%-54.7% token reduction
- 1.8x-2.6x faster inference
- 2.7x-3.7x cost savings
Sources