Matrix-Driven Instant Review: Confident Detection and Reconstruction of LLM Plagiarism on PC

Ruichong Zhang · 2025

Read Matrix-Driven Instant Review: Confident Detection and Reconstruction of LLM Plagiarism on PC by Ruichong Zhang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-06309.txt.

Introduction

Ruichong Zhang and Daniel Goldstein presented a matrix method for telling whether one language model was copied from another in a 2025 paper. The head title frames the work as finding weight homology between models. The method needs no test runs of the models and works on modest hardware, since it compares one pair of weight matrices at a time.

Matrix tools and large deviation theory turn the comparison into a match score with a rigorous significance value. That lets the team rebuild how weights in one model map to weights in another. On a public benchmark the method scored perfectly on both ranking and accuracy measures across source models. The authors pitch it as a practical guard against unattributed reuse of model weights.

Catalog record

CREATORRuichong Zhang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE58 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-06309

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Matrix-Driven Instant Review: Confident Detection and Reconstruction of LLM Plagiarism on PC Ruichong Zhang Abstract In recent years, concerns about intellectual property (IP) in large language models (LLMs) have grown significantly. Plagiarizing other LLMs (through direct weight copying, upcycling, pruning, or continual pretraining) and claiming authorship without properly attributing to the original license, is a serious misconduct that can lead to significant financial and reputational harm to the original developers. However, existing methods for detecting LLM plagiarism fall short in key areas. They fail to accurately reconstruct weight correspondences, lack the ability to compute statistical significance measures such as p p -values, and may mistakenly flag models trained on similar data as being related. To address these limitations, we propose Matrix-Driven Instant Review (MDIR), a novel method that leverages matrix analysis and Large Deviation Theory. MDIR achieves accurate reconstruction of weight relationships, provides rigorous p p -value estimation, and focuses exclusively on weight similarity without requiring full model inference. Experimental results demonstrate that MDIR reliably detects plagiarism even after extensive transformations, such as random permutations and continual pretraining with trillions of tokens. Moreover, all detections can be performed on a single PC within an hour, making MDIR both efficient and accessible. 1 Introduction In recent years, the rapid advancement of large language models (LLMs) has resulted in a proliferation of models trained on vast amounts of data, achieving impressive performance across various tasks. However, this progress has also raised concerns about the potential misuse of model weights, such as direct copying, upcycling [ 52 , 19 ] , pruning [ 25 , 1 ] , and continual pretraining, followed by claims of authorship without proper attribution to the original developers. These practices, commonly referred to as "plagiarizing" other LLMs, severely undermine the integrity of the research community and erode public trust. The complexity and size of LLMs make it particularly challenging to detect such plagiarism in a timely and efficient manner. Existing methods for detecting model similarity can be broadly classified into two main categories: retrieval-based methods and representation-based methods. Retrieval-based Methods. Retrieval-based methods [ 49 ] rely on vendors designing specific key-value pairs {(k_{i},v_{i})}_{i=1,\cdots,N} for language model watermarking. These (k_{i},v_{i}) pairs are typically generated randomly (e.g., (k=’396f4b37...’, v=’586a99’) ) to ensure they rarely appear in conventional conversational scenarios. During pretraining, vendors inject these key-value pairs into the training data to maximize the conditional probability p_{\theta}(v_{i}|k_{i}) . If a downstream model \theta^{\prime} exhibits significantly higher p_{\theta^{\prime}}(v_{i}|k_{i}) than a random guessing baseline (or lower perplexity; e.g., perplexity =16 for random hexadecimal digits), vendors may confidently claim the origin of the downstream model. This method has proven effective in identifying model plagiarism, such as the case of Llama3-V plagiarizing MiniCPM-o v2.6 [ 33 ] , where rare oracle bone inscriptions were used as keys and their corresponding modern Chinese characters as values. However, this approach requires access to vendor-specific keys or prompts, which introduces significant limitations. In practice, it is often impractical for users to obtain or know the internal training data of vendors. Representation-based Methods. Representation-based methods, such as REEF (Representation Encoding Fingerprint) [ 55 ] , HuRef (Human-Readable Fingerprint) [ 53 ] , and Intrinsic