Hybrid Reasoning Agents

Jiaru Zou, Yunzhe Qi, Sirui Chen, Federico Bianchi, Shang Zhu, Pan Lu et al.

arXiv 2610.18861 45 pages Read the paper

The short version

The problem: Today’s AI agents can explain their thinking in words and call tools, but producing every thought as text is slow and costly. Systems that think in compact internal representations can be faster, yet they struggle to interact with tools over multiple turns and are harder to inspect.
The idea: HybridAgent combines the two: it can think in an internal continuous representation, then switch to ordinary text when it needs to call a tool or show an important step. For example, it can do routine calculation internally, use text to run a Python tool, and retain both the internal thinking and tool results for its next turn. It also uses dedicated training methods to learn these transitions and to practice reasoning with its own internal representations.
What it got: Across 11 tests, it improved performance by an average of 10.4% over the compared methods. It also used 35.5%–54.7% fewer tokens and ran 1.8×–2.6× faster.

Modules from this paper

HybridLink: Bridging Latent Reasoning and Text

Hybrid agents need to move between continuous internal steps and readable text; HybridLink maps hidden states back into the model’s input-embedding space while learned anchors mark entry and exit.

hybrid reasoninglatent reasoninglanguage modelsagents

Persistent Working Memory for Multi-Turn Hybrid Context

Hybrid turns mix text, latent thoughts, and tool observations that cannot be reconstructed from text alone; the agent preserves their accumulated key-value context and resumes from it across turns.

working-memoryhybrid-agentsmulti-turn-generation

Policy Learning over Hybrid Trajectories

Hybrid trajectories contain continuous latent steps and environment observations with no token likelihood, so HybridRL factors policy probability over sampled text positions and replays the recorded context to update them.

reinforcement-learninghybrid-agentspolicy-optimizationtrajectory-replay

Self-Conditioned HybridSFT

Hybrid latent training can fail when the model only sees target embeddings; HybridSFT feeds its own predicted latent inputs back during a second training pass.

hybrid-reasoningsupervised-fine-tuninglatent-generation