arXiv subjects

Victor S. Sheng

Publications and source records attributed to Victor S. Sheng.

At least 19 recordsLinked to original sources

arXiv2026

A Behavioral Fingerprint for Large Language Models: Provenance Tracking via Refusal Vectors

Protecting the intellectual property of large language models (LLMs) is a critical challenge due to the proliferation of unauthorized derivative models. We introduce a novel fingerprinting framework that leverages the behavioral patterns induced by safety alignment, applying the concept of refusal vectors for LLM provenance tracking. These vectors, extracted from directional patterns in a model's internal representations when processing harmful versus harmless prompts, serve as robust behavioral fingerprints. Our contribution lies in developing a fingerprinting system around this concept and conducting extensive validation of its effectiveness for IP protection. We demonstrate that these behavioral fingerprints are highly robust against common modifications, including finetunes, merges, and quantization. Our experiments show that the fingerprint is unique to each model family, with low cosine similarity between independently trained models. In a large-scale identification task across 76 offspring models, our method achieves 100\% accuracy in identifying the correct base model family. Furthermore, we analyze the fingerprint's behavior under alignment-breaking attacks, finding that while performance degrades significantly, detectable traces remain. Finally, we propose a theoretical framework to transform this private fingerprint into a publicly verifiable, privacy-preserving artifact using locality-sensitive hashing and zero-knowledge proofs.

cs.CR↗

arXiv2026

Controlling Chat Style in Language Models via Single-Direction Editing

Controlling stylistic attributes in large language models (LLMs) remains challenging, with existing approaches relying on either prompt engineering or post-training alignment. This paper investigates this challenge through the lens of representation engineering, testing the hypothesis that distinct stylistic attributes - from emotional tone to linguistic structure - are encoded as linear directions in the model's activation space. We provide strong empirical evidence for this hypothesis across a wide range of styles and, based on this finding, present a lightweight, training-free method for precise style control. Our approach supports linear style composition, enhances safety by ablating undesirable behaviors, and, as confirmed by experiments on over a dozen models, achieves high style adherence while preserving core capabilities at minimal computational cost.

cs.CL↗

arXiv2025

Quantum parallel information exchange (QPIE) hybrid network with transfer learning

Quantum machine learning (QML) has emerged as an innovative framework with the potential to uncover complex patterns by leveraging quantum systems ability to simulate and exploit high-dimensional latent spaces, particularly in learning tasks. Quantum neural network (QNN) frameworks are inherently sensitive to the precision of gradient calculations and the computational limitations of current quantum hardware as unitary rotations introduce overhead from complex number computations, and the quantum gate operation speed remains a bottleneck for practical implementations. In this study, we introduce quantum parallel information exchange (QPIE) hybrid network, a new non-sequential hybrid classical quantum model architecture, leveraging quantum transfer learning by feeding pre-trained parameters from classical neural networks into quantum circuits, which enables efficient pattern recognition and temporal series data prediction by utilizing non-clifford parameterized quantum gates thereby enhancing both learning efficiency and representational capacity. Additionally, we develop a dynamic gradient selection method that applies the parameter shift rule on quantum processing units (QPUs) and adjoint differentiation on GPUs. Our results demonstrate model performance exhibiting higher accuracy in ad-hoc benchmarks, lowering approximately 88% convergence rate for extra stochasticity time-series data within 100-steps, and showcasing a more unbaised eigenvalue spectrum of the fisher information matrix on CPU/GPU and IonQ QPU simulators.

quant-ph↗

arXiv2025

The Backpropagation of the Wave Network

This paper provides an in-depth analysis of Wave Network, a novel token representation method derived from the Wave Network, designed to capture both global and local semantics of input text through wave-inspired complex vectors. In complex vector token representation, each token is represented with a magnitude component, capturing the global semantics of the entire input text, and a phase component, encoding the relationships between individual tokens and the global semantics. Building on prior research that demonstrated the effectiveness of wave-like operations, such as interference and modulation, during forward propagation, this study investigates the convergence behavior, backpropagation characteristics, and embedding independence within the Token2Wave framework. A detailed computational complexity analysis shows that Token2Wave can significantly reduce video memory usage and training time compared to BERT. Gradient comparisons for the [CLS] token, total input text, and classifier parameters further highlight Token2Wave's unique characteristics. This research offers new insights into wave-based token representations, demonstrating their potential to enable efficient and computationally friendly language model architectures.

