You-Liang Huang, Yi Pan, Cheng Ling Tan, Y. Charlie Hu · · 2026
DOI: 10.1145/3858610.3859272
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Today, platforms of Model-as-a-Service (MaaS) provides access to various foundation models to users. Users trust that the platform indeed provides the exact service as requested. However, this trust may be misplaced for benign or malicious intentions, where results from providers are not exactly the same as what the user requested. To detect such violations, this paper introduces Espeon, a large language model (LLM) inference verification framework. It enables reproducible deterministic inference through batch-invariant kernels and includes different verification variants for flexible computation and storage budgets of verification. We evaluated Espeon with single-point manipulation cases and Espeon's performance of serving and verification. Espeon can detect such cases, and under the full verification mode it can achieve a speed-up of 4.56X compared to serving time.
No comments yet — start the discussion below.