Definition
Plain language
A training pipeline that builds AI research agents using fully synthetic tasks graded against detailed checklists.
As stated in the literature
A deep-research agent training framework from Ohio State and Amazon that uses LLM-generated rubric trees as the unifying primitive for both task synthesis and reward computation, combined with a context condenser, SFT on teacher trajectories, and GRPO with a capped fact-checking bonus.
Why it matters: Tying both the task and the grading rubric to the same structure means you can scale up research-agent training without hand-writing every problem and every rubric.
For example, the framework can synthesize a multi-hop research question along with a tree of subcriteria a good answer must satisfy, and then grade an agent's response against each leaf check.
Heard on the show
“… The paper we're working through is "QUEST: Training Frontier Deep Research Agents with Fully Synthetic Tasks," from Jian Xie and a long list …”Episode 082 — Training a Deep Research Agent on 8,000 Synthetic Tasks: The Rubric Tree Trick