Tool-call Monitoring on BFCL (Berkeley Function Calling Leaderboard) zero-shot transfer
87.6AgreementGemma 3 27B
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Gemma 3 27BMonitoring Probes=Probe 1 & Probe 2, Transfer Mode=zero-shot, Training Setup=Nemotron only, no retraining or threshold tuning2026.05 | 87.6 | 0.2 | 98.5 | 50 | 1.17 | 77.7 | |
| GPT-OSS 20BMonitoring Probes=Probe 1 & Probe 2, Transfer Mode=zero-shot, Training Setup=Nemotron only, no retraining or threshold tuning2026.05 | 83.6 | 10.8 | 56.3 | 49.2 | 1.69 | 87.7 |