@inproceedings{rafiuddin-khan-2025-learning,
    title = "Learning What to Remember: Adaptive Probabilistic Memory Retention for Memory-Efficient Language Models",
    author = "Rafiuddin, S M  and
      Khan, Muntaha Nujat",
    editor = "Christodoulopoulos, Christos  and
      Chakraborty, Tanmoy  and
      Rose, Carolyn  and
      Peng, Violet",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2025",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/author-page-yu-wang-polytechnic/2025.findings-emnlp.212/",
    doi = "10.18653/v1/2025.findings-emnlp.212",
    pages = "3969--3981",
    ISBN = "979-8-89176-335-7",
    abstract = "Transformer attention scales quadratically with sequence length $O(n^2)$, limiting long-context use. We propose \textit{Adaptive Retention}, a probabilistic, layer-wise token selection mechanism that learns which representations to keep under a strict global budget $M$. Retention is modeled with Bernoulli gates trained via a Hard-Concrete/variational relaxation and enforced with a simple top-$M$ rule at inference, making the method differentiable and drop-in for standard encoders. Across classification, extractive QA, and long-document summarization, keeping only 30{--}50{\%} of tokens preserves $\geq 95\%$ of full-model performance while cutting peak memory by $\sim 35${--}45{\%} and improving throughput by up to $\sim 1.8\times$. This architecture-agnostic approach delivers practical long-context efficiency without modifying base attention or task heads."
}