Research map / Architectures and efficiency / Attention and transformers
Compressive Transformers for Long-Range Sequence Modelling
Jack W. Rae, Anna Potapenko, Siddhant M. Jayakumar, Chloe Hillier, Timothy P. Lillicrap
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 |
|---|---|---|---|
| compress old memories into a secondary memory | TransformerXL-style long-context language model · PG-19 test | perplexity: 36.3 -> 33.6 | Improved |
| compress old memories into a secondary memory | 24-layer TransformerXL · Enwik8 | BPC: 0.98 -> 0.97 | Improved |
| compress old memories into a secondary memory | 18-layer TransformerXL · WikiText-103 test | perplexity: 18.1 -> 17.1 | Improved |
| use convolutional compression with attention-reconstruction loss instead of convolution with BPTT (ablation) | Compressive Transformer · Enwik8 | BPC: 0.996 -> 0.973 | Improved |
| use attention-reconstruction loss rather than auto-encoding loss for convolutional compression (ablation) | Compressive Transformer · Enwik8 | BPC: 0.984 -> 0.973 | Improved |
| use convolutional compression rather than max pooling (ablation) | Compressive Transformer · Enwik8 | BPC: 0.986 -> 0.973 | Improved |
| increase compressed memory size from 512 to 3072 (ablation) | Compressive Transformer · Enwik8 test | BPC: 1.01 -> 0.97 | Improved |
| increase compressed memory size from 1536 to 4096 (ablation) | Compressive Transformer · Enwik8 test | BPC: 0.97 -> 1.00 | Hurt |
| increase compressed memory size from 256 to 1536 (ablation) | Compressive Transformer · WikiText-103 test | perplexity: 18.2 -> 17.1 | Improved |
| reduce the frequency of optimizer updates near the end of training | TransformerXL baseline · Enwik8 | BPC: 0.995 -> 0.984 | Improved |
Closest papers by meaning
- FastLongSpeech: Enhancing Large Speech-Language Models for Efficient Long-Speech Processing (NeurIPS 2025 · similarity 0.87)
- LeaPformer: Enabling Linear Transformers for Autoregressive and Simultaneous Tasks via Learned Proportions (ICML 2024 · similarity 0.87)
- Beyond Attention: Breaking the Limits of Transformer Context Length with Recurrent Memory (AAAI 2024 · similarity 0.86)
- DeciMamba: Exploring the Length Extrapolation Potential of Mamba (ICLR 2025 · similarity 0.86)
- Transformers Get Stable: An End-to-End Signal Propagation Theory for Language Models (ICML 2024 · similarity 0.86)
- SoLA: Leveraging Soft Activation Sparsity and Low-Rank Decomposition for Large Language Model Compression (AAAI 2025 · 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.
