Research map / Architectures and efficiency / Attention and transformers
Token Turing Machines
Michael S. Ryoo, Keerthana Gopalakrishnan, Kumara Kahatapitiya, Ted Xiao, Kanishka Rao, Austin Stone, Yao Lu, Julian Ibarz, Anurag Arnab
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 TTM to ViViT-B | ViViT-B · Charades v1 localize validation | mAP: 23.18 -> 26.34 | Improved |
| Add TTM to ViViT-L | ViViT-L · Charades v1 localize validation | mAP: 26.01 -> 28.79 | Improved |
| Use TTM-Transformer with 3136 input tokens rather than a temporal Transformer with 3360 tokens | ViViT-B · Charades v1 localize validation | mAP: 25.53 -> 26.34 | Improved |
| Add TTM per video | ViViT-B · AVA v2.2 | mAP: 25.2 -> 27.9 | Improved |
| Add TTM per box | ViViT-B · AVA v2.2 | mAP: 25.2 -> 31.5 | Improved |
| Use TTM memory for policy history | BC-Zero EfficientNet · Real-robot evaluation | task success rate: 80.08 -> 89.26 | Improved |
| Replace Transformer processing unit with MLP (ablation) | TTM · Charades v1 localize validation | mAP: 26.34 -> 23.34 | Hurt |
| Replace Transformer processing unit with MLPMixer (ablation) | TTM · Charades v1 localize validation | mAP: 26.34 -> 26.14 | No clear effect |
| Replace MLP summarization with latent-query summarization (ablation) | TTM · Charades v1 localize validation | mAP: 26.34 -> 26.75 | Improved |
| Replace learned token summarization with pooling (ablation) | TTM · Charades v1 localize validation | mAP: 26.34 -> 25.75 | Hurt |
| Replace bounded memory update with concatenation of observed input tokens (ablation) | TTM · Charades v1 localize validation | mAP: 26.34 -> 20.97 | Hurt |
| Replace token summarization write with erase-and-add memory writes (ablation) | TTM · Charades v1 localize validation | mAP: 26.34 -> 25.86 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MViT — M³ViT: Mixture-of-Experts Vision Transformer for Efficient Multi-task Learning with Model-Accelerator Co-design (NeurIPS 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Seer: Language Instructed Video Prediction with Latent Diffusion Models (ICLR 2024 · similarity 0.85)
- PRISE: LLM-Style Sequence Compression for Learning Temporal Action Abstractions in Control (ICML 2024 · similarity 0.85)
- A Large Recurrent Action Model: xLSTM enables Fast Inference for Robotics Tasks (ICML 2025 · similarity 0.84)
- Token Bottleneck: One Token to Remember Dynamics (NeurIPS 2025 · similarity 0.84)
- Continuous Thought Machines (NeurIPS 2025 · similarity 0.84)
- Distilling Sequential Computation in Transformer Language Models (NeurIPS 2026 · similarity 0.84)
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.
