Stanford, Meta team unveils Hybrid Reasoning Agents to cut AI token use
The technique computes intermediate steps in latent space, cutting generated tokens by 35.5-54.7% and reporting 1.8-2.6x faster reasoning.
A joint research team from Stanford University, Meta, Together AI and the University of Illinois Urbana-Champaign announced Hybrid Reasoning Agents, an AI reasoning technique that handles intermediate reasoning in a latent space of numeric vectors rather than writing out every step as text tokens, according to digitaltoday.co.kr.
The method pairs a latent mode for internal problem solving with a text mode for tool calls and final answers. A technique called HybridLink connects the two, and applying the approach requires adding a supervised fine-tuning method, HybridSFT, and a reinforcement learning method, HybridRL, to pretrained models.
The team tested 4B, 8B and 14B parameter models based on Qwen3 across 11 benchmarks covering math, science and search. Generated tokens fell by 35.5-54.7% and reasoning speed rose 1.8-2.6 times compared with existing methods.
Problem-solving performance improved by an average of 10.4%, and the estimated compute cost of the 4B model dropped to about 27% of the conventional method. The team released the code and training data on GitHub and Hugging Face.
The results appear in a paper that has not yet undergone peer review.
Quick answers
What is Hybrid Reasoning Agents?
It is an AI reasoning technique from Stanford University, Meta, Together AI and the University of Illinois Urbana-Champaign that performs intermediate reasoning in latent space instead of generating every step as text tokens.
How much faster is the method?
In tests on 4B, 8B and 14B parameter Qwen3-based models across 11 benchmarks, generated tokens fell by 35.5-54.7% and reasoning speed increased 1.8-2.6 times versus existing methods.
Is the code available?
Yes. The team released the code and training data on GitHub and Hugging Face. The paper has not yet been peer reviewed.