🎉 Congratulations on EMNLP 2025 Paper Acceptance!
Image credit: EMNLPWe are overjoyed to share that our paper, “QUITO-X: A New Perspective on Context Compression from the Information Bottleneck Theory,” has been accepted to Findings of EMNLP 2025! 🎉
In this work, we propose a novel context compression method based on information bottleneck theory and cross-attention, enabling efficient token selection that preserves task-relevant content. 🔍
Our experiments on four major QA datasets (DROP, CoQA, SQuAD, Quoref) demonstrate that QUITO-X improves compression rates by nearly 25% over previous state-of-the-art methods, and in some cases—when removing 25% of tokens—the Exact Match (EM) scores even exceed those with full, uncompressed context! 💥
We look forward to sharing more insights and presenting this work at EMNLP 2025. Huge thanks to our amazing collaborators and supporters! 🙏
Stay tuned for further updates as the conference approaches! ✨