cs.CL↗

arXiv2025

Multi-Task Program Error Repair and Explanatory Diagnosis

Program errors can occur in any type of programming, and can manifest in a variety of ways, such as unexpected output, crashes, or performance issues. And program error diagnosis can often be too abstract or technical for developers to understand, especially for beginners. The goal of this paper is to present a novel machine-learning approach for Multi-task Program Error Repair and Explanatory Diagnosis (mPRED). A pre-trained language model is used to encode the source code, and a downstream model is specifically designed to identify and repair errors. Programs and test cases will be augmented and optimized from several perspectives. Additionally, our approach incorporates a "chain of thoughts" method, which enables the models to produce intermediate reasoning explanations before providing the final correction. To aid in visualizing and analyzing the program structure, we use a graph neural network for program structure visualization. Overall, our approach offers a promising approach for repairing program errors across different programming languages and providing helpful explanations to programmers.

cs.SE↗

arXiv2025

CodeVision: Detecting LLM-Generated Code Using 2D Token Probability Maps and Vision Models

The rise of large language models (LLMs) like ChatGPT has significantly improved automated code generation, enhancing software development efficiency. However, this introduces challenges in academia, particularly in distinguishing between human-written and LLM-generated code, which complicates issues of academic integrity. Existing detection methods, such as pre-trained models and watermarking, face limitations in adaptability and computational efficiency. In this paper, we propose a novel detection method using 2D token probability maps combined with vision models, preserving spatial code structures such as indentation and brackets. By transforming code into log probability matrices and applying vision models like Vision Transformers (ViT) and ResNet, we capture both content and structure for more accurate detection. Our method shows robustness across multiple programming languages and improves upon traditional detectors, offering a scalable and computationally efficient solution for identifying LLM-generated code.

cs.SE↗

arXiv2024

Affiliation-based Local Community Detection across Multiple Networks

Real-world networks are often constructed from different sources or domains, including various types of entities and diverse relationships between networks, thus forming multi-domain networks. A single network typically fails to capture the complete graph structure and the diverse relationships among multiple networks. Consequently, leveraging multiple networks is crucial for a comprehensive detection of community structures. Most existing local community detection methods discover community structures by integrating information from different views on multi-view networks. However, methods designed for multi-view networks are not suitable for multi-domain networks. Therefore, to mine communities from multiple networks, we propose a Local Algorithm for Multiple networks with node Affiliation, called LAMA, which is suitable for both multi-view and multi-domain networks. The core idea of LAMA is to optimize node affiliations by maximizing the quality of communities within each network while ensuring consistency in community structures across multiple networks. The algorithm iteratively optimizes node affiliations and expands the community outward based on affiliations to detect the community containing the seed node. Experimental results show that LAMA outperforms comparison algorithms on two synthetic datasets and five real datasets.

cs.SI↗

arXiv2024

Wave Network: An Ultra-Small Language Model

We propose an innovative token representation and update method in a new ultra-small language model: the Wave network. Specifically, we use a complex vector to represent each token, encoding both global and local semantics of the input text. A complex vector consists of two components: a magnitude vector representing the global semantics of the input text, and a phase vector capturing the relationships between individual tokens and global semantics. Experiments on the AG News text classification task demonstrate that, when generating complex vectors from randomly initialized token embeddings, our single-layer Wave Network achieves 90.91% accuracy with wave interference and 91.66% with wave modulation - outperforming a single Transformer layer using BERT pre-trained embeddings by 19.23% and 19.98%, respectively, and approaching the accuracy of the pre-trained and fine-tuned BERT base model (94.64%). Additionally, compared to BERT base, the Wave Network reduces video memory usage and training time by 77.34% and 85.62% during wave modulation. In summary, we used a 2.4-million-parameter small language model to achieve accuracy comparable to a 100-million-parameter BERT model in text classification.

