HybridAgents is an AI agent design that reasons in a hidden, non-linguistic latent space and switches to plain text only when it needs tools or environment interaction. The authors report lower token use, faster inference and 10.4 percent better performance, based on their own unverified evaluation.
Researchers have proposed a new kind of AI agent that reasons in a hidden, non-linguistic form and switches to plain text only when it needs to use tools or interact with its environment.
The approach, called Hybrid Reasoning Agents, or HybridAgents, is described in a paper posted to the arXiv preprint server on October 13. The authors, including Jiaru Zou and James Zou, say most large language model agents still think and act entirely through text. That works, but it is slow and expensive. Latent reasoning, which happens inside the model's internal representations, is more efficient. It has lacked native access to tools, however, and offers little visibility into how the model reached its answer.
How does HybridAgents switch between latent reasoning and text?
HybridAgents tries to combine the two. The system moves between latent reasoning and text-based interaction within a single generation process. A lightweight component called HybridLink handles those transitions, and the agent keeps a continuous working memory that carries context across multiple turns. According to the authors, this design preserves tool-integrated capabilities and makes the model's latent thoughts directly inspectable.
How was HybridAgents trained?
Training happens in two stages. The first, HybridSFT, teaches the underlying model to coordinate with HybridLink so that tool-integrated hybrid generation stays stable. The second, HybridRL, teaches the agent when to think in latent space and when to act through text.
How well did HybridAgents perform in tests?
The team tested the system at three model sizes across 11 benchmarks covering mathematics, science, medicine, search and agentic reasoning. They compared it with frontier LLM agents and with agentic frameworks that use either text-based or latent methods. HybridAgents, the authors write, consistently improved performance by 10.4 percent. Token use fell by 35.5 to 54.7 percent, inference ran 1.8 to 2.6 times faster, and cost savings ranged from 2.7 to 3.7 times.
Those figures come from the authors' own evaluation and have not been independently verified. The paper is a preprint, so it has not gone through the formal peer review that accompanies many journal publications. The authors say code and data are available online.
Este artículo fue escrito con la asistencia de IA.
News Factory APP - noticias agénticas para impulsar tu SEO y AEO.