A new open-source model matches near-frontier performance on complex research tasks by correcting a training data problem its authors say most labs have overlooked.
Researchers released S1-DeepResearch, a framework for training AI agents to handle the full arc of research work — not just retrieving facts, but planning an inquiry, integrating evidence across sources, and writing structured reports. The core claim is that existing training datasets lean heavily on closed-ended question answering, which teaches a model to look things up but not to reason through them. Their alternative builds what the authors call "agentic trajectories" from a blend of closed Q&A and open-ended exploration, then validates those trajectories across multiple quality dimensions. The result, S1-DeepResearch-32B, achieves top or near-top scores across 20 benchmarks covering complex reasoning, instruction following, report generation, file handling, and tool use.
"Deep research" has become one of AI's more aggressively marketed terms, with major labs attaching it to products that still function mostly as sophisticated search. This paper makes a concrete methodological argument for why those products hit a ceiling: if you train on search data, you get a search agent. The capability gap the researchers identify — synthesis, planning, structured output — is precisely what users complain about when AI-generated reports feel shallow and disconnected.
The model is open-source, which at least means the methodology can be independently reproduced and stress-tested — a bar closed labs rarely have to clear.