cs.CL↗

arXiv2024

Neuro-Symbolic AI: Explainability, Challenges, and Future Trends

Explainability is an essential reason limiting the application of neural networks in many vital fields. Although neuro-symbolic AI hopes to enhance the overall explainability by leveraging the transparency of symbolic learning, the results are less evident than imagined. This article proposes a classification for explainability by considering both model design and behavior of 191 studies from 2013, focusing on neuro-symbolic AI, hoping to inspire scholars who want to understand the explainability of neuro-symbolic AI. Precisely, we classify them into five categories by considering whether the form of bridging the representation differences is readable as their design factor, if there are representation differences between neural networks and symbolic logic learning, and whether a model decision or prediction process is understandable as their behavior factor: implicit intermediate representations and implicit prediction, partially explicit intermediate representations and partially explicit prediction, explicit intermediate representations or explicit prediction, explicit intermediate representation and explicit prediction, unified representation and explicit prediction. We also analyzed the research trends and three significant challenges: unified representations, explainability and transparency, and sufficient cooperation from neural networks and symbolic learning. Finally, we put forward suggestions for future research in three aspects: unified representations, enhancing model explainability, ethical considerations, and social impact.

cs.AI↗

arXiv2024

Bridging the Gap: Representation Spaces in Neuro-Symbolic AI

Neuro-symbolic AI is an effective method for improving the overall performance of AI models by combining the advantages of neural networks and symbolic learning. However, there are differences between the two in terms of how they process data, primarily because they often use different data representation methods, which is often an important factor limiting the overall performance of the two. From this perspective, we analyzed 191 studies from 2013 by constructing a four-level classification framework. The first level defines five types of representation spaces, and the second level focuses on five types of information modalities that the representation space can represent. Then, the third level describes four symbolic logic methods. Finally, the fourth-level categories propose three collaboration strategies between neural networks and symbolic learning. Furthermore, we conducted a detailed analysis of 46 research based on their representation space.

cs.AI↗

arXiv2024

Logic Error Localization in Student Programming Assignments Using Pseudocode and Graph Neural Networks

Pseudocode is extensively used in introductory programming courses to instruct computer science students in algorithm design, utilizing natural language to define algorithmic behaviors. This learning approach enables students to convert pseudocode into source code and execute it to verify their algorithms' correctness. This process typically introduces two types of errors: syntax errors and logic errors. Syntax errors are often accompanied by compiler feedback, which helps students identify incorrect lines. In contrast, logic errors are more challenging because they do not trigger compiler errors and lack immediate diagnostic feedback, making them harder to detect and correct. To address this challenge, we developed a system designed to localize logic errors within student programming assignments at the line level. Our approach utilizes pseudocode as a scaffold to build a code-pseudocode graph, connecting symbols from the source code to their pseudocode counterparts. We then employ a graph neural network to both localize and suggest corrections for logic errors. Additionally, we have devised a method to efficiently gather logic-error-prone programs during the syntax error correction process and compile these into a dataset that includes single and multiple line logic errors, complete with indices of the erroneous lines. Our experimental results are promising, demonstrating a localization accuracy of 99.2% for logic errors within the top-10 suspected lines, highlighting the effectiveness of our approach in enhancing students' coding proficiency and error correction skills.

