ACT3D: RL with Outcome Rewards Teaches VLMs to Actively Explore 3D Worlds

UCSD and Georgia Tech introduce ACT3D, which trains vision-language models to actively think and explore in 3D using outcome-reward reinforcement learning, achieving an average accuracy of 53.2% across six benchmarks.

2026-10-19 ~ 2026-10-20 · 2 related posts

• ACT3D trains VLMs to actively explore 3D worlds with RL, hitting 53.2% accuracy — UCSanDiego · 2026-10-19
• Act3D trains VLMs to explore 3D with outcome-only RL, hitting 53.2% across six benchmarks — _akhaliq · 2026-10-20