The MineAmongUs sandbox and ARIA harness enable testing of strategic deception in VLM agents through joint verbal and non-verbal social interactions. Unlike text-only testbeds, this 3D environment captures sensorimotor deception, such as physical movement, in multi-agent settings.
HOW THIS AFFECTS YOU
●
researcherYou can now evaluate how non-verbal actions contribute to agentic deception in embodied environments.
●
policyThis provides more realistic benchmarks for evaluating the alignment and safety of socially interacting AI agents.