Fingerprint [ 19 ] , analyze similarities between internal representations of language models. Similarities can be examined both in terms of weights and activation distributions under the same input. An ideal similarity measure should remain robust to simple attacks, such as dimension permutation, scaling, or orthogonal transformations. Moreover, this measure should remain stable throughout training, even after continual pretraining with trillions of tokens. To the best of our knowledge, these methods primarily focus on proving similarity through "fingerprints" but lack the ability to identify, trace, or reconstruct the specific process of model plagiarism. Our proposed method overcomes these limitations by operating at the matrix level, directly computing similarity without requiring access to vendor-specific data. Furthermore, we provide deeper insights by identifying exact relationships, such as permutations and transformations, during model copying. We also introduce statistical significance via p -values, a critical component for making robust statistical inferences that is absent in prior research. Our advancements address a key gap in current methodologies, offering both efficiency and interpretability in detecting model similarity. Our method is grounded in a strong mathematical foundation, leveraging Singular Value Decomposition (SVD) and polar decomposition to analyze model matrices. We further utilize Large Deviation Theory (LDT) and random matrix theory to estimate p -values. These mathematical tools provide a rigorous framework for detecting similarities at a fundamental level, making it exceedingly difficult for cheaters to bypass our detection. Cheaters attempting to copy models would face significant challenges in obfuscating this watermark. Our method also democratizes the model verification process. Without requiring access to vendor-specific prompts or specialized hardware, anyone with a standard PC can participate in the verification process. 2 Preliminaries Our work primarily builds upon the foundation of matrix analysis. To avoid excessive technicality or verbosity, we only introduce the key tools and properties used in this paper, potentially omitting or abbreviating proofs. Interested readers are encouraged to refer to [ 21 ] for a comprehensive overview of matrix analysis. 2.1 Matrix Analysis: Singular Value and Polar Decompositions Singular Value Decomposition. The singular value decomposition (SVD) of a matrix A\in\mathbb{R}^{m\times n} is given by A=USV^{T}, where U\in\mathbb{R}^{m\times m} and V\in\mathbb{R}^{n\times n} are orthogonal matrices, and S\in\mathbb{R}^{m\times n} is a diagonal matrix with non-negative singular values on the diagonal: \sigma_{1}\geq\sigma_{2}\geq\cdots\geq\sigma_{r}>0\quad(r=\mathrm{rank}(A)). We denote \sigma_{i}(A) as the i -th singular value of A . Polar Decomposition. The polar decomposition of A has two common but distinct forms: the left decomposition A=PW and the right decomposition A=WQ , where P=(AA^{T})^{1/2} and Q=(A^{T}A)^{1/2} are symmetric positive semidefinite matrices, and W\in\mathbb{R}^{m\times n} (with orthonormal columns) is shared between both decompositions. We define the orthogonal part of A as \mathrm{Ortho}(A):=W . When A is full-rank, the orthogonal part of A is unique. Connection to SVD. The orthogonal part W in the polar decomposition can be obtained from the SVD: W=UV^{T}, where U and V are derived from the SVD of A . The symmetric factors satisfy P=USU^{T} and Q=VSV^{T} . However, when A is not invertible (with many singular values either exactly zero or close to zero), the computation of W=UV^{T} becomes ill-conditioned. SVD and Spectral Calculus. Assume A\in\mathbb{R}^{m\times n} has SVD A=USV^{T} . Then: AA^{T}=U(SS^{T})U^{T},\quad A^{T}A=V(S^{T}S)V^{T}. For any polynomial function f\in\mathbb{R}[x] and g(x)=xf(x^{2}) , we have: \displaystyle f(AA^{T})A \displaystyle=Uf(SS^{T})U^{T}USV^{T}=Uf(SS^{T})SV^{T} \displaystyle=Ug(S)V^{T} \displaystyle=USf(S^{T}S)V^{T}=USV^{T}Vf(S^{T}S)V^{T} \displaystyle=Af(A^{T}A). In fact, g(S) can represent any odd polynomial. For any odd function G , we may find a polynomial f such that f(x) agrees with G(x)/x on all nonzero diagonal entries of SS^{T} . In particular, setting G(x)=\mathrm{sign}(x) , this spectral calculus yields the most “sensible” orthogonal part of A . We adopt the methods in [ 3 ] for a fast and relatively accurate implementation. Trace Maximization Property. The orthogonal part W solves the optimization problem: \max_{\{X|X^{T}X=I\}}\left(\Tr(AX^{T})\right). The maximum value is the sum of all the singular values of A , i.e., \Tr(P)=\sum_{i=1}^{r}\sigma_{i} . Orthogonal Invariance of Singular Values. The singular values \{\sigma_{i}\} of A are invariant under both left and right orthogonal transformations: \sigma_{i}(UAV)=\sigma_{i}(A),\quad\forall U,V\text{ orthogonal and }i=1,\cdots,\min(m,n). Thus, any function depending on these singular values remains invariant under orthogonal transformations. Examples include:

Read the complete text