Research map / Architectures and efficiency / Attention and transformers
Proxy Compression for Language Modeling
Lin Zheng, Xinyu Li, Qian Liu, Xiachong Feng, Lingpeng Kong
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 |
|---|---|---|---|
| train with proxy (tokenizer) compression mixed with raw bytes instead of pure byte-level | EvaByte 14B · HumanEval-Plus | pass@1: 24.4 -> 30.5 | Improved |
| train with proxy (neural) compression instead of pure byte-level | EvaByte 14B · MBPP-Plus | pass@1: 42.1 -> 49.2 | Improved |
| proxy (tokenizer) vs tokenizer-based model at 0.5B | EvaByte 0.5B · HumanEval-Plus | pass@1: 17.7 -> 12.2 | Hurt |
| proxy (tokenizer) vs tokenizer-based model at 14B | EvaByte 14B · MBPP-Plus | pass@1: 48.1 -> 49.5 | Improved |
| longer training (320B symbols) with neural proxy vs byte-level | EvaByte 7B · HumanEval-Plus | pass@1: 14.6 -> 21.3 | Improved |
| add proxy (tokenizer) compression vs byte-level at 1.5B, 320B symbols | EvaByte 1.5B · HumanEval-Plus | pass@1: 9.1 -> 12.8 | Improved |
| add always-on translation pairing instead of no pairs (tokenizer proxy) (ablation) | Proxy (Tokenizer) 1.5B · HumanEval-Plus oracle translation | oracle-translation pass@1: 45.7 -> 96.3 | Improved |
| use pairing in warm-up only instead of always-on (tokenizer proxy) (ablation) | Proxy (Tokenizer) 1.5B · HumanEval-Plus | pass@1: 17.0 -> 20.7 | Improved |
| warm-up-only pairing vs no pairs (neural proxy) (ablation) | Proxy (Neural) 1.5B · HumanEval-Plus | pass@1: 14.6 -> 18.9 | Improved |
| mix gzip-compressed data into training (increase gzip proportion) (ablation) | EvaByte 1.5B · HumanEval-Plus | pass@1 | Hurt |
| use proxy (tokenizer) rather than tokenizer-based model | EvaByte 7B · ReCode HumanEval | robust pass@1: 14.9 -> 19.1 | Improved |
| use neural proxy instead of byte-level | EvaByte 7B · ReCode HumanEval | robust pass@1: 18.7 -> 19.8 | Improved |
Closest papers by meaning
- Language Modeling Is Compression (ICLR 2024 · similarity 0.89)
- Accuracy is Not All You Need (NeurIPS 2024 · similarity 0.88)
- zip2zip: Inference-Time Adaptive Tokenization via Online Compression (NeurIPS 2025 · similarity 0.87)
- Compact Language Models via Pruning and Knowledge Distillation (NeurIPS 2024 · similarity 0.87)
- ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference (NeurIPS 2025 · similarity 0.87)
- BitStack: Any-Size Compression of Large Language Models in Variable Memory Environments (ICLR 2025 · similarity 0.87)
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.
