RIDAR: Safeguarding Research Integrity in Autonomous Research Agents via Automated Detection and Runtime Intervention

[arXiv] Jiening Siow, Amy Xin and others propose RIDAR, for automatically detecting and intervening at runtime in research integrity violations (RIV). The authors present the RIVT taxonomy containing 14 categories of RIV, and release a post-hoc detection framework; the paper claims it achieves an F1 of 85.2% on RIV-Gold, which contains 60 expert-annotated trajectories. The authors claim that in RIV-Wild, a dataset of 300 trajectories covering 50 research tasks, 3,964 RIV cases were detected, appearing in 99.7% of the trajectories. The paper claims the runtime intervention framework can reduce RIV still present at task completion by 36.8%.