tttLRM: Test-Time Training for Long Context and Autoregressive 3D Reconstruction
Chen Wang, Hao Tan, Wang Yifan, Zhiqin Chen, Yuheng Liu, Kalyan Sunkavalli, Sai Bi, Lingjie Liu, Yiwei Hu, University of Pennsylvania · 2026
Read tttLRM: Test-Time Training for Long Context and Autoregressive 3D Reconstruction by Chen Wang, Hao Tan, Wang Yifan, Zhiqin Chen, Yuheng Liu, Kalyan Sunkavalli, Sai Bi, Lingjie Liu, Yiwei Hu, University of Pennsylvania (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-20160.txt.
Description
Rebuilding three-dimensional scenes from images usually needs many views and heavy compute. Long image sequences strain context limits. A 2026 arXiv paper by Chen Wang and eight co-authors proposes tttLRM, a large three-dimensional model. It uses a test-time training layer for long-context, step-by-step rebuilding with linear complexity.
The framework packs many image views into the test-time training layer's fast weights. That forms an implicit three-dimensional model in latent space. It decodes into direct formats such as Gaussian Splats for downstream use. An online learning variant supports rebuilding and polish from streaming views. Early training on novel view creation transfers well to direct three-dimensional building, raising rebuilding quality with faster learning. Broad experiments show better one-pass three-dimensional Gaussian rebuilding, beating leading approaches on both objects and scenes.
Catalog record
| CREATOR | Chen Wang, Hao Tan, Wang Yifan, Zhiqin Chen, Yuheng Liu, Kalyan Sunkavalli, Sai Bi, Lingjie Liu, Yiwei Hu, University of Pennsylvania |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 44 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-20160