Sources
MineAmongUs puts deception research in a 3D body, and separates what the model does from what the harness lets it do
arXiv 2608.30428 (submitted 2026-08-31) points out that social-deduction testbeds for agent deception are text-only and run on one fixed agent configuration, so an observed behavior cannot be attributed to the model rather than the surrounding scaffold. MineAmongUs is a 3D multimodal Among Us sandbox where imposter agents must deceive through joint verbal and non-verbal action, paired with ARIA, a configurable VLM-agent harness exposing five cognitive-component ablation axes. The harness-versus-model confound they name applies well beyond deception research.
↳ Follow the thread