Zhenhua Wang

Zhenhua Wang, Wei Xie, Kai Chen et al.

Large language models (LLMs), such as ChatGPT, have emerged with astonishing capabilities approaching artificial general intelligence. While providing convenience for various societal needs, LLMs have also lowered the cost of generating harmful content. Consequently, LLM developers have deployed semantic-level defenses to recognize and reject prompts that may lead to inappropriate content. Unfortunately, these defenses are not foolproof, and some attackers have crafted "jailbreak" prompts that temporarily hypnotize the LLM into forgetting content defense rules and answering any improper questions. To date, there is no clear explanation of the principles behind these semantic-level attacks and defenses in both industry and academia. This paper investigates the LLM jailbreak problem and proposes an automatic jailbreak method for the first time. We propose the concept of a semantic firewall and provide three technical implementation approaches. Inspired by the attack that penetrates traditional firewalls through reverse tunnels, we introduce a "self-deception" attack that can bypass the semantic firewall by inducing LLM to generate prompts that facilitate jailbreak. We generated a total of 2,520 attack payloads in six languages (English, Russian, French, Spanish, Chinese, and Arabic) across seven virtual scenarios, targeting the three most common types of violations: violence, hate, and pornography. The experiment was conducted on two models, namely the GPT-3.5-Turbo and GPT-4. The success rates on the two models were 86.2% and 67%, while the failure rates were 4.7% and 2.2%, respectively. This highlighted the effectiveness of the proposed attack method. All experimental code and raw data will be released as open-source to inspire future research. We believe that manipulating AI behavior through carefully crafted prompts will become an important research direction in the future.

Xiaohua Wang, Zhenghua Wang, Xuan Gao et al.

Retrieval-augmented generation (RAG) techniques have proven to be effective in integrating up-to-date information, mitigating hallucinations, and enhancing response quality, particularly in specialized domains. While many RAG approaches have been proposed to enhance large language models through query-dependent retrievals, these approaches still suffer from their complex implementation and prolonged response times. Typically, a RAG workflow involves multiple processing steps, each of which can be executed in various ways. Here, we investigate existing RAG approaches and their potential combinations to identify optimal RAG practices. Through extensive experiments, we suggest several strategies for deploying RAG that balance both performance and efficiency. Moreover, we demonstrate that multimodal retrieval techniques can significantly enhance question-answering capabilities about visual inputs and accelerate the generation of multimodal content using a "retrieval as generation" strategy.

Xiaolei Zhang, Yi Shen, Zhenhua Wang

Satellite power system is a complex, highly interconnected hybrid system that exhibit nonlinear and mode switching behaviors. Directed factor graph is an inference model for fault diagnosis using probabilistic reasoning techniques. A novel approach for constructing the directed factor graph structure based on hybrid bond graph model is proposed. The system components status and their fault symptoms are treated as hypothesis and evidences respectively. The cause-effect relations between hypothesis and evidences are identified and concluded though qualitative equations and causal path analysis on hybrid bond graph model. A power supply module of a satellite power system is provided as case study to show the feasibility and validity of the proposed method.

Zhenhua Wang, Ming Ren, Dong Gao et al.

Entity extraction is critical in the intelligent advancement across diverse domains. Nevertheless, a challenge to its effectiveness arises from the data imbalance. This paper proposes a novel approach by viewing the issue through the quantitative information, recognizing that entities exhibit certain levels of commonality while others are scarce, which can be reflected in the quantifiable distribution of words. The Zipf's Law emerges as a well-suited adoption, and to transition from words to entities, words within the documents are classified as common and rare ones. Subsequently, sentences are classified into common and rare ones, and are further processed by text generation models accordingly. Rare entities within the generated sentences are then labeled using human-designed rules, serving as a supplement to the raw dataset, thereby mitigating the imbalance problem. The study presents a case of extracting entities from technical documents, and experimental results from two datasets prove the effectiveness of the proposed method. Furthermore, the significance of Zipf's law in driving the progress of AI is discussed, broadening the reach and coverage of Informetrics. This paper presents a successful demonstration of extending Informetrics to interface with AI through Zipf's Law.

Zhenhua Wang, Bin Wang, Ming Ren et al.

