@inproceedings{hu-etal-2026-lil,
    title = "Lil: Less is Less When Applying Post-Training Sparse-Attention Algorithms in Long-Decode Stage",
    author = "Hu, Junhao  and
      Li, Fangze  and
      Xu, Mingtao  and
      Meng, Feifan  and
      Zhao, Shiju  and
      Hu, Tiancheng  and
      Peng, Ting  and
      Liu, Anmin  and
      Huang, Wenrui  and
      Liu, Chenxu  and
      Hua, Ziyue  and
      Xie, Tao",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl/2026.findings-acl.91/",
    pages = "1898--1912",
    ISBN = "979-8-89176-395-1",
    abstract = "Large language models (LLMs) demonstrate strong capabilities across a wide range of complex tasks and are increasingly deployed at scale, placing significant demands on inference efficiency. Prior work typically decomposes inference into prefill and decode stages, with the decode stage dominating total latency. To reduce time and memory complexity in the decode stage, a line of work introduces sparse-attention algorithms. In this paper, we show, both empirically and theoretically, that sparse attention can paradoxically increase end-to-end complexity: information loss often induces significantly longer sequences, a phenomenon we term ``Less is Less'' (Lil). To mitigate the Lil problem, we propose an early-stopping algorithm that detects the threshold where information loss exceeds information gain during sparse decoding. Our early-stopping algorithm reduces token consumption by up to 90{\%} with a marginal accuracy degradation of less than 2{\%} across reasoning-intensive benchmarks."
}