cs.CY↗

arXiv2024

LecPrompt: A Prompt-based Approach for Logical Error Correction with CodeBERT

Logical errors in programming don't raise compiler alerts, making them hard to detect. These silent errors can disrupt a program's function or cause run-time issues. Their correction requires deep insight into the program's logic, highlighting the importance of automated detection and repair. In this paper, we introduce LecPrompt to localize and repair logical errors, an prompt-based approach that harnesses the capabilities of CodeBERT, a transformer-based large language model trained on code. First, LecPrompt leverages a large language model to calculate perplexity and log probability metrics, pinpointing logical errors at both token and line levels. Through statistical analysis, it identifies tokens and lines that deviate significantly from the expected patterns recognized by large language models, marking them as potential error sources. Second, by framing the logical error correction challenge as a Masked Language Modeling (MLM) task, LecPrompt employs CodeBERT to autoregressively repair the identified error tokens. Finally, the soft-prompt method provides a novel solution in low-cost scenarios, ensuring that the model can be fine-tuned to the specific nuances of the logical error correction task without incurring high computational costs. To evaluate LecPrompt's performance, we created a method to introduce logical errors into correct code and applying this on QuixBugs to produce the QuixBugs-LE dataset. Our evaluations on the QuixBugs-LE dataset for both Python and Java highlight the impressive capabilities of our method, LecPrompt. For Python, LecPrompt achieves a noteworthy 74.58% top-1 token-level repair accuracy and 27.4% program-level repair accuracy. In Java, LecPrompt delivers a 69.23\% top-1 token-level repair accuracy and 24.7% full program-level repair accuracy.

cs.SE↗

arXiv2024

Signal Watermark on Large Language Models

As Large Language Models (LLMs) become increasingly sophisticated, they raise significant security concerns, including the creation of fake news and academic misuse. Most detectors for identifying model-generated text are limited by their reliance on variance in perplexity and burstiness, and they require substantial computational resources. In this paper, we proposed a watermarking method embedding a specific watermark into the text during its generation by LLMs, based on a pre-defined signal pattern. This technique not only ensures the watermark's invisibility to humans but also maintains the quality and grammatical integrity of model-generated text. We utilize LLMs and Fast Fourier Transform (FFT) for token probability computation and detection of the signal watermark. The unique application of signal processing principles within the realm of text generation by LLMs allows for subtle yet effective embedding of watermarks, which do not compromise the quality or coherence of the generated text. Our method has been empirically validated across multiple LLMs, consistently maintaining high detection accuracy, even with variations in temperature settings during text generation. In the experiment of distinguishing between human-written and watermarked text, our method achieved an AUROC score of 0.97, significantly outperforming existing methods like GPTZero, which scored 0.64. The watermark's resilience to various attacking scenarios further confirms its robustness, addressing significant challenges in model-generated text authentication.

cs.CR↗

arXiv2024

FreqMark: Frequency-Based Watermark for Sentence-Level Detection of LLM-Generated Text

The increasing use of Large Language Models (LLMs) for generating highly coherent and contextually relevant text introduces new risks, including misuse for unethical purposes such as disinformation or academic dishonesty. To address these challenges, we propose FreqMark, a novel watermarking technique that embeds detectable frequency-based watermarks in LLM-generated text during the token sampling process. The method leverages periodic signals to guide token selection, creating a watermark that can be detected with Short-Time Fourier Transform (STFT) analysis. This approach enables accurate identification of LLM-generated content, even in mixed-text scenarios with both human-authored and LLM-generated segments. Our experiments demonstrate the robustness and precision of FreqMark, showing strong detection capabilities against various attack scenarios such as paraphrasing and token substitution. Results show that FreqMark achieves an AUC improvement of up to 0.98, significantly outperforming existing detection methods.

cs.CL↗

arXiv2024

Federated Learning of Large Language Models with Parameter-Efficient Prompt Tuning and Adaptive Optimization

Federated learning (FL) is a promising paradigm to enable collaborative model training with decentralized data. However, the training process of Large Language Models (LLMs) generally incurs the update of significant parameters, which limits the applicability of FL techniques to tackle the LLMs in real scenarios. Prompt tuning can significantly reduce the number of parameters to update, but it either incurs performance degradation or low training efficiency. The straightforward utilization of prompt tuning in the FL often raises non-trivial communication costs and dramatically degrades performance. In addition, the decentralized data is generally non-Independent and Identically Distributed (non-IID), which brings client drift problems and thus poor performance. This paper proposes a Parameter-efficient prompt Tuning approach with Adaptive Optimization, i.e., FedPepTAO, to enable efficient and effective FL of LLMs. First, an efficient partial prompt tuning approach is proposed to improve performance and efficiency simultaneously. Second, a novel adaptive optimization method is developed to address the client drift problems on both the device and server sides to enhance performance further. Extensive experiments based on 10 datasets demonstrate the superb performance (up to 60.8\% in terms of accuracy) and efficiency (up to 97.59\% in terms of training time) of FedPepTAO compared with 9 baseline approaches. Our code is available at https://github.com/llm-eff/FedPepTAO.

