Hybrid Reasoning Agents: 35–55% fewer tokens and up to 2.6× faster inference

Hybrid Reasoning Agents (HybridAgents) present a unified approach that lets an agent alternate between internal, high-efficiency latent reasoning and external, text-based interaction while retaining native tool access and transparent chain-of-thought. The system inserts a lightweight HybridLink to switch seamlessly between latent and textual modes and maintains a continuous working memory that preserves hybrid context losslessly across turns. Two post-training procedures optimize this architecture: HybridSFT trains the backbone to coordinate reliably with HybridLink for tool-integrated hybrid generation, and HybridRL teaches the agent when to perform internal latent thinking versus produce text actions, enabling adaptive division of labor between thinking and acting.

Evaluated at three model scales on 11 benchmarks spanning mathematics, science, medicine, search, and agentic tasks, HybridAgents consistently outperform frontier text-only and latent frameworks by an average of 10.4% while dramatically reducing resource use. Reported gains include 35.5%-54.7% fewer tokens, 1.8×-2.6× faster inference, and 2.7×-3.7× lower computational cost, demonstrating that combining inspectable latent thought with text-based toolability yields both accuracy and efficiency improvements. Code and data are provided for reproduction.