TTT3R: Test-Time Training Enhances Length Generalization in 3D Reconstruction

rsasaki0109 · x · 2026-07-31

This is an ICLR 2026 paper focusing on 3D reconstruction. While modern Recurrent Neural Networks (RNNs) are competitive due to linear-time complexity, their performance degrades significantly beyond the training context length.

The authors revisit 3D reconstruction foundation models from a Test-Time Training perspective, framing them as online learning problems. By leveraging alignment confidence between the memory state and incoming observations, they derive a closed-form learning rate for memory updates. This training-free intervention, termed TTT3R, effectively balances retaining historical information and adapting to new observations, substantially improving length generalization.

Related event: TTT3R: Enhancing 3D Reconstruction via Test-Time Training(2 posts)→

Original post →

More from Multimodal

Multimodal channel →