Hybrid Reasoning Agents Bridge Latent Reasoning and Text-Based Agent Interactions

Architecture: HybridLink and Working Memory

A new research paper titled "Hybrid Reasoning Agents" introduces an approach that enables LLM agents to dynamically navigate between latent reasoning and text-based interaction throughout a unified generation process. The work, submitted to arXiv on 13 October 2026, is authored by Jiaru Zou, Yunzhe Qi, Sirui Chen, Federico Bianchi, Shang Zhu, Pan Lu, Jingrui He, and James Zou.

The system addresses a key challenge in modern LLM agents: these agents still reason and interact through text, while latent reasoning offers an efficient internal medium but lacks native tool access and reasoning transparency. The proposed Hybrid Reasoning Agents (HybridAgents) bridge these two directions. Architecturally, a lightweight component called HybridLink enables seamless transitions between text and latent modes, while the agent's continuous working memory preserves lossless hybrid context across multiple turns. These designs preserve tool-integrated agentic capabilities while making latent thoughts directly inspectable.

Post-Training: HybridSFT and HybridRL

The researchers developed dedicated post-training algorithms for hybrid reasoning. 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.

Evaluation and Results

HybridAgent was evaluated at 3 model scales across 11 benchmarks covering mathematics, science, medicine, search, and agentic reasoning. Compared with frontier LLM agents and agentic frameworks spanning both text-based and latent methods, the system consistently improved performance by 10.4%. It also delivered 35.5% to 54.7% token reduction, 1.8x to 2.6x faster inference, and 2.7x to 3.7x cost savings.

The paper is 45 pages long and includes 12 figures. Code and data are made available online. The work is categorized under Computation and Language (cs.CL), Artificial Intelligence (cs.AI), and Machine Learning (cs.LG).

Sources: arxiv.org