PhD students in Dr. Weikuan Yu’s group publish a paper in the EMNLP main conference
Department of Computer Science
“SALT: Salience-Aware Lexical Trie for Long-Context Compression” by Oteo Mamo and Hyunjin Yi, PhD students in the Department of Computer Science at Florida State University, has been accepted to the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026). The work was conducted in collaboration with Joydhriti Choudhury, Shangqian Gao, and Dr. Weikuan Yu.
SALT: Salience-Aware Lexical Trie for Long-Context Compression
The research introduces SALT, a new approach for making large language models (LLMs) more efficient when processing long documents and prompts. Rather than simply keeping the highest-ranked sentences, SALT is designed to preserve information from different themes throughout a document, helping prevent less frequent but important information from being lost during compression.
Experimental results show that SALT provides a strong balance between accuracy, speed, and memory efficiency across several long-context benchmarks. The method can also support repeated questions about the same document without having to process the full document again, making it well suited for efficient long-context and multi-turn LLM applications.