1 min read

Controlling Deception using Representation Engineering

AI SafetyRepEDeception DetectionLLMs

Explored ways to detect and mitigate deceptive behavior in large language models using Representation Engineering (RepE) as part of the Apart Lab Fellowship.

This work analyzed activation-space directions associated with deceptive completions, examining how RepE-based interventions could reduce deceptive outputs without degrading general model capability.

Back to top