This article was compiled and organized by BlockWeeks
Six Yale University computer scientists have proposed a protocol called Maverick, with the goal of enabling open-source AI models to run primarily on "someone else's machine" while not exposing the user's raw input to that machine. By design, a user's phone or other low-compute device outsources the heavy computation to an external server; the server only receives a scrambled copy of the input, and the device then verifies the result after it is returned.
The related paper was uploaded to arXiv on September 9, 2026. It is still a preprint and has not yet been accepted by a conference or journal, and the paper does not include a code link.
Core Problem: How to Trust a Machine You Do Not Control
Verifying whether a machine you cannot control has honestly completed a computation is exactly the problem that zero-knowledge proof systems and some decentralized compute networks have long sought to solve. Maverick's additional goal is to verify the result while keeping the server from ever seeing the user's input.
The paper points out that the zero-knowledge machine learning system DeepProve can check computation results but exposes the input to the server. Maverick attempts to achieve both "verifiability" and "input confidentiality" at the same time.
One of the paper's authors, assistant professor Fan Zhang, and doctoral student Wenhao Wang list the Initiative for Cryptocurrencies and Contracts (IC3) as their affiliation, and part of Wenhao Wang's research is funded by IC3.
Regarding blockchain markets that sell AI inference services, host Kain Warwick described the difficulty of trusting other people's machines on a program this month: "People are selling you inference services, but you don't really know where it comes from."
How Maverick Works: Mask First, Then Spot-Check
Maverick's basic flow is as follows: the user's device hides the input with a layer of seemingly random masking, then sends most of the model's mathematical computation to the server; the server completes the computation without knowing the original input, and after returning the result, the user's device removes the mask and recovers the answer.
To prevent the server from cutting corners or acting maliciously, the device does not re-execute the entire computation; instead, it runs a fast randomized test to spot-check the result returned by the server.
The paper's author Charalampos Papamanthou lists only Yale University in the paper, but according to his Yale personal page, he also serves as chief scientist at Lagrange Labs. Lagrange builds zero-knowledge proof systems, including DeepProve, and Papamanthou is also a co-author of the DeepProve paper.
Performance Data: The 45x Speedup Has Strict Preconditions
In the paper's end-to-end tests, the largest performance improvement appeared when the user's device prepared the mask in advance: prompt token processing speed reached up to 45 times that of running the model locally on the device itself. If the mask is constructed on the fly during execution, the peak speedup drops to 17 times.
These numbers come from specific test conditions: the user's device uses only one processing thread, the server is a CPU and is allowed to use up to 128 threads, the prompt length is 8 tokens, and network latency is not included.
Another test timed a short request on a simulated network, with the device using two threads and the server using 64 threads. At 50 milliseconds of round-trip latency, the time to first output under the two privacy modes was only 1.42 times and 1.46 times that of local execution, respectively; when round-trip latency rose to 100 milliseconds, both modes were slower than local execution.
The tests used only one model, Qwen3-4B.
Key Bottleneck: Client-Side Storage
The authors call Maverick's client-side storage requirement "a key bottleneck of our method." This means that although the protocol shifts computational pressure to the server, the user's device still needs to bear non-negligible storage overhead, which may limit its practical deployment on devices such as low-end phones.
In terms of positioning, Maverick seeks to enter a heating intersection: AI inference outsourcing, privacy protection, and verifiable computation. Whether it can move from a preprint to a deployable system still depends on code implementation, testing with larger models, and whether client-side overhead can be further compressed.






