Hybrid agents: thinking in latent space, acting through text

Modern LLM agents reason and interact with their environment through text. That is convenient for transparency and tool access, but costly in tokens and compute. The alternative is reasoning in a "latent" space: it is a more efficient internal medium, but it lacks direct tool access and is opaque to any observer.

A team of eight authors led by Jiaru Zou has proposed a way to combine both. The paper "Hybrid Reasoning Agents" appeared on arXiv on 13 October 2026.

How it works

The proposed Hybrid Reasoning Agents dynamically move between latent reasoning and text-based interaction within a single, unified generation process. A lightweight component called HybridLink handles the seamless transitions between the two modes, while the agent's continuous working memory preserves context losslessly across multiple turns.

By design, this keeps the agent's tool-integrated abilities intact while making its latent "thoughts" directly inspectable — addressing the main objection to latent reasoning, namely its opacity.

Two dedicated training algorithms back the architecture. HybridSFT stabilizes tool-integrated hybrid generation by teaching the backbone model to coordinate with HybridLink. HybridRL then scales reasoning by learning when to think in latent space and when to act through text.

What the numbers show

The authors evaluated the approach at three model scales across eleven benchmarks spanning mathematics, science, medicine, search, and agentic reasoning. They compared it against frontier LLM agents and frameworks, both text-based and latent.

By their account, HybridAgent consistently improves performance by 10.4% while cutting token use by 35.5–54.7%, running inference 1.8–2.6 times faster, and lowering cost by a factor of 2.7–3.7. Code and data have been released publicly.

The 45-page preprint with 12 figures has not yet been peer-reviewed, so the reported figures should be read as the authors' own results rather than independently confirmed ones.

Source: Hacker News