Markets
BTC $59,060 +1.97% ETH $1,665 +2.23% BNB $540 +3.06% XRP $0.9153 +2.65% SOL $69.19 +2.55% LINK $8.56 +2.87% BTC $59,060 +1.97% ETH $1,665 +2.23% BNB $540 +3.06% XRP $0.9153 +2.65% SOL $69.19 +2.55% LINK $8.56 +2.87% TOTAL $2.06T BTC.D 57.6%

Blockchain Technology

Yale Researchers Pitch a Way to Run AI Models Without Showing the Server Your Prompt

The scheme hides the input from the rented machine and verifies its output with a randomised test, but latency gains shrink on slow networks.

Six Yale computer scientists have described a protocol that lets a phone or laptop run an open-source AI model mostly on someone else’s machine while keeping the user’s input hidden from that machine. The device masks the prompt, sends the heavy arithmetic out, and strips the mask from the answer when it returns.

The paper, called Maverick, went up on arXiv on Sept. 9, 2026, as a preprint. It has not been accepted at a conference or journal and does not link to code. Qwen3-4B was the only model tested.

The design splits the problem in two. Hiding the input is one part: the user’s device applies a random-looking transformation that the server cannot invert. Checking the work is the other: instead of redoing the computation, the device runs a quick randomised test on the returned result to catch a server that computed dishonestly.

That pairing is what separates it from existing approaches. DeepProve, a zero-knowledge machine-learning system, verifies results but leaves the input exposed to the server, according to the paper. Decentralised compute networks attack the same trust problem from the market side; as one host put it this month, buyers are purchasing inference and “don’t really know where it’s coming from.”

The numbers are encouraging in narrow conditions. When the user’s device prepared its masks in advance, prompt tokens moved up to 45 times faster than running the model locally. Building masks during the run cut the peak gain to 17 times. Those results come from one processing thread on the device working against a CPU server allowed up to 128 threads, with an eight-token prompt and no network latency counted.

Add a network and the advantage narrows sharply. In a simulated test with two device threads, 64 server threads and a 50-millisecond round trip, the two private modes returned their first output just 1.42 and 1.46 times faster than running locally. At 100 milliseconds they were slower than local execution.

The authors are direct about the limits. Client-side storage is described as “a key bottleneck of our approach,” and the speed tests cover one short prompt on one small model. Latency, not throughput, decides whether masking a prompt is worth the trouble, and at realistic mobile round trips the margin is thin.

The affiliations show where the work sits. Two authors, assistant professor Fan Zhang and PhD student Wenhao Wang, list the Initiative for CryptoCurrencies and Contracts, which part-funds Wang’s research. Co-author Charalampos Papamanthou is chief scientist at Lagrange Labs, which builds zero-knowledge proof systems including DeepProve. The overlap is not incidental: the privacy guarantee comes from the same cryptographic toolbox blockchains use to prove a computation was performed without re-running it.

Share this article

Viber uses viber://forward?text= when the app is installed.