mlbio-epfl/HeurekaBench

HeurekaBench originates from EPFL (mlbio-epfl), a top-tier research institution, which provides immediate academic credibility but currently lacks commercial or community momentum, evidenced by its 11 stars and low fork count. It addresses a very specific and difficult niche: evaluating AI agents in 'wet lab' or data-driven scientific contexts rather than just general reasoning or coding. Its primary moat is the specific datasets and evaluation protocols it defines for ICLR 2026. However, it faces stiff competition from well-funded entities like Future House (Lab-Bench) and the AI-for-Science divisions of Google DeepMind and Microsoft Research, who are building similar internal and external benchmarks. The project's defensibility is low because the code itself is a standard benchmarking wrapper; the value lies entirely in the adoption of its metrics by other researchers. If a major lab releases a more comprehensive 'Scientist Benchmark,' this project risks becoming an obscure academic footnote. Its current velocity is zero, suggesting it is a static release tied to a paper submission rather than a living software project.