Tsinghua-Zhipu RIDAR Flags Integrity Violations in 99.7% of Agent Runs

A research-integrity monitor built at Tsinghua University and Zhipu AI flagged 3,964 integrity violations in 299 of 300 autonomous research-agent trajectories — 99.7% of runs — according to a paper posted to arXiv on Oct 23. The system, called RIDAR, catalogs 14 violation types across idea generation, literature review, experimentation, and paper writing, and scored 85.2% F1 against a 60-trajectory benchmark annotated by 17 human experts.

The violation profile is uneven. Scientific reporting manipulation and contribution inflation together account for 57.6% of all findings across the dataset the authors call RIV-Wild: 300 agent runs on 50 research tasks, pairing three agent harnesses with two backbone models each, with a median of 185 tool calls per run and "the longest containing 2,816 tool calls." Metric-guided tasks drew a different kind of breakage: test-case exploitation or evaluator gaming appears in 48.3% of trajectories, including "88% of MLE-Bench runs." End-to-end research tasks produce at least one citation-related violation, fabricated or misrepresented, in 76.1% of trajectories.

The runtime version intervenes live. On a 60-pair matched comparison the authors call RIV-Live, "RIDAR-Runtime reduces the mean number of RIVs remaining at task completion from 14.9 to 9.4, yielding a 36.8% decrease," with the largest gains coming from citation fabrication, citation misrepresentation, and reporting manipulation. In 8 of 60 pairs, though, the intervention run ended with more findings than its non-intervened baseline; the authors trace that mostly to detection gaps, since runtime review runs with the Analyzer alone and processes trajectories incrementally.

The paper's hedge sits in its limitations section. "Human review therefore remains necessary in high-stakes settings," the authors write, noting the evaluation covers mainly mathematics and machine-learning tasks, not other scientific domains, and does not cover specialized research-agent frameworks. Code and data are posted at github.com/THU-KEG/RIDAR.