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

RIDAR. Safeguarding Research Integrity in Autonomous Research Agents via Automated Detection and Runtime Intervention. By Jiening Siow and nine others, from Tsinghua University, Renmin University of China, and one other institution. Published on the twenty-second of October. Its code repository is worked out: about fifty source files, with a readme and a licence.

Autonomous research agents are being used to automate large portions of scientific work. The trouble is that they are prone to what the authors call research integrity violations: behaviours that compromise the validity, the reproducibility, or the faithfulness of the research process. An agent can reach the end of a task while the route it took has quietly undermined the result.

RIDAR is a framework for catching those violations automatically and for intervening while the agent is still running. Its foundation is a taxonomy called RIVT, which lays out fourteen types of violation spanning the full research lifecycle. Everything else is built on that vocabulary.

There are two detectors. The first, RIDAR-post-hoc, reads an agent's trajectory after the fact and finds violations within it. The second, RIDAR-runtime, works during execution, dynamically detecting violations and correcting them before the task is finished. To measure detection, the authors built RIV-Gold, a benchmark of sixty trajectories annotated by experts.

On RIV-Gold, the post-hoc detector reaches eighty-five point two percent F1. The authors then pointed it at RIV-Wild, a dataset of three hundred agent trajectories drawn from fifty distinct research tasks. It found three thousand nine hundred and sixty-four violation instances.

Those instances turned up in ninety-nine point seven percent of the trajectories, which is the basis for the claim that violations are prevalent in current agents. The runtime framework then reduces the violations still present at task completion by thirty-six point eight percent.

Code and data are to be released. The arc of the work runs from naming the problem, to measuring how often it happens in the wild, to pushing the number down while the agent is still working, and the survey of the wild set suggests the problem is close to universal.