Drift-Bench¶
"Drift-Bench: Cooperative Breakdowns in Conversational AI." arXiv:2602.02455, 2026.
Key findings used in wiki¶
- Introduces a benchmark specifically designed to measure cooperative breakdowns in multi-turn dialogue
- Demonstrates that conversational AI systems fail to maintain cooperative alignment over extended interactions
- Breakdowns occur even in non-adversarial settings where users are cooperative
- Complements PBSuite by showing that failure is not limited to adversarial scenarios
- Provides evaluation methodology that influenced InvisibleBench's cooperative breakdown scoring