Modern LLM agents reason and interact through text, while latent reasoning offers an efficient internal medium but lacks native tool access and reasoning transparency. Hybrid Reasoning Agents (Hybrid Agents) bridge these two directions with a unified generation process. Architecturally, a lightweight Hybrid Link 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. Building on this architecture, dedicated post-training algorithms for hybrid reasoning were developed: Hybrid SFT stabilizes tool-integrated hybrid generation by teaching the backbone to coordinate with Hybrid Link, then Hybrid RL scales reasoning by learning when to think in latent space and when to act through text. Hybrid Agent 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, Hybrid Agent consistently improves performance by 10.4%, while delivering 35.5%-54.7% token reduction, 1.8x-2.6x faster inference, and 2.7x-3.7x cost savings. Code and data are available at this https URL. The research addresses the challenge of building agents that think beyond human language yet still act through it in the external environment, offering a practical solution that combines the efficiency of latent reasoning with the accessibility of text-based interaction.
The unified approach represents a significant advancement in making latent thoughts directly inspectable while maintaining tool-integrated agentic capabilities.
Source: Hacker News · Summarized by HeadlinesBriefing