Research map / Architectures and efficiency / Attention and transformers
Test-Time Training with KV Binding Is Secretly Linear Attention
Junchen Liu, Sven Elflein, Or Litany, Zan Gojcic, Ruilong Li
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 |
|---|---|---|---|
| increase inner-loop iterations at inference (ablation) | pretrained LaCT · LLM and NVS | perplexity / PSNR | Hurt |
| replace gradient descent with gradient ascent in inner loop (retrained) (ablation) | LaCT-LLM · Book-3 perplexity | perplexity: 16.43 -> 16.19 | No clear effect |
| replace gradient descent with gradient ascent in inner loop (retrained) (ablation) | ViTTT · ImageNet classification | top-1 accuracy: 79.34 -> 79.61 | No clear effect |
| replace query with key at output (ablation) | LaCT-NVS · NVS | PSNR: 25.94 -> 25.95 | No clear effect |
| replace query with key at output (ablation) | ViTTT · image classification | top-1 accuracy: 79.34 -> 79.18 | No clear effect |
| update only last-layer inner-loop parameters (ablation) | LaCT-LLM · Book-3 perplexity | perplexity: 16.43 -> 15.93 | Improved |
| update only last-layer inner-loop parameters (ablation) | ViTTT · image classification | top-1 accuracy: 79.34% -> 79.63% | Improved |
| remove weight normalization (ablation) | LaCT-LLM Variant 1 · Book-3 perplexity | perplexity: 15.93 -> 16.31 | Hurt |
| replace multi-layer MLP with single linear layer (ablation) | LaCT-NVS · NVS | PSNR: 25.93 -> 25.71 | Hurt |
| remove per-token learnable learning rate (ablation) | LaCT-LLM Variant 3 · Book-3 perplexity | perplexity: 16.23 -> 16.12 | No clear effect |
| remove momentum in inner-loop SGD (ablation) | LaCT-LLM Variant 4 · Book-3 perplexity | perplexity: 16.12 -> 15.97 | No clear effect |
| add gradient orthogonalization (ablation) | linear attention (Variant 6) on LaCT-LLM · Book-3 perplexity | perplexity: 16.80 -> 15.97 | Improved |
Closest papers by meaning
- Learning to (Learn at Test Time): RNNs with Expressive Hidden States (ICML 2025 · similarity 0.88)
- Rethinking Expressivity and Efficiency in Test-Time Training (NeurIPS 2026 · similarity 0.88)
- Learning What to Remember: Test-Time Training via Context Distillation (NeurIPS 2026 · similarity 0.88)
- In-Place Test-Time Training (ICLR 2026 · similarity 0.86)
- Test-Time Training Provably Improves Transformers as In-context Learners (ICML 2025 · similarity 0.86)
- Raven: High-Recall Sequence Modeling via Sparse Memory Routing (NeurIPS 2026 · similarity 0.86)
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.
