Hybrid Reasoning Agents mix latent reasoning with text-based tool use

The paper starts from a gap in current agent design: modern LLM agents reason and act through text, while latent reasoning is more efficient internally but lacks native tool access and reasoning transparency. HybridAgents moves between latent reasoning and text-based interaction throughout a single unified generation process. A lightweight component called HybridLink handles transitions between text and latent modes, and a continuous working memory keeps hybrid context across multiple turns.

The authors also introduce two post-training algorithms. HybridSFT teaches the backbone model to coordinate with HybridLink for tool-integrated generation, and HybridRL learns when to think in latent space and when to act through text. According to the paper, latent thoughts remain directly inspectable.

The model was evaluated at three scales across 11 benchmarks covering mathematics, science, medicine, search and agentic reasoning. The authors report 1.8x to 2.6x faster inference and 2.7x to 3.7x cost savings compared with the systems they tested. The paper links code and data.

If the reported results hold, agent builders may be able to get tool-using reasoning with fewer tokens and lower cost per task. The numbers are the authors' own claims in a preprint, so they have not been independently verified. Check the released code against your own tool setup before expecting the same gains.