Researchers propose AI agent that reasons in latent space and acts through text

A research team has published a paper on arXiv, the open scientific preprint repository, describing a new type of large language model agent called Hybrid Reasoning Agents, or HybridAgents. These AI agents do their reasoning inside the model's internal, non-verbal (latent) space, while interacting with the external environment, such as tools, through text. The paper was submitted on 13 October 2026 in the computation and language (cs.CL) category.

The problem being addressed

Current language-model agents both think and act in text. Reasoning in latent space is more efficient, but it lacks direct tool access and makes the reasoning process harder to inspect. The authors aim to bring the two approaches together.

How the system works

Within a single generation process, the system switches dynamically between latent reasoning and text interaction. A lightweight component called HybridLink handles the transitions. The agent's continuous working memory keeps the hybrid context intact across multiple turns. According to the authors, this retains tool-using capabilities while allowing the latent thoughts to be inspected directly.

Two training algorithms were developed. HybridSFT stabilizes tool-integrated hybrid generation by teaching the model to coordinate with HybridLink. HybridRL, based on reinforcement learning, teaches the agent when to think in latent space and when to act through text.

Results

HybridAgent was evaluated at three model scales on 11 benchmarks covering mathematics, science, medicine, search and agentic reasoning. Compared with leading text-based and latent agents and frameworks, the authors report an average performance improvement of 10.4%. At the same time, it uses 35.5%–54.7% fewer tokens, runs inference 1.8 to 2.6 times faster and cuts costs by a factor of 2.7 to 3.7.

The paper runs to 45 pages, and the authors state that the code and data are publicly available.