Post-shift accepted-trajectory risk estimation on Live agent RAG tool benchmark
5.5Post-shift Accepted Trajectory RiskToolChain-CRC
Evaluation Results
| Method | Links | |
|---|---|---|
| ToolChain-CRCCalibration Strategy=Trajectory calibration2026.06 | 5.5 | |
| Prompt/retrieval-onlyCalibration Strategy=Prompt/retrieval-only2026.06 | 14.5 | |
| Fixed trajectoryCalibration Strategy=Fixed trajectory2026.06 | 22.9 | |
| Final-onlyCalibration Strategy=Final-answer-only2026.06 | 35.5 |