Hazard and operability analysis (HAZOP) is the paradigm of industrial safety that can reveal the hazards of process from its node deviations, consequences, causes, measures and suggestions, and such hazards can be considered as hazard events (HaE). The classification research on HaE has much irreplaceable pragmatic values. In this paper, we present a novel deep learning model termed DLF through multifractal to explore HaE classification where the motivation is that HaE can be naturally regarded as a kind of time series. Specifically, first HaE is vectorized to get HaE time series by employing BERT. Then, a new multifractal analysis method termed HmF-DFA is proposed to win HaE fractal series by analyzing HaE time series. Finally, a new hierarchical gating neural network (HGNN) is designed to process HaE fractal series to accomplish the classification of HaE from three aspects: severity, possibility and risk. We take HAZOP reports of 18 processes as cases, and launch the experiments on this basis. Results demonstrate that compared with other classifiers, DLF classifier performs better under metrics of precision, recall and F1-score, especially for the severity aspect. Also, HmF-DFA and HGNN effectively promote HaE classification. Our HaE classification system can serve application incentives to experts, engineers, employees, and other enterprises. We hope our research can contribute added support to the daily practice in industrial safety.

Qi Qian, Chengsong Huang, Jingwen Xu et al.

The rapid proliferation of benchmarks for evaluating large language models (LLMs) has created an urgent need for systematic methods to assess benchmark quality itself. We propose Benchmark^2, a comprehensive framework comprising three complementary metrics: (1) Cross-Benchmark Ranking Consistency, measuring whether a benchmark produces model rankings aligned with peer benchmarks; (2) Discriminability Score, quantifying a benchmark's ability to differentiate between models; and (3) Capability Alignment Deviation, identifying problematic instances where stronger models fail but weaker models succeed within the same model family. We conduct extensive experiments across 15 benchmarks spanning mathematics, reasoning, and knowledge domains, evaluating 11 LLMs across four model families. Our analysis reveals significant quality variations among existing benchmarks and demonstrates that selective benchmark construction based on our metrics can achieve comparable evaluation performance with substantially reduced test sets.

Feiran Zhang, Yixin Wu, Zhenghua Wang et al.

Vision-Language Models (VLMs) have demonstrated remarkable progress in multimodal tasks, but remain susceptible to hallucinations, where generated text deviates from the underlying visual content. Existing hallucination detection methods primarily rely on output logits or external verification tools, often overlooking their internal mechanisms. In this work, we investigate the outputs of internal attention heads, postulating that specific heads carry the primary signals for truthful generation.However, directly probing these high-dimensional states is challenging due to the entanglement of visual-linguistic syntax and noise. To address this, we propose VIB-Probe, a novel hallucination detection and mitigation framework leveraging the Variational Information Bottleneck (VIB) theory. Our method extracts discriminative patterns across layers and heads while filtering out semantic nuisances through the information bottleneck principle. Furthermore, by leveraging the gradients of our VIB probe, we identify attention heads with strong causal influence on hallucinations and introduce an inference-time intervention strategy for hallucination mitigation. Extensive experiments across diverse benchmarks demonstrate that VIB-Probe significantly outperforms existing baselines in both settings. Our code will be made publicly available.

Zhenhua Wang, Ming Ren, Dong Gao et al.

Hazards can be exposed by HAZOP as text information, and studying their classification is of great significance to the development of industrial informatics, which is conducive to safety early warning, decision support, policy evaluation, etc. However, there is no research on this important field at present. In this paper, we propose a novel model termed DLGM via deep learning for hazard classification. Specifically, first, we leverage BERT to vectorize the hazard and treat it as a type of time series (HTS). Secondly, we build a grey model FSGM(1, 1) to model it, and get the grey guidance in the sense of the structural parameters. Finally, we design a hierarchical-feature fusion neural network (HFFNN) to investigate the HTS with grey guidance (HTSGG) from three themes, where, HFFNN is a hierarchical structure with four types of modules: two feature encoders, a gating mechanism, and a deepening mechanism. We take 18 industrial processes as application cases and launch a series of experiments. The experimental results prove that DLGM has promising aptitudes for hazard classification and that FSGM(1, 1) and HFFNN are effective. We hope our research can contribute added value and support to the daily practice in industrial safety.

Muling Wu, Qi Qian, Wenhao Liu et al.

Large Language Models (LLMs) have achieved remarkable performance across various reasoning tasks, yet post-training is constrained by inefficient sample utilization and inflexible difficulty samples processing. To address these limitations, we propose Customized Curriculum Learning (CCL), a novel framework with two key innovations. First, we introduce model-adaptive difficulty definition that customizes curriculum datasets based on each model's individual capabilities rather than using predefined difficulty metrics. Second, we develop "Guided Prompting," which dynamically reduces sample difficulty through strategic hints, enabling effective utilization of challenging samples that would otherwise degrade performance. Comprehensive experiments on supervised fine-tuning and reinforcement learning demonstrate that CCL significantly outperforms uniform training approaches across five mathematical reasoning benchmarks, confirming its effectiveness across both paradigms in enhancing sample utilization and model performance.

