mlsanityData to 14 Oct 2026

Optimizing Korean-Centric LLMs via Token Pruning

arXiv 2604.16235 ↗First posted 17 Apr 2026cs.CLcited by 1

On the map: Efficient LLM inference and serving → KV cache optimization

Abstract (arXiv)

This paper presents a systematic benchmark of state-of-the-art multilingual large language models (LLMs) adapted via token pruning - a compression technique that eliminates tokens and embedding parameters corresponding to languages irrelevant to the target application. Focusing on Korean-centric natural language processing (NLP) tasks, we evaluate architectures including Qwen3, Gemma-3, Llama-3, and Aya across three vocabulary configurations: Original, English-Korean (EnKo), and English-Korean-Chinese (EnKoZh). Performance is assessed using established benchmarks for general aptitude, cultural literacy, instruction following, and machine translation. Our findings indicate that token pruning significantly improves generation stability by eliminating language confusion, and in the case of machine translation, frequently enhances performance on Korean-specific tasks. While instruction-following capabilities display architecture-dependent variance linked to latent cross-lingual representations, the significant reduction in vocabulary size validates token pruning as a highly effective optimization strategy for memory-constrained, domain-specific deployments, despite modest gains in inference latency.

Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Placed in KV cache optimization, one of 12 directions in Efficient LLM inference and serving. Builds on 11 of its 17 references that are in the map. First posted on arXiv 17 Apr 2026.

Builds on

11 of 17 references.

References · 11

  1. 2024The Llama 3 Herd of Modelscited by 14,715
  2. 2025Qwen3 Technical Reportcited by 8,820
  3. 2023xCOMET: Transparent Machine Translation Evaluation through Fine-grained Error Detectioncited by 265
  4. 2024KMMLU: Measuring Massive Multitask Language Understanding in Koreancited by 91
  5. 2025WMT24++: Expanding the Language Coverage of WMT24 to 55 Languages & Dialectscited by 79
  6. 2025Token Pruning in Multimodal Large Language Models: Are We Solving the Right Problem?cited by 76
  7. 2024Understanding and Mitigating Language Confusion in LLMscited by 75
  8. 2023HAE-RAE Bench: Evaluation of Korean Knowledge in Language Modelscited by 47
  9. 2025Seed-X: Building Strong Multilingual Translation LLM with 7B Parameterscited by 21
  10. 2024RedWhale: An Adapted Korean LLM Through Efficient Continual Pretrainingcited by 8
  11. 2024Dynamic Vocabulary Pruning in Early-Exit LLMscited by 4

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026Language Models are not Equally Robust to Non-Canonical Tokenization across Languagescited by 2

Similar, not cited

Close in wording, no citation either way

  1. 2025Pruning Laws for Large Language Models
  2. 2024Compact Language Models via Pruning and Knowledge Distillation
  3. 2023Compresso: Structured Pruning with Collaborative Prompting Learns Compact Large Language Models
  4. 2025Leveraging KV Similarity for Online Structured Pruning in LLMs
  5. 2025Systematic Evaluation of Optimization Techniques for Long-Context Language Models
  6. 2025Bridging the Linguistic Divide: A Survey on Leveraging Large Language Models for Machine Translation
  7. 2024Aggressive Post-Training Compression on Extremely Large Language Models
  8. 2025KORMo: Korean Open Reasoning Model for Everyone