Report
Act3D reportedly averages 53.2% across six 3D benchmarks
HuggingPapers says Act3D uses outcome-only reinforcement learning to decide when and how to move its camera.
TLDR
HuggingPapers describes Act3D as a vision-language model that starts with one image and uses outcome-only reinforcement learning to decide when and how to move its camera to gather evidence. It reports a 53.2% average across six 3D benchmarks, +22% over its Qwen base and above Gemini 3.6 Flash.
Combined views
3.6K
2 Sources, first seen ago
