Dr. Bench: A Multidimensional Evaluation for Deep Research Agents, from Answers to Reports

Dr. Bench addresses a critical gap in the 'Deep Research' agent space: evaluating long-form, multi-step reasoning and structured report generation rather than just short-form answers. However, its defensibility is currently low (3/10) due to its status as an academic artifact with zero GitHub stars and limited community traction beyond initial forks. While the multi-dimensional scoring mechanism is a novel combination of existing evaluation techniques, benchmarks in the AI space only gain a moat through massive adoption or unique, human-verified datasets. Currently, it faces high frontier risk as OpenAI (with OpenAI Deep Research) and Google are defining the capabilities of these agents and will likely release their own proprietary or sponsored benchmarks (e.g., following the pattern of SWE-bench or GPQA). Competitors include established benchmarks like GAIA, Tau-bench, or industry-standard evaluation frameworks like LangSmith. Without a leaderboard or a significant push for adoption by frontier labs, this project remains a reference implementation likely to be displaced by 2025 as agent evaluation standards consolidate.