Research map / Architectures and efficiency / Attention and transformers
Unbounded cache model for online language modeling with open vocabulary
Edouard Grave, Moustapha M Cisse, Armand Joulin
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 the unbounded cache to the static language model | Static model trained on News 2007 · News 2007–2011 | perplexity: 220.9 -> 166.5; 237.6 -> 191.4; 256.2 -> 202.6; 259.7 -> 204.8; 268.8 -> 214.3 | Improved |
| Use a unigram cache instead of only the static model | Static model trained on News 2007 · News 2007–2011 | perplexity: 220.9 -> 220.3; 237.6 -> 235.9; 256.2 -> 252.6; 259.7 -> 256.1; 268.8 -> 264.3 | No clear effect |
| Use a local cache instead of only the static model | Static model trained on News 2007 · News 2007–2011 | perplexity: 220.9 -> 218.9; 237.6 -> 234.5; 256.2 -> 250.5; 259.7 -> 256.2; 268.8 -> 265.2 | No clear effect |
| Add the unbounded cache instead of the strongest cache baseline | Static model trained on News · News → Commentary | perplexity: 288.5 -> 191.1 | Improved |
| Add the unbounded cache instead of the strongest cache baseline | Static model trained on News · News → Web | perplexity: 581.1 -> 383.4 | Improved |
| Add the unbounded cache instead of the strongest cache baseline | Static model trained on News · News → Wiki | perplexity: 316.5 -> 337.4 | Hurt |
| Add the unbounded cache instead of the strongest cache baseline | Static model trained on News · News → Books | perplexity: 240.3 -> 237.2 | Improved |
| Add the unbounded cache instead of the strongest cache baseline | Static model trained on Web · Web → News | perplexity: 519.2 -> 306.3 | Improved |
| Add the unbounded cache instead of the strongest cache baseline | Static model trained on Web · Web → Commentary | perplexity: 391.3 -> 234.9 | Improved |
| Add the unbounded cache instead of the strongest cache baseline | Static model trained on Web · Web → Wiki | perplexity: 321.5 -> 340.2 | Hurt |
| Add the unbounded cache instead of the strongest cache baseline | Static model trained on Web · Web → Books | perplexity: 235.8 -> 223.6 | Improved |
| Add the unbounded cache instead of the strongest cache baseline | Static model trained on Wiki · Wiki → News | perplexity: 532.8 -> 318.7 | Improved |
Closest papers by meaning
- Training Free Exponential Context Extension via Cascading KV Cache (ICLR 2025 · similarity 0.89)
- MiniCache: KV Cache Compression in Depth Dimension for Large Language Models (NeurIPS 2024 · similarity 0.88)
- LaCache: Ladder-Shaped KV Caching for Efficient Long-Context Modeling of Large Language Models (ICML 2025 · similarity 0.88)
- EntropyCache: Decoded Token Entropy Guided KV Caching for Diffusion Language Models (NeurIPS 2026 · similarity 0.88)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.88)
- Cache Me If You Must: Adaptive Key-Value Quantization for Large Language Models (ICML 2025 · similarity 0.88)
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.
