Researchers introduce 3D-VCD, a new technique to reduce false predictions in AI agents operating in 3D environments by contrasting predictions against distorted scene variations.

3D-VCD is the first inference-time visual contrastive decoding framework designed specifically to combat hallucinations in 3D embodied AI agents

The method creates distorted 3D scene graphs using semantic and geometric perturbations like object category substitutions and coordinate corruption

By comparing predictions between original and distorted 3D contexts, the approach suppresses unreliable tokens and grounds decisions in actual spatial and geometric information

Existing 2D hallucination mitigation methods fail for 3D embodied reasoning because errors stem from object presence and spatial layout rather than pixel-level issues