Gespeichert in:
| Hauptverfasser: | , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | https://arxiv.org/abs/2605.29729 |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866913170079088640 |
|---|---|
| author | Krakovna, Victoria Lindner, David Ho, Lewis Farquhar, Sebastian Shah, Rohin |
| author_facet | Krakovna, Victoria Lindner, David Ho, Lewis Farquhar, Sebastian Shah, Rohin |
| contents | We introduce scheming honeypot evaluations, a framework for testing whether models will pursue instrumental goals if given the opportunity. Our scheming honeypot evaluations take the form of coding tasks in Google's alignment research codebases. In a real internal deployment setting, Gemini models do not demonstrate unprompted scheming. If prompts explicitly encourage agency (situational awareness or goal-directedness) and/or give the model a hidden goal, models sometimes scheme or attempt sabotage. Validating the realism of our setting, models show low rates of evaluation awareness, usually due to agency prompts rather than the environments. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_29729 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Realistic honeypot evaluations for scheming propensity Krakovna, Victoria Lindner, David Ho, Lewis Farquhar, Sebastian Shah, Rohin Machine Learning We introduce scheming honeypot evaluations, a framework for testing whether models will pursue instrumental goals if given the opportunity. Our scheming honeypot evaluations take the form of coding tasks in Google's alignment research codebases. In a real internal deployment setting, Gemini models do not demonstrate unprompted scheming. If prompts explicitly encourage agency (situational awareness or goal-directedness) and/or give the model a hidden goal, models sometimes scheme or attempt sabotage. Validating the realism of our setting, models show low rates of evaluation awareness, usually due to agency prompts rather than the environments. |
| title | Realistic honeypot evaluations for scheming propensity |
| topic | Machine Learning |
| url | https://arxiv.org/abs/2605.29729 |