HybridAgent preprint reports lower token use and faster inference

Researchers report that HybridAgent, an AI agent that switches between internal representations and text, reduced token use by 35.5%-54.7% and delivered 1.8x-2.6x faster inference in benchmark evaluations. The preprint also reports a 10.4% performance improvement and 2.7x-3.7x cost savings against the agents and frameworks tested. A lightweight component called HybridLink lets the system switch reasoning modes while retaining context across turns. Dedicated training teaches it when to reason internally and when to use text to interact with tools. The evaluation covers 3 model scales and 11 benchmarks spanning mathematics, science, medicine, search and agentic reasoning; the reported gains apply to those comparisons.