Maverick: Private and Verifiable LLM Inference Made Practical via Matrix-Vector Multiplication Delegation
Maverick claims practical private, verifiable LLM inference by delegating matrix-vector multiplication with almost no server overhead.
The paper targets users who want open-source LLM inference without exposing prompts to a third-party provider or blindly trusting the result. Its protocol verifies delegated matrix-vector multiplication, a dominant LLM operation, and adds LPN-based pseudorandom masking for input privacy. The authors built an end-to-end prototype and evaluated it on Qwen3-4B, reporting up to 17x throughput gains over local inference with online mask generation and up to 45x with precomputed masks. Source: ArXiv · AI/CL/LG's note.
The paper targets users who want open-source LLM inference without exposing prompts to a third-party provider or blindly trusting the result. Its protocol verifies delegated matrix-vector multiplication, a dominant LLM operation, and adds LPN-based pseudorandom masking for input privacy. The authors built an end-to-end prototype and evaluated it on Qwen3-4B, reporting up to 17x throughput gains over local inference with online mask generation and up to 45x with precomputed masks. Source: ArXiv · AI/CL/LG's note.
score 5