Research map / Large language models / Language, safety and interpretability
Byte-Level Machine Reading Across Morphologically Varied Languages
Tom Kenter, Llion Jones, Daniel Hewlett
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 |
|---|---|---|---|
| read bytes instead of words (multi-level RNN) | RNN · WikiReading Turkish | mean F1: 0.6365 -> 0.6956 | Improved |
| read bytes instead of words (encoder-transformer-decoder) | Enc-trans-dec · WikiReading Russian | mean F1: 0.5759 -> 0.5808 | Improved |
| read bytes instead of words (encoder-transformer-decoder) | Enc-trans-dec · WikiReading English | mean F1: 0.7365 -> 0.7182 | Hurt |
| use hybrid word-byte reading instead of word-level | Word-byte · WikiReading English | mean F1: 0.7365 -> 0.7418 | Improved |
| use hybrid word-byte reading instead of word-level | Word-byte · WikiReading Turkish | mean F1: 0.6365 -> 0.6654 | Improved |
| add bidirectional encoder instead of unidirectional RNN | byte-level RNN · WikiReading Turkish | mean F1: 0.6956 -> 0.6627 | Hurt |
| use convolutional-recurrent encoder instead of RNN | byte-level encoder · WikiReading Turkish | mean F1: 0.6956 -> 0.5753 | Hurt |
| let decoder attend document encoder states (encoder-transformer-decoder) instead of memory states | byte-level memory network · WikiReading Turkish | mean F1: 0.6899 -> 0.6956 | Improved |
| use GRU instead of LSTM cells | byte-level models · hyperparameter search | mean F1 | Improved |
| tie input and output embeddings | byte-level models · hyperparameter search | mean F1 | Improved |
Closest papers by meaning
- Hierarchical Autoregressive Transformers: Combining Byte- and Word-Level Processing for Robust, Adaptable Language Models (ICLR 2025 · similarity 0.84)
- Over-Tokenized Transformer: Vocabulary is Generally Worth Scaling (ICML 2025 · similarity 0.84)
- Exploiting Vocabulary Frequency Imbalance in Language Model Pre-training (NeurIPS 2025 · similarity 0.83)
- Scaling Laws with Vocabulary: Larger Models Deserve Larger Vocabularies (NeurIPS 2024 · similarity 0.83)
- From Bytes to Ideas: Language Modeling with Autoregressive U-Nets (NeurIPS 2025 · similarity 0.83)
- Language Models as Hierarchy Encoders (NeurIPS 2024 · similarity 0.83)
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.
