Evaluations & experimentsGoogle DeepMind
Testing scheming with realistic coding tasks
DeepMind built evaluations around coding tasks in its alignment codebases. In that setting, Gemini did not exhibit unprompted scheming.
- Published
- Added to the archive
- Source checked on
- Original title
- Realistic honeypot evaluations for scheming propensity
Details
Prompts explicitly encouraging agency (situational awareness or goal-directedness) or assigning hidden goals elicited some such behavior.