Research map / Architectures and efficiency / Attention and transformers
Fast Spatial Memory with Scalable Elastic Test-Time Training
Ziqiao Ma, Xueyang Yu, Haoyu Zhen, Yuncong Yang, Joyce Chai, Chuang Gan
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 |
|---|---|---|---|
| add elastic consolidation with streaming-EMA anchor (4 chunks) (ablation) | LaCT · Stereo4D test (ablation, 12-layer FSM-LVSM) | PSNR: 26.908 -> 29.922 (MAS); 29.989 (SI) | Improved |
| add elastic consolidation with streaming-EMA anchor (4 chunks) (ablation) | LaCT · Stereo4D test (ablation, 12-layer FSM-LVSM) | LPIPS: 0.0988 -> 0.0519 (MAS) | Improved |
| use 4 chunks instead of 1 chunk without elasticity (ablation) | LaCT · Stereo4D test (ablation, 12-layer FSM-LVSM) | PSNR: 26.021 -> 26.908 | Improved |
| use streaming-EMA anchor instead of streaming (no EMA) anchor (ablation) | LaCET · Stereo4D test (ablation, 12-layer FSM-LVSM) | PSNR: 26.960 -> 29.922 | Improved |
| use streaming-EMA anchor instead of global anchor (ablation) | LaCET · Stereo4D test (ablation, 12-layer FSM-LVSM) | PSNR: 28.347 -> 29.922 | Improved |
| streaming anchor without EMA vs no elasticity (ablation) | LaCET · Stereo4D test (ablation, 12-layer FSM-LVSM) | PSNR: 26.908 -> 26.960 | No clear effect |
| SI vs EWC vs MAS Fisher estimate (ablation) | LaCET · Stereo4D test (ablation, 12-layer FSM-LVSM) | PSNR: 29.989 (SI); 29.781 (EWC); 29.922 (MAS) | No clear effect |
| increase batch size from 1 to 16 (ablation) | LaCET · Stereo4D test (ablation, 12-layer FSM-LVSM) | PSNR: 29.922 -> 29.928 | No clear effect |
| train with 1 chunk but test with 4 chunks (ablation) | LaCET · Stereo4D test (ablation, 12-layer FSM-LVSM) | PSNR: 29.922 -> 21.993 | Hurt |
| elastic consolidation reduces train-test gap (ablation) | LaCET · Stereo4D test (ablation, 12-layer FSM-LVSM) | training l2 loss (x10^3): 1.80 -> 2.36 | Mixed |
| choose anchor policy when chunk is full sequence (ablation) | LaCET · Stereo4D test (ablation, 12-layer FSM-LVSM) | PSNR: 26.965 vs 26.021 | No clear effect |
| FSM-LVSM vs best rendering-based baseline on 4D NVS | FSM-LVSM · Stereo4D test (7109 scenes) | PSNR: 27.19 -> 32.16 (vs MoVieS) | Improved |
Closest papers by meaning
- Test-Time Training Done Right (ICLR 2026 · similarity 0.89)
- Rethinking the State Update Gate for Long-Sequence Recurrent 3D Reconstruction (NeurIPS 2026 · similarity 0.88)
- DC-DiT: Adaptive Compute and Elastic Inference for Visual Generation via Dynamic Chunking (NeurIPS 2026 · similarity 0.87)
- BAT3R: Robust Online 3D Reconstruction with Bayesian Adaptive State Updates (NeurIPS 2026 · similarity 0.86)
- 4D-LRM: Large Space-Time Reconstruction Model From and To Any View at Any Time (NeurIPS 2025 · similarity 0.86)
- Elastic ViTs from Pretrained Models without Retraining (NeurIPS 2025 · similarity 0.85)
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.
