Research map / Architectures and efficiency / Attention and transformers
RMAAT: Astrocyte-Inspired Memory Compression and Replay for Efficient Long-Context Transformers
Md Zesun Ahmed Mia, Malyaban Bal, Abhronil Sengupta
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 memory retention factor (ablation) | RMAAT · LRA Retrieval (8K) | accuracy (%): 80.5 -> 83.2 | Improved |
| add memory retention factor (ablation) | RMAAT · LRA Text (4K) | accuracy (%): 64.9 -> 65.9 | Improved |
| train with AMRB instead of BPTT (ablation) | RMAAT · LRA Retrieval (8K) | peak memory (GB): 15.0 -> 3.4 | Improved |
| train with AMRB instead of BPTT (ablation) | RMAAT · LRA Text (4K) | peak memory (GB): 22.0 -> 5.1 | Improved |
| add retention factor and AMRB to RLT (ablation) | RLT · LRA Retrieval (8K) | accuracy (%): 78.4 -> 79.2 | Improved |
| add astromorphic attention (H_astro, P) on top of RLT+AMRB (ablation) | RLT + AMRB · LRA Retrieval (8K) | accuracy (%): 79.2 -> 83.2 | Improved |
| process longer total context (more segments) (ablation) | RMAAT · LRA Retrieval | accuracy (%): 65.3 -> 71.5 -> 83.2 | Improved |
| RMAAT recurrence/memory vs non-recurrent AT | AT · LRA average | accuracy (%): 56.4 -> 68.0 | Improved |
| RMAAT vs RMT with standard attention and BPTT | RMT · LRA Retrieval | training speed: 1x -> 1.73x | Improved |
| RMAAT vs RMT on ListOps | RMT · LRA ListOps (2K) | accuracy (%): 37.4 -> 38.9 | No clear effect |
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.
- Luna-256 — Luna: Linear Unified Nested Attention (NeurIPS 2021) · 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
- MA-Net: Rethinking Neural Unit in the Light of Astrocytes (AAAI 2024 · similarity 0.84)
- GliaNet: Adaptive Neural Network Structure Learning with Glia-Driven (CVPR 2025 · similarity 0.83)
- SPICED: A Synaptic Homeostasis-Inspired Framework for Unsupervised Continual EEG Decoding (NeurIPS 2025 · similarity 0.82)
- The Expressive Leaky Memory Neuron: an Efficient and Expressive Phenomenological Neuron Model Can Solve Long-Horizon Tasks. (ICLR 2024 · similarity 0.82)
- Predictive Attractor Models (NeurIPS 2024 · similarity 0.82)
- Exponential Dynamic Energy Network for High Capacity Sequence Memory (NeurIPS 2025 · similarity 0.82)
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.
