Hybrid Reasoning Agents lifts benchmarks, runs 2.1× faster

The starting point the researchers describe is that an AI agent must think out loud, token by token, before it can search, run code, or act, while humans ponder silently and speak only when it matters. Hybrid Reasoning Agents changes that by having the thinking step run in latent space, a numerical space where data is converted into a form the AI can handle. In their design, inference alternates between a text mode that handles output and tool calls and a latent mode that handles thinking. The latent-mode thinking is also formatted so people can read it, which keeps the internal step inspectable.

The method is not a drop-in change. It requires post-training a pretrained model with HybridSFT, a supervised learning method, and HybridRL, a reinforcement learning method. The reported results came from that post-training: benchmark scores improved, processing became up to 2.1× faster, and applying the method to Qwen3 4B cut cost to one 3.7th. The speed and cost figures point most directly at teams running compact models in production, though the post-training requirement suggests adoption would come as a deliberate model-building step rather than a simple switch.