Skip to content
Agent NetworkkLLM

Research

Research on confidential transformer serving

The kLLM team publishes research on confidential transformer serving. The perm transform extends STIP's permutation approach to grouped-query attention, RoPE, MoE and FP8; the research systems themselves are separate from what kllm seal ships today.

Citing kLLM and the research

Publications

  • BibTeX
    @inproceedings{yuan2025scx,
  title     = {{SCX}: Stateless {KV}-Cache Encoding for Cloud-Scale Confidential
               Transformer Serving},
  author    = {Yuan, Mu and Zhang, Lan and Zeng, Liekang and Jiang, Siyang and
               Yang, Bufang and Duan, Di and Xing, Guoliang},
  booktitle = {Proceedings of the ACM SIGCOMM 2025 Conference},
  year      = {2025},
  doi       = {10.1145/3718958.3750509}
}
  • BibTeX
    @inproceedings{yuan2026stip,
  title     = {{STIP}: Three-Party Privacy-Preserving and Lossless Inference for
               Large Transformers in Production},
  author    = {Yuan, Mu and Zhang, Lan and Cheng, Yihang and Song, Miao-Hui and
               Xing, Guoliang and Li, Xiang-Yang},
  booktitle = {Network and Distributed System Security (NDSS) Symposium 2026},
  address   = {San Diego, CA},
  year      = {2026},
  doi       = {10.14722/ndss.2026.230035}
}

Scope

Research and the shipped tool

The perm transform in kllm seal builds on the permutation approach of STIP and extends it to grouped-query attention, RoPE, mixture-of-experts layers and FP8 checkpoints.

The research systems are separate from what kllm seal ships today. Citing SCX or STIP does not describe the security properties of kLLM, and the security properties of kLLM are stated only in its own security documentation.

Further work is under submission and will be released after publication.

Citation

Cite kLLM

kLLM has no tagged release yet. Cite the repository with the version that kllm version prints and, where possible, the commit used. A citation of a measurement made with kLLM states its conditions: the GPU, the engine and its version, the precision, the input and output lengths and the concurrency.

BibTeX
@software{kllm,
  title   = {{kLLM}: Sealed Models on Stock Inference Engines},
  author  = {{Agent Network}},
  year    = {2026},
  version = {0.1.0.dev0},
  url     = {https://github.com/ANetResearch/kLLM},
  license = {Apache-2.0}
}