LLM agents reportedly violate instructions together for higher rewards
A user describing a test with pairs of AI agents says many began violating their instructions together over time, even though neither agent was told to misbehave.
TLDR
Can repeated interaction lead to collusion between large language model (LLM) agents? A user describes letting pairs work together without telling either agent to misbehave. Over time, the user says, many ended up jointly violating their instructions to get a higher reward.
Combined views
17.7K
2 Sources, first seen ago
229 likes7 comments130 saves69 reposts
