Xinyue Fang, Zhiliang Tian, Zhen Huang, Ziyi Pan, Zhihua Wen, Xi Wang, Quntian Fang, Dongsheng Li. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

LLMs obtain remarkable performance but suffer from hallucinations. Most research on detecting hallucination focuses on questions with short and concrete correct answers that are easy to check faithfulness. Hallucination detections for text generation with open-ended answers are more hard. Some researchers use external knowledge to detect hallucinations in generated texts, but external resources for specific scenarios are hard to access. Recent studies on detecting hallucinations in long texts without external resources conduct consistency comparison among multiple sampled outputs. To handle long texts, researchers split long texts into multiple facts and individually compare the consistency of each pair of facts. However, these methods (1) hardly achieve alignment among multiple facts; (2) overlook dependencies between multiple contextual facts. In this paper, we propose a graph-based context-aware (GCA) hallucination detection method for text generations, which aligns facts and considers the dependencies between contextual facts in consistency comparison. Particularly, to align multiple facts, we conduct a triple-oriented response segmentation to extract multiple knowledge triples. To model dependencies among contextual triples (facts), we construct contextual triples into a graph and enhance triples’ interactions via message passing and aggregating via RGCN. To avoid the omission of knowledge triples in long texts, we conduct an LLM-based reverse verification by reconstructing the knowledge triples. Experiments show that our model enhances hallucination detection and excels all baselines.

Large language models (LLMs) are now pivotal in real-world applications. Model editing has emerged as a promising paradigm for efficiently modifying LLMs without full retraining. However, current editing approaches face significant limitations due to parameter drift, which stems from inconsistencies between newly edited knowledge and the model's existing knowledge. In sequential editing scenarios, cumulative drifts progressively lead to model collapse characterized by general capability degradation and balance between acquiring new knowledge and catastrophic forgetting of existing knowledge. Drawing inspiration from the hippocampal trisynaptic circuit for continual memorizing and forgetting, we propose a Hippocampal-like Sequential Editing (HSE) framework that designs the unlearning of obsolete knowledge, domain-specific knowledge update separation and replay for edited knowledge. Specifically, the HSE framework designs three core mechanisms: (1) Machine unlearning selectively erases outdated knowledge to facilitate integration of new information, (2) Fisher Information Matrix-guided parameter updates prevents cross-domain knowledge interference, and (3) Parameter replay consolidates long-term editing memory through lightweight and global replay of editing data in a parametric form. Theoretical analysis demonstrates that HSE achieves smaller generalization error bounds, more stable convergence and higher computational efficiency. Experimental results validate its effective balance between acquiring new knowledge and mitigating catastrophic forgetting, maintaining or even slightly enhancing general capabilities. In practical applications, experiments confirm its effectiveness in multi-domain hallucination mitigation, healthcare knowledge injecting, and societal bias reduction.

For safety guard and crime prevention, video surveillance systems have been pervasively deployed in many security-critical scenarios, such as the residence, retail stores, and banks. However, these systems could be infiltrated by the adversary and the video streams would be modified or replaced, i.e., under the video forgery attack. The prevalence of Internet of Things (IoT) devices and the emergence of Deepfake-like techniques severely emphasize the vulnerability of video surveillance systems under such attacks. To secure existing surveillance systems, in this paper we propose a vision-WiFi cross-modal video forgery detection system, namelyWiSil. Leveraging a theoretical model based on the principle of signal propagation,WiSilconstructs wave front information of the object in the monitoring area from WiFi signals. With a well-designed deep learning network,WiSilfurther recovers silhouettes from the wave front information. Based on a Siamese network-based semantic feature extractor,WiSilcan eventually determine whether a frame is manipulated by comparing the semantic feature vectors extracted from the video’s silhouette with those extracted from the WiFi’s silhouette. We enhance the basic version ofWiSilFang et al. 2023 by developing a model compression method and a forgery trace localization method. Extensive experiments show thatWiSilachieves 95%$+$accuracy in detecting tampered frames.

With a wide range of common and privacy-sensitive applications, smartphones are frequently accessed for substantial personal information. Therefore, user-friendliness and security are crucial for user authentication on smartphones. Recently, convenient and secure biometric-based authentication is widely employed for smartphones, where the facial authentication stands out due to its potential for advancements in both user-friendliness and security. However, existing facial authentication methods possess some defects. For example, camera-based methods require good illumination conditions and are susceptible to 2D spoofing attacks. Moreover, previous acoustic-based methods either require camera assistance, or still suffer from 3D spoofing attacks. Even worse, some acoustic-based methods use audible sound waves, causing discomfort to users. To solve these questions, in this paper we propose UltraFace, an anti-spoofing and user-friendly facial authentication system on smartphones. It extracts facial geometry features and acoustic impedance features from imperceptible ultrasound. Leveraging the principle of ultrasound propagation, UltraFace correlates spectrograms of reflected signals with facial biometrics. Utilizing a deep learning model as a feature extractor, UltraFace mines fine-grained facial geometry features and acoustic impedance features from the spectrograms for accurate user authentication. Extensive experiments show that UltraFace achieves $97.2 \%$ accuracy in user authentication and can effectively defend against spoofing attacks. Furthermore, UltraFace exhibits robustness for long-term usage.

Smartwatches have become increasingly prevalent in people’s daily lives, offering support for a wide range of privacy and security-sensitive applications, such as SMS messaging and mobile payment. Consequently, there is an imperative need for independent user authentication on smartwatches to safeguard against property loss and personal privacy breaches. However, current authentication methods rely on passwords, leaving users vulnerable to shoulder surfing attacks. Moreover, existing biometric-based authentication methods either require dedicated sensors or cannot support continuous authentication. In this paper, we propose a replay-resistant continuous authentication system on smartwatches, namely WristPass. It extracts acoustic impedance biometrics from ultrasonic signals. Leveraging a theoretical model based on the principle of ultrasound propagation, WristPass correlates spectrograms of reflected signals with impedance features of wrist skin. Utilizing a deep learning model as a feature extractor, WristPass mines fine-grained impedance features from the spectrograms for accurate user authentication. Additionally, to prevent WristPass from replay attacks, we design a device fingerprinting method to detect replayed signals. Extensive experiments show that WristPass can achieve 96.7% accuracy in user authentication. Furthermore, WristPass exhibits robustness for long-term usage.