Hybrid Reasoning Agents use hidden thought for tool-based work

The paper describes agents that do most of their reasoning in a hidden internal form, using ordinary text when calling tools or giving a final answer. Its authors include researchers from Stanford, Together AI, UIUC, and Meta; the Reddit post links to the paper on arXiv.