knowledge-verification

The paper introduces KnownLieBench, a benchmark to evaluate emergent deception in LLM agents under conflicting incentives by verifying knowledge before assessing deceptive behavior.