CITATION.cff

cff-version: 1.2.0
message: "If you use this work, please cite it as below."
title: "Certified Sparse Attention: Runtime, Label-Free Verification of Fidelity in Sparse-Attention LLM Serving"
abstract: >-
  A serving-layer system that detects, at runtime and without labels, when
  sparse-attention KV-cache compression has diverged from dense execution, and
  bounds that divergence with an anytime-valid statistical certificate. Includes
  a label-free divergence detector built from discarded block scores, sampled
  dense verification via time-uniform confidence sequences, and an elastic
  scheduling formulation. Detectability and proxy-validity are confirmed at 7B
  parameters across two model architectures (Qwen2.5, Mistral).
authors:
  - family-names: Patil
    given-names: "Archana Suresh"
    alias: ArchanaChetan07
repository-code: "https://github.com/ArchanaChetan07/sparse-attention"
url: "https://github.com/ArchanaChetan07/sparse-attention"
license: MIT
version: 1.0.0
date-released: "2026-08-26"
keywords:
  - sparse attention
  - LLM inference
  - KV cache compression
  - long-context serving
  - anytime-valid inference
  - confidence sequences
  - runtime verification