Stanford and Together AI's HybridAgent reasons in latent space and scores higher on GAIA

Stanford and Together AI researchers introduce HybridAgent, an agent that reasons in hidden vectors and switches to text only for tool calls and key decisions.

A small adapter feeds the model's internal state back as its next input, so it can think without writing words. The paper reports 2.1x faster inference than matched text agents and cuts decoding compute by 62.5% to 72.9%.

On the GAIA benchmark, the 14B version scored 52.4, compared with 43.7 for the same model trained as a text-only agent with ARPO. The figures come from the authors' paper, which the Rohan Paul post on X reports; they have not been independently verified.