arXiv subjects

Linghui Li

Publications and source records attributed to Linghui Li.

6 recordsLinked to original sources

arXiv2026

Robust Harmful Features Under Jailbreak Attacks: Mechanistic Evidence from Attention Head Specialization in Large Language Models

Jailbreak attacks bypass LLM safety alignment, yet their mechanisms remain poorly understood. We provide evidence that attacks do not comprehensively eliminate safety features, but instead selectively suppress specific attention heads. We identify two functionally differentiated types: Adversarially Compromised Heads (ACHs) concentrated in early layers, which are suppressed under attacks, and Safety-Aligned Heads (SAHs) in mid-layers, which maintain robust activations even when attacks succeed. Ablation studies support the causal role of ACHs and the contribution of SAHs to robust activations: suppressing a small number of ACHs is sufficient to induce jailbreak-like behavior on normally refused inputs, while removing SAHs substantially weakens mid-layer safety activations. Token-level attribution further shows that ACH suppression is driven specifically by attack-template tokens, providing a mechanistic account of why attacks can bypass refusal decisions through ACH suppression while leaving internal safety signals sustained by SAHs -- a phenomenon we term Robust Harmful Features. To validate the practical significance of this robustness, we show that simply reading these persistent activations -- without any training -- yields competitive aggregate detection performance with strong adversarial robustness.

cs.CR↗

arXiv2026

RefiningGPT: Specialized language Models for Automated Refinery Unit-level Process Diagram Synthesis

Applying LLMs to complex industrial processes remains challenging due to the semantic gap between natural language design intents and the rigorous physical logic of engineering. In the field of petroleum refining engineering, a critical bottleneck is the automated synthesis of Unit-level Process Diagrams (UPDs), which serve as the topological bridge connecting abstract requirements to concrete unit operations. In this paper, we propose RefineGPT, a domain-specialized agent for autonomous refinery design.RefineGPT adopts a hierarchical architecture in which a supervised fine-tuned small language model is responsible for selecting units that satisfy design requirements, while a large language model is used to connect these units to generate the final topology. To enable supervised training, we develop a pipeline that extracts latent process motifs from noisy, unstructured legacy topologies and synthesizes high-quality rationale-based Chain-of-Thought (CoT) training data. Empirical validation demonstrates that RefineGPT achieves substantial improvements in topological consistency and chemical engineering feasibility, establishing a high-fidelity pathway for AI-augmented industrial process synthesis.

cs.CE↗

arXiv2025

Towards Principled Learning for Re-ranking in Recommender Systems

As the final stage of recommender systems, re-ranking presents ordered item lists to users that best match their interests. It plays such a critical role and has become a trending research topic with much attention from both academia and industry. Recent advances of re-ranking are focused on attentive listwise modeling of interactions and mutual influences among items to be re-ranked. However, principles to guide the learning process of a re-ranker, and to measure the quality of the output of the re-ranker, have been always missing. In this paper, we study such principles to learn a good re-ranker. Two principles are proposed, including convergence consistency and adversarial consistency. These two principles can be applied in the learning of a generic re-ranker and improve its performance. We validate such a finding by various baseline methods over different datasets.

cs.IR↗

arXiv2024

Generalize Your Face Forgery Detectors: An Insertable Adaptation Module Is All You Need

A plethora of face forgery detectors exist to tackle facial deepfake risks. However, their practical application is hindered by the challenge of generalizing to forgeries unseen during the training stage. To this end, we introduce an insertable adaptation module that can adapt a trained off-the-shelf detector using only online unlabeled test data, without requiring modifications to the architecture or training process. Specifically, we first present a learnable class prototype-based classifier that generates predictions from the revised features and prototypes, enabling effective handling of various forgery clues and domain gaps during online testing. Additionally, we propose a nearest feature calibrator to further improve prediction accuracy and reduce the impact of noisy pseudo-labels during self-training. Experiments across multiple datasets show that our module achieves superior generalization compared to state-of-the-art methods. Moreover, it functions as a plug-and-play component that can be combined with various detectors to enhance the overall performance.

cs.CV↗

arXiv2024

Boosting the Targeted Transferability of Adversarial Examples via Salient Region & Weighted Feature Drop

Deep neural networks can be vulnerable to adversarially crafted examples, presenting significant risks to practical applications. A prevalent approach for adversarial attacks relies on the transferability of adversarial examples, which are generated from a substitute model and leveraged to attack unknown black-box models. Despite various proposals aimed at improving transferability, the success of these attacks in targeted black-box scenarios is often hindered by the tendency for adversarial examples to overfit to the surrogate models. In this paper, we introduce a novel framework based on Salient region & Weighted Feature Drop (SWFD) designed to enhance the targeted transferability of adversarial examples. Drawing from the observation that examples with higher transferability exhibit smoother distributions in the deep-layer outputs, we propose the weighted feature drop mechanism to modulate activation values according to weights scaled by norm distribution, effectively addressing the overfitting issue when generating adversarial examples. Additionally, by leveraging salient region within the image to construct auxiliary images, our method enables the adversarial example's features to be transferred to the target category in a model-agnostic manner, thereby enhancing the transferability. Comprehensive experiments confirm that our approach outperforms state-of-the-art methods across diverse configurations. On average, the proposed SWFD raises the attack success rate for normally trained models and robust models by 16.31% and 7.06% respectively.

cs.IR↗

arXiv2019

On the parameter estimation of ARMA(p,q) model by approximate Bayesian computation

In this paper, the parameter estimation of ARMA(p,q) model is given by approximate Bayesian computation algorithm. In order to improve the sampling efficiency of the algorithm, approximate Bayesian computation should select as many statistics as possible with parameter information in low dimension. Firstly, we use the autocorrelation coefficient of the first p+q order sample as the statistic and obtain an approximate Bayesian estimation of the AR coefficient, transforming the ARMA(p,q) model into the MA(q) model. Considering the first q order sample autocorrelation functions and sample variance as the statistics, the approximate Bayesian estimation of MA coefficient and white noise variances can be given. The method mentioned above is more accurate and powerful than the maximum likelihood estimation, which is verified by the numerical simulations and experiment study.

stat.CO↗