Hybrid Reasoning Agents

Hybrid Reasoning Agents. By Jiaru Zou and seven others, from Stanford University, Together AI, and two other institutions. Published on the thirteenth of October. Its code repository is worked out: about forty source files, with a readme and a licence.

The paper opens with a question. How can an agent think beyond human language, yet still act through it in the external environment? Two directions have each answered half of that. Modern LLM agents reason and interact through text, which keeps them able to call tools and keeps their reasoning readable. Latent reasoning offers a more efficient internal medium, but it lacks native tool access, and it lacks reasoning transparency. One side can act and be inspected. The other thinks more efficiently and can do neither.

Hybrid Reasoning Agents, or HybridAgents, try to keep both. The agent dynamically navigates between latent reasoning and text-based interaction, and it does so throughout a single, unified generation process rather than in separate systems stitched together. Thinking can happen in latent space; acting in the world happens through text; and the agent moves between the two as it goes.

Two architectural pieces make that possible. A lightweight HybridLink enables seamless transitions between text and latent modes. The agent's continuous working memory preserves lossless hybrid context across multiple turns, so what was thought in one mode is not lost when the agent switches to the other. According to the authors, these designs preserve tool-integrated agentic capabilities while making latent thoughts directly inspectable.

On top of the architecture sit two dedicated post-training algorithms. HybridSFT comes first. It stabilizes tool-integrated hybrid generation by teaching the backbone to coordinate with HybridLink. HybridRL follows, and scales reasoning by learning when to think in latent space and when to act through text. The first stage teaches the model to use the bridge at all; the second teaches it when to cross.

HybridAgent is evaluated at three model scales across eleven benchmarks, covering mathematics, science, medicine, search, and agentic reasoning. The comparison set is frontier LLM agents and agentic frameworks, spanning both text-based and latent methods. Against those, HybridAgent consistently improves performance by ten point four percent.

The efficiency figures come alongside. Token use falls by between thirty-five point five and fifty-four point seven percent. Inference runs between one point eight and two point six times faster. Cost savings range from two point seven to three point seven times. Code and data are released.

The framing is a refusal to choose. Text was the medium for acting and being read; latent space was the medium for thinking cheaply. HybridAgents treat them as two modes of one agent, with a link between them and a memory that carries across, and then train the agent to decide for itself which mode each moment calls for.