DEEPRUBRIC: Evidence-Tree Rubric Supervision for Efficient Reinforcement Learning of Deep Research Agents

A new data-construction framework called DeepRubric trains deep research agents more efficiently by reversing how evaluation rubrics are generated, its creators report. The method builds an evidence tree from a seed topic and synthesizes aligned query–rubric pairs, cutting the reinforcement-learning compute needed to match prior open models by roughly 13 times [1][2]. Deep research agents are large language models that produce long-form reports by searching and reasoning over retrieved evidence [1][2]. Reinforcement learning with rubric-based rewards has been used to improve these agents by optimizing them against checkable criteria that translate report quality into a reward signal [2]. The efficiency of that approach depends on whether the rubrics reliably capture the task scope and the evidence the agent needs to gather [2]. Most existing work asks an LLM to generate rubrics for a given query, but when the model fails to infer the underlying information needs, the resulting rubrics can be incomplete and reduce RL efficiency [2]. The DeepRubric framework, described in a paper posted to the arXiv preprint server on 15 June 2026, reverses that workflow [1][2]. Instead of inferring evaluation criteria for a given query, it first determines what an evidence-backed report should be evaluated on and then synthesizes aligned query–rubric pairs from those evaluation targets [2]. Starting from a sampled seed topic, DeepRubric builds an evidence tree by recursively expanding evidence-backed sub-questions; the leaves of that tree serve as atomic, verifiable evaluation targets [2]. The tree is then used to synthesize the training query and its rubrics, so the reward evaluates exactly the information the query requests [2]. The authors used the framework to construct 9,000 query–rubric supervision examples and trained an 8-billion-parameter model, DeepRubric-8B, with rubric-based Group Relative Policy Optimization [2]. Across three benchmarks, the model achieved performance comparable to prior open state-of-the-art deep research models while consuming roughly 13 times fewer RL GPU-hours [1][2]. arXiv, where the paper appeared, is an open-access repository of electronic preprints that has hosted scientific papers since August 1991 and now receives about 24,000 submissions per month [6]. The repository’s arXivLabs program, launched in 2020, provides a formal framework for community-developed tools that appear on article pages, such as citation explorers and recommender systems [4][5].