Research map / Architectures and efficiency / Attention and transformers
MEGABYTE: Predicting Million-byte Sequences with Multiscale Transformers
LILI YU, Daniel Simig, Colin Flaherty, Armen Aghajanyan, Luke Zettlemoyer, Mike Lewis
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 |
|---|---|---|---|
| remove the local model (ablation) | Megabyte · arXiv | bits-per-byte: 1.263 -> 0.6871 | Improved |
| add the global model (ablation) | Megabyte · arXiv | bits-per-byte: 1.373 -> 0.6871 | Improved |
| add cross-patch attention (ablation) | Megabyte · ImageNet256 | bits-per-byte: 3.259 -> 3.158 | Improved |
| add cross-patch attention (ablation) | Megabyte · arXiv | bits-per-byte: 0.6781 -> 0.6871 | Hurt |
| add CNN patch encoder (ablation) | Megabyte · ImageNet256 | bits-per-byte: 3.158 -> 3.155 | No clear effect |
| strided inference | Megabyte · PG-19 test | bits-per-byte: 0.9079 -> 0.8926 | Improved |
| strided plus sliding window inference | Megabyte · PG-19 test | bits-per-byte: 0.9079 -> 0.8751 | Improved |
| patch size 192 versus 48 or 768 (ablation) | Megabyte · ImageNet256 | bits-per-byte: 3.158 vs 3.178 / 3.186 | No clear effect |
| patch scan instead of raster scan (ablation) | Megabyte · ImageNet256 | bits-per-byte: 3.428 -> 3.158 | Improved |
| larger global model with smaller local model (ablation) | Megabyte · PG-19 | bits-per-byte: 1.014 -> 0.991 | Improved |
| double context length to 16,384 (ablation) | Megabyte · PG-19 | bits-per-byte: 0.8751 -> 0.8787 | No clear effect |
| Megabyte vs byte-level Transformer, same compute | Megabyte · PG-19 | bits-per-byte: 1.057 -> 1.000 | Improved |
Later papers that compared with this method
2 later accepted papers report MegaByte as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- SpaceByte: Towards Deleting Tokenization from Large Language Modeling (NeurIPS 2024 · better on 3 of 3 shared results)
- Megalodon: Efficient LLM Pretraining and Inference with Unlimited Context Length (NeurIPS 2024 · 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
- SpaceByte: Towards Deleting Tokenization from Large Language Modeling (NeurIPS 2024 · similarity 0.87)
- Compression via Pre-trained Transformers: A Study on Byte-Level Multimodal Data (ICML 2025 · similarity 0.87)
- BitStack: Any-Size Compression of Large Language Models in Variable Memory Environments (ICLR 2025 · similarity 0.86)
- Megalodon: Efficient LLM Pretraining and Inference with Unlimited Context Length (NeurIPS 2024 · similarity 0.85)
- mPLUG-Owl3: Towards Long Image-Sequence Understanding in Multi-Modal Large Language Models (ICLR 2025 · similarity 0.85)
- The Hyperfitting Phenomenon: Sharpening and Stabilizing LLMs for Open-Ended Text Generation (ICLR 2025 · similarity 0.85)
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.
