Hybrid Reasoning Agents combine latent thinking with text-based action

Together Computer has introduced Hybrid Reasoning Agents, a new architecture that lets language models think in compressed latent space while acting through text. The approach uses a lightweight component called HybridLink to transition between internal reasoning and external text-based interaction, preserving tool integration and reasoning transparency.

The team developed dedicated training methods: HybridSFT teaches models to coordinate with HybridLink, and HybridRL learns when to reason in latent space versus act through text. Testing across 11 benchmarks in math, science, medicine, search and reasoning tasks, HybridAgent improved performance by 10.4% over frontier LLM agents while cutting token use by 35.5% to 54.7%, achieving 1.8x to 2.6x faster inference and 2.7x to 3.7x cost savings.

Code and models are open-source on GitHub and Hugging Face.