Zhenghua Wang, Yiran Ding, Changze Lv et al.

Although large language models (LLMs) have achieved significant progress in handling long-context inputs, they still suffer from the ``lost-in-the-middle'' problem, where crucial information in the middle of the context is often underrepresented or lost. Our extensive experiments reveal that this issue may arise from the rapid long-term decay in Rotary Position Embedding (RoPE). To address this problem, we propose a layer-specific positional encoding scaling method that assigns distinct scaling factors to each layer, slowing down the decay rate caused by RoPE to make the model pay more attention to the middle context. A specially designed genetic algorithm is employed to efficiently select the optimal scaling factors for each layer by incorporating Bezier curves to reduce the search space. Through comprehensive experimentation, we demonstrate that our method significantly alleviates the ``lost-in-the-middle'' problem. Our approach results in an average accuracy improvement of up to 20% on the Key-Value Retrieval dataset. Furthermore, we show that layer-specific interpolation, as opposed to uniform interpolation across all layers, enhances the model's extrapolation capabilities when combined with PI and Dynamic-NTK positional encoding schemes.

Zhenhua Wang, Daiyu Zhang, Ming Ren et al.

Textual geographic information is indispensable and heavily relied upon in practical applications. The absence of clear distribution poses challenges in effectively harnessing geographic information, thereby driving our quest for exploration. We contend that geographic information is influenced by human behavior, cognition, expression, and thought processes, and given our intuitive understanding of natural systems, we hypothesize its conformity to the Gamma distribution. Through rigorous experiments on a diverse range of 24 datasets encompassing different languages and types, we have substantiated this hypothesis, unearthing the underlying regularities governing the dimensions of quantity, length, and distance in geographic information. Furthermore, theoretical analyses and comparisons with Gaussian distributions and Zipf's law have refuted the contingency of these laws. Significantly, we have estimated the upper bounds of human utilization of geographic information, pointing towards the existence of uncharted territories. Also, we provide guidance in geographic information extraction. Hope we peer its true countenance uncovering the veil of geographic information.

Zhenhua Wang, Ming Ren, Dong Gao

In this world full of uncertainty, where the fabric of existence weaves patterns of complexity, multifractal emerges as beacons of insight, illuminating them. As we delve into the realm of text mining that underpins various natural language processing applications and powers a range of intelligent services, we recognize that behind the veil of text lies a manifestation of human thought and cognition, intricately intertwined with the complexities. Building upon the foundation of perceiving text as a complex system, this study embarks on a journey to unravel the hidden treasures within, armed with the proposed multifractal method that deciphers the multifractal attributes embedded within the text landscape. This endeavor culminates in the birth of our novel model, which also harnesses the power of the proposed activation function to facilitate nonlinear information transmission within its neural network architecture. The success on experiments anchored in real-world technical reports covering the extraction of technical term and classification of hazard events, stands as a testament to our endeavors. This research venture not only expands our understanding of text mining but also opens new horizons for knowledge discovery across various domains.

Zhenhua Wang, Dong Gao, Haozhe Liu et al.

Deep learning oriented named entity recognition (DNER) has gradually become the paradigm of knowledge discovery, which greatly promotes domain intelligence. However, the current activation function of DNER fails to treat gradient vanishing, no negative output or non-differentiable existence, which may impede knowledge exploration caused by the omission and incomplete representation of latent semantics. To break through the dilemma, we present a novel activation function termed KDAC. Detailly, KDAC is an aggregation function with multiple conversion modes. The backbone of the activation region is the interaction between exponent and linearity, and the both ends extend through adaptive linear divergence, which surmounts the obstacle of gradient vanishing and no negative output. Crucially, the non-differentiable points are alerted and eliminated by an approximate smoothing algorithm. KDAC has a series of brilliant properties, including nonlinear, stable near-linear transformation and derivative, as well as dynamic style, etc. We perform experiments based on BERT-BiLSTM-CNN-CRF model on six benchmark datasets containing different domain knowledge, such as Weibo, Clinical, E-commerce, Resume, HAZOP and People's daily. The evaluation results show that KDAC is advanced and effective, and can provide more generalized activation to stimulate the performance of DNER. We hope that KDAC can be exploited as a promising activation function to devote itself to the construction of knowledge.