Tool-Risk Prediction on Nemotron core (held-out test)
90.3Tool-Risk Accuracy (tool rows)GPT-OSS 20B
Evaluation results
| Method | Links | ||
|---|---|---|---|
| GPT-OSS 20BInstruction-tuned (IT)=true, Dominant SAE layers=23, 19, Number of SAE features=1000, Probe regularization=elastic net2026.05 | 90.3 | 64 | |
| Gemma 3 27BInstruction-tuned (IT)=true, Dominant SAE layers=40, 53, 31, Number of SAE features=1000, Probe regularization=elastic net2026.05 | 88.5 | 62 |