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

Researchers developed Hybrid Reasoning Agents that combine latent (internal) and text-based reasoning to improve LLM agent efficiency, achieving 10.4% better performance while reducing token usage by 35-55% and speeding up inference by 1.8-2.6×. The system uses a lightweight HybridLink component to seamlessly switch between internal reasoning and external text-based interactions, maintaining tool access and reasoning transparency while enabling significant cost savings of 2.7-3.7×.