MOAI preprint reports 52.8% reduction in encrypted transformer evaluation time

Published Wednesday, November 11, 2026 · 2:37 AM EST

Researchers propose MOAI, a framework for running transformer models on encrypted inputs, and report a 52.8% reduction in evaluation time compared with THOR, an existing non-interactive encrypted inference system. Homomorphic encryption lets a cloud provider compute on encrypted inputs, but its computational overhead remains a bottleneck. MOAI arranges encrypted data consistently across layers to avoid format conversions and reduces costly operations called rotations. Its Softmax and LayerNorm algorithms eliminate 2,448 rotations in BERT-base inference. Applied to Powerformer, which simplifies those functions, MOAI reduces evaluation time by 55.7%. The preprint reports an amortized time of 2.36 minutes per input in a single GPU environment; the implementation is open source.

Sources

cs.CR updates on arXiv.org

Published by Tech & Business, a media brand covering technology and business. This story was sourced from cs.CR updates on arXiv.org and reviewed by the T&B editorial agent team.

Back to Newswire

Keep reading

Full wire

Security AI

Researchers find Grok Build exposes data from over 35,000 apps

Researchers found that Grok Build, SpaceXSI's AI coding agent, is leaking data from over 35,000 apps through shareable links, Ars Technica reported. The exposed material includes proprietary business information and other sensitive secrets, with some belonging to SpaceXSI itself.

Infrastructure Security

Ginux shifts to a userspace kernel with a Rocky Linux baseline

Ginux has moved active development from an independent bootable kernel to a userspace kernel that runs as an ordinary Linux process, according to its project documentation.

AI Science

Medical AI training reduces false lesion reports but misses critical external cases

Researchers report in a preprint that training MedScope-4B on images with lesions erased reduced its tendency to report findings absent from an image.

AI Science

Plumbline preprint reports 61 training problems found across five AI models

Researchers report that Plumbline, a system used in production at Amazon to validate large language model training, found 61 problems across three frameworks and five models. Of those problems, 32 were silent failures.

AI Science

TimeOmni-o1 combines text reasoning and time series forecasting in one architecture

Researchers report that TimeOmni-o1 combines a language model with a pretrained time series model to interpret text and numerical sequences together and generate both text and forecasts.

AI Science

Tree-Ising preprint reports cheaper weight repairs that improve math accuracy

Researchers report in a preprint that Tree-Ising repairs language model weights with four forward-pass requests and zero LLM tokens. On Qwen3.5-0.8B, the method scored 54.6% on MATH-500, compared with 51.7% for an earlier agent-generated repair under the same evaluation protocol.