AI & Agents · via Hacker News
Hybrid Reasoning Agents think in latent space and act through text
Researchers from Stanford, UIUC and Together AI trained Qwen3-based agents that switch between hidden 'latent' thinking and normal text for tool calls. They report about 10% better scores across 11 benchmarks, with 35–55% fewer tokens and roughly 2x faster inference, while the latent steps can still be read back as text.