The new HybridAgents framework, detailed in a preprint submitted on October 13, 2026, allows AI agents to toggle between latent and text-based reasoning modes. By performing internal reasoning without decoding thoughts into text, the system aims to reduce computational expenses.
Researchers report that the framework achieves compute cost savings ranging from 63% to 73%. Additionally, the model demonstrated an average performance improvement of 10.4% compared to existing text-based and latent LLM agents across 11 benchmarks.
The study evaluated the framework using Qwen3-Base models with 4B, 8B, and 14B parameters across domains such as math, science, and medicine. The internal reasoning process can be visualized by developers as readable pseudo-text.
The project, authored by researchers including Jiaru Zou from Stanford and Together AI, is now publicly available. The code is hosted on GitHub under the repository togethercomputer/HybridAgent, and the models are accessible via Hugging Face.
