ResearchArena: Evaluating Sabotage and Monitoring in Automated AI R&D
Published in arXiv preprint, 2026
We introduce ResearchArena, a framework spanning four long-horizon AI R&D tasks (safety post-training, capabilities post-training, CUDA-kernel optimization and inference-server optimization) that evaluates whether frontier agents can covertly sabotage the artifacts they produce, and whether monitors can catch them before deployment. Sabotage hidden in the training data is the hardest to catch.
Recommended citation: Libon, L., Rank, B., Yeon, J., Schmotz, D., Qin, J., Donnelly, D., Prinzhorn, D. & Andriushchenko, M. (2026). "ResearchArena: Evaluating Sabotage and Monitoring in Automated AI R&D." arXiv preprint arXiv:2607.19321.
Download Paper