Behavior Error Detection on Personal-VCL-Bench Behavior
66.22AccuracyAgentic Context Bank
Evaluation Results
| Method | Links | |
|---|---|---|
| Agentic Context BankContext regime=Ours, Backbone=Gemma-4-31B2026.05 | 66.22 | |
| Visual-ctx (kmax)Context regime=Visual-ctx (kmax), Backbone=Gemma-4-31B2026.05 | 59.46 | |
| Language-ctx (kmax)Context regime=Language-ctx (kmax), Backbone=Gemma-4-31B2026.05 | 58.11 | |
| No-contextContext regime=No-context, Backbone=Gemma-4-31B2026.05 | 52.43 |