cs.LG↗

arXiv2023

Intent Contrastive Learning with Cross Subsequences for Sequential Recommendation

The user purchase behaviors are mainly influenced by their intentions (e.g., buying clothes for decoration, buying brushes for painting, etc.). Modeling a user's latent intention can significantly improve the performance of recommendations. Previous works model users' intentions by considering the predefined label in auxiliary information or introducing stochastic data augmentation to learn purposes in the latent space. However, the auxiliary information is sparse and not always available for recommender systems, and introducing stochastic data augmentation may introduce noise and thus change the intentions hidden in the sequence. Therefore, leveraging user intentions for sequential recommendation (SR) can be challenging because they are frequently varied and unobserved. In this paper, Intent contrastive learning with Cross Subsequences for sequential Recommendation (ICSRec) is proposed to model users' latent intentions. Specifically, ICSRec first segments a user's sequential behaviors into multiple subsequences by using a dynamic sliding operation and takes these subsequences into the encoder to generate the representations for the user's intentions. To tackle the problem of no explicit labels for purposes, ICSRec assumes different subsequences with the same target item may represent the same intention and proposes a coarse-grain intent contrastive learning to push these subsequences closer. Then, fine-grain intent contrastive learning is mentioned to capture the fine-grain intentions of subsequences in sequential behaviors. Extensive experiments conducted on four real-world datasets demonstrate the superior performance of the proposed ICSRec model compared with baseline methods.

cs.IR↗

arXiv2023

Meta-optimized Joint Generative and Contrastive Learning for Sequential Recommendation

Sequential Recommendation (SR) has received increasing attention due to its ability to capture user dynamic preferences. Recently, Contrastive Learning (CL) provides an effective approach for sequential recommendation by learning invariance from different views of an input. However, most existing data or model augmentation methods may destroy semantic sequential interaction characteristics and often rely on the hand-crafted property of their contrastive view-generation strategies. In this paper, we propose a Meta-optimized Seq2Seq Generator and Contrastive Learning (Meta-SGCL) for sequential recommendation, which applies the meta-optimized two-step training strategy to adaptive generate contrastive views. Specifically, Meta-SGCL first introduces a simple yet effective augmentation method called Sequence-to-Sequence (Seq2Seq) generator, which treats the Variational AutoEncoders (VAE) as the view generator and can constitute contrastive views while preserving the original sequence's semantics. Next, the model employs a meta-optimized two-step training strategy, which aims to adaptively generate contrastive views without relying on manually designed view-generation techniques. Finally, we evaluate our proposed method Meta-SGCL using three public real-world datasets. Compared with the state-of-the-art methods, our experimental results demonstrate the effectiveness of our model and the code is available.

cs.IR↗

arXiv2023

Frequency Enhanced Hybrid Attention Network for Sequential Recommendation

The self-attention mechanism, which equips with a strong capability of modeling long-range dependencies, is one of the extensively used techniques in the sequential recommendation field. However, many recent studies represent that current self-attention based models are low-pass filters and are inadequate to capture high-frequency information. Furthermore, since the items in the user behaviors are intertwined with each other, these models are incomplete to distinguish the inherent periodicity obscured in the time domain. In this work, we shift the perspective to the frequency domain, and propose a novel Frequency Enhanced Hybrid Attention Network for Sequential Recommendation, namely FEARec. In this model, we firstly improve the original time domain self-attention in the frequency domain with a ramp structure to make both low-frequency and high-frequency information could be explicitly learned in our approach. Moreover, we additionally design a similar attention mechanism via auto-correlation in the frequency domain to capture the periodic characteristics and fuse the time and frequency level attention in a union model. Finally, both contrastive learning and frequency regularization are utilized to ensure that multiple views are aligned in both the time domain and frequency domain. Extensive experiments conducted on four widely used benchmark datasets demonstrate that the proposed model performs significantly better than the state-of-the-art approaches.

cs.IR↗