# R24: Maverick: Private and Verifiable LLM Inference Made Practical via Matrix-Vector Multiplication Delegation

Maverick delegates matrix-vector multiplication with transparent preprocessing, information-theoretic batch verification, and LPN-based pseudorandom input masking while adding virtually no server-side verification overhead. PLLM status: planned.

[View canonical HTML](https://pllm.run/research/papers/r24-maverick/)

Document ID: `pllm.docs.research.papers.r24-maverick`  
Release: `0.1.0`

**Status:** `planned`
**Year:** 2026
**Authors:** Ben Merbaum, Mohammad Amin Raeisi, Wenhao Wang, Charalampos Papamanthou, Katerina Sotiraki, Fan Zhang
**External source:** [Maverick: Private and Verifiable LLM Inference Made Practical via Matrix-Vector Multiplication Delegation](https://arxiv.org/abs/2609.10264v1)

## Source contribution

Maverick delegates matrix-vector multiplication with transparent preprocessing, information-theoretic batch verification, and LPN-based pseudorandom input masking while adding virtually no server-side verification overhead.

## PLLM perspective

The delegated linear operation is close to PLLM's prepared public-weight path, but the protocols and claims are not interchangeable. PLLM currently uses one-time masks and preloaded `W·r-s` corrections under an honest-but-curious, non-colluding preparation/inference assumption. It does not implement Maverick's verification or LPN masking construction.

An independent PLLM reproduction should first specify the paper's exact matrix-vector, masking, and verification interfaces; implement bounded native reference primitives; and compare them with the same Qwen3-4B workload and party/resource accounting. Similar data flow is not evidence of implementation, correctness, privacy, or verifiability.

This work is planned. PLLM does not yet implement Maverick's delegation, LPN masking, or batch verification, and claims no reproduction or matched benchmark.

The source's boundary remains binding: The paper's input-privacy and verifiability claims depend on its exact transparent-preprocessing, LPN, and verification constructions. Similarity to PLLM's prepared matrix path does not transfer those claims to PLLM.

## PLLM modules and documentation

Target module identifiers: `protocol.delegated_matvec`, `gate.batch_verification`, `preparation.lpn_masking`. These identifiers describe scoped work unless this page explicitly links an implementation source. Dependencies: None.

- [Research backlog](/research/backlog/)
- [Reimplementation backlog](/research/backlog/)
- [Clean-room workflow](/research/clean-room/)
- [Masked linear protocol](/sdk/pipeline/protocols/masked-linear/)

## Backlog

This source is assigned to reimplementation priority 1, **Delegated matrix-vector execution**. Its source-registry order is 24. Follow current gates and unfinished work in the [research backlog](/research/backlog/).