Tool-use Agent Robustness on τ-bench
6.9Behavioral Uncertainty (BU)PI Detector
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| PI DetectorBackbone LLM=GPT-4o2025.10 | 6.9 | 5.65 | 0 | |
| NoneBackbone LLM=GPT-4o2025.10 | 51.73 | 47.4 | 56.09 | |
| SpotlightingBackbone LLM=GPT-4o2025.10 | 51.74 | 46.74 | 52.6 | |
| Repeat promptBackbone LLM=GPT-4o2025.10 | 52.17 | 46.09 | 52.67 | |
| SanitizerBackbone LLM=GPT-4o2025.10 | 59.09 | 63.91 | 0 |