Research map / Architectures and efficiency / Attention and transformers
Test-Time Training Done Right
Tianyuan Zhang, Sai Bi, Yicong Hong, Kai Zhang, Fujun Luan, Songlin Yang, Kalyan Sunkavalli, William T. Freeman, Hao Tan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use large-chunk TTT instead of full attention for prefill | LaCT · NVS prefill 48 views | prefill time: 16.1 s -> 1.4 s | Improved |
| scale up fast-weight state size (ablation) | LaCT · novel view synthesis and language modeling | PSNR / loss | Improved |
| use Muon test-time optimizer instead of vanilla GD or momentum (ablation) | LaCT · NVS and language modeling | PSNR / loss | Improved |
| use Muon update (vs GD) (ablation) | LaCT · 3B language model training throughput | tokens/s: 5.0K -> 4.3K | Hurt |
| use nonlinear SwiGLU fast weights instead of linear (ablation) | LaCT · NVS and language modeling | PSNR / loss | Improved |
| use large-chunk linear recurrence instead of per-token Mamba-2 with same state size (ablation) | LaCT linear · novel view synthesis | PSNR | Improved |
| use linear large-chunk recurrence instead of GLA SWA with same state size (ablation) | LaCT linear · language modeling | loss | Hurt |
| combine large nonlinear state with Muon (ablation) | LaCT · language modeling | per-position loss and retrieval accuracy | Improved |
| use LaCT instead of sliding window attention | LaCT · autoregressive video diffusion | denoising loss | Improved |
| use LaCT instead of 3D Gaussian Splatting under sparse views | LaCT · DL3DV scenes | PSNR | Improved |
Closest papers by meaning
- Test Time Training for Long Videos via Frame Forgetting Network (ECCV 2026 · similarity 0.91)
- Test-Time Training on Video Streams (JMLR 2025 · similarity 0.91)
- Learning What to Remember: Test-Time Training via Context Distillation (NeurIPS 2026 · similarity 0.91)
- In-Place Test-Time Training (ICLR 2026 · similarity 0.90)
- Rethinking Expressivity and Efficiency in Test-Time Training (NeurIPS 2026 · similarity 0.90)
- Let's (not) just put things in Context: Test-time Training for Long-context LLMs (ICLR 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
