DR Tulu-8B: Open-Source AI for Long-Form Research via Reinforcement Learning with Evolving Rubrics

Researchers have introduced Reinforcement Learning with Evolving Rubrics (RLER), a novel training method for deep research agents, and developed DR Tulu-8B, the first fully open model trained for open-ended, long-form research. Unlike prior open models that rely on short-form QA tasks with verifiable rewards, RLER constructs and maintains rubrics that co-evolve with the policy model during training, incorporating newly explored information from search and contrasting model responses for better fact-checking. DR Tulu-8B substantially outperforms existing open deep research models across four long-form benchmarks in science, healthcare, and general domains. The work is detailed in a paper on arXiv (2511.19399).