HybridAgent

Together researchers propose Hybrid Reasoning Agents

A paper introduces HybridAgents, which switch between latent-space reasoning and text interaction via a lightweight HybridLink and keep hybrid context in continuous working memory, trained with HybridSFT and HybridRL. Across 3 model scales and 11 benchmarks it improves performance by 10.4% over frontier LLM agents while delivering 35.5%-54.7% token savings; code is open-sourced.