# Yale Researchers Describe Maverick for Private, Verifiable AI Computing

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/62260
Published: 2026-11-20T18:44:42.000Z
Updated: 2026-11-20T18:44:42.000Z
Section: Technology

> Tests on Qwen3-4B showed up to 45 times faster prompt processing than running the model on a user’s device when masks were prepared in advance.

Six Yale computer scientists described Maverick, a protocol designed to keep prompts private while allowing users to check computations performed by outside AI servers. The system was tested only with Qwen3-4B, an open-source model.

Maverick sends a masked version of a user’s input to the server. The user’s device later strips the mask from the answer and runs a randomized check of the server’s computation.

A preprint posted Sept. 9, 2026, describes the protocol. It has not been accepted by a conference or journal, and the paper does not link to code. Assistant professor Fan Zhang and PhD student Wenhao Wang list the Initiative for CryptoCurrencies and Contracts (IC3) as an affiliation.

In one test, prompt tokens were processed up to 45 times faster than when the model ran on the user’s device, with masks prepared in advance. Preparing masks during the run reduced the maximum speedup to 17 times. The test used one processing thread on the user’s device, a CPU server with up to 128 threads and an eight-token prompt; it did not include network delays.

A separate simulated-network test measured one short request using two device threads and 64 server threads. At a 50-millisecond round trip, the two private modes returned their first output 1.42 and 1.46 times faster than local processing. At a 100-millisecond round trip, both were slower than local processing.

The authors describe client storage needs as a key bottleneck. Maverick is designed to keep the prompt hidden from the server and let the user check whether the server performed the computation. DeepProve, a zero-knowledge machine-learning system, checks results but leaves the input visible to the server.


