HybridAgents framework reduces AI compute costs by up to 73%

A new research paper introduces HybridAgents, a framework designed to significantly reduce the computational costs of AI reasoning agents. By allowing models to toggle between a latent mode—where reasoning occurs internally without decoding thoughts into text—and a text mode for external actions, the system achieves substantial efficiency gains.

Research conducted by authors from Stanford and Together AI indicates that the framework can reduce compute costs by approximately 63% to 73%. Additionally, the system reported an average performance increase of 10.4% over existing text-based and latent LLM agents when tested on Qwen3-Base models across math, science, and medicine benchmarks.

Technical improvements include a HybridLink module for mode switching, continuous working memory, and specialized post-training methods known as HybridSFT and HybridRL. The researchers have made the work public via GitHub and Hugging Face.