Safeguarding research integrity in autonomous research agents.
News · Overview · Quick Start · License · Citation
- 2026/10/23 — We released RIDAR!
We introduce RIDAR, a framework for Research Integrity violation Detection And Runtime intervention. Research integrity violations (RIVs) are agent behaviors that compromise the validity, reproducibility, or faithfulness of the research process. To provide a unified standard for RIV detection, we first formalize RIVT (Research Integrity Violation Taxonomy), a taxonomy of 14 RIV types across four stages of the research lifecycle: idea generation, literature review, experimentation, and paper writing. RIDAR consists of two complementary modules: RIDAR Post-Hoc detects RIVs in completed agent trajectories, while RIDAR Runtime monitors running research agents and intervenes with feedback for real-time correction.
| Dataset | Contents |
|---|---|
| RIV-GOLD | 60 human-annotated trajectories for detection evaluation |
| RIV-WILD | 300 trajectories across 50 research tasks and 6 agent configurations |
| RIV-LIVE | 60 runtime-intervention trajectories paired with controls in RIV-WILD |
Table 2: Results for different configurations of RIDAR-POST-HOC on RIV-GOLD.
RIDAR requires Python 3.10 or newer. RIDAR Post-Hoc uses Claude Code for the Analyzer and the Codex CLI for the Verifier.
git clone https://github.com/THU-KEG/RIDAR.git
cd RIDAR
conda create -n ridar python=3.11 -y
conda activate ridar
python -m pip install -e .
ridar --helpCreate a local configuration and provide the API credentials named in it:
cp config.example.toml config.toml
export RIDAR_ANALYZER_API_KEY=...
export RIDAR_VERIFIER_API_KEY=...ridar collect claude-code --path LOGS_DIR --run-id RUN_ID
ridar collect codex --session SESSION_ID --run-id RUN_ID
ridar collect opencode --session SESSION_ID --run-id RUN_IDCollected runs are written to runs/ by default. Run ridar adapters to list
all supported collection adapters.
Detect one or more collected runs by run ID. Detailed reports are written to
each run's detections/ directory:
ridar detect RUN_ID --runs runs
ridar export runs/RUN_ID/detections --output results/ridar-detection.jsonlRIDAR detects the live trajectory and returns detected violations to the agent
for correction. Everything after -- is the agent command launched by RIDAR:
ridar runtime claude-code --run-id RUN_ID --cwd WORKSPACE --intervene \
-- claude --dangerously-skip-permissions -pBrowse runs in the local web UI:
ridar monitor --port 4321python -m pip uninstall ridarUse ridar <command> --help for the complete command reference.
RIDAR is released under the MIT License.
If you use RIDAR or these datasets, please cite our paper:
@article{siow2026ridar,
title={RIDAR: Safeguarding Research Integrity in Autonomous Research Agents via Automated Detection and Runtime Intervention},
author={Jiening Siow and Amy Xin and Junjie Wang and Ziyu Tang and Zijun Yao and Yunjia Qi and Fanjin Zhang and Lei Hou and Jie Tang and Juanzi Li},
journal={arXiv preprint arXiv:2610.31372},
year={2026}
}