Ziyou Jiang, Mingyang Li, Junjie Wang, Yuekai Huang, Jie Huang, Zhiyuan Chang, Zhaoyang Li, Qing Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Visual entailment (VE) is a multimodal reasoning task consisting of image-sentence pairs whereby a promise is defined by an image, and a sentence describes a hypothesis. The goal is to predict whether the image semantically entails the sentence. VE systems have been widely adopted in many downstream tasks such as image caption and visual question answering. However, the robustness of VE systems still faces significant challenges. One of the reasons is that the VE system suffers object-confusing defect when some similar objects exist. It outputs a positive prediction inferred by an erroneous object relationship, which will result in a fault negative prediction if the noised object does not exist. Previous approaches generate tests primarily relied on some general perturbations, such as simulating noise or weather interference in images, or substituting synonyms or rewriting sentences in texts. To test the object-confusing defect in VE systems, it requires perceiving and understanding key objects and entities and maintain the semantic relevance between cross-modal inputs, making it challenging to generate effective tests with high quality. Therefore, we propose VEglue , an object-aligned joint erasing approach for VE systems testing. It first aligns the object regions in the premise and object descriptions in the hypothesis to identify linked and un-linked objects. Then, based on the alignment information, three metamorphic relations are designed to jointly erase the objects of the two modalities. We evaluate VEglue on four widely used VE systems involving two public datasets, and the results demonstrate that VEglue could detect 11,609 issues on average with a 52.5% Issue Finding Rate (IFR). Furthermore, we leverage the tests generated by VEglue to retrain the VE systems, which largely improves model performance (50.8% increase in accuracy) on newly generated tests without sacrificing the accuracy on the original test set.

Information theft attacks pose a significant risk to Large Language Model (LLM) tool-learning systems. Adversaries can inject malicious commands through compromised tools, manipulating LLMs to send sensitive information to these tools, which leads to potential privacy breaches. However, existing attack approaches are black-box oriented and rely on static commands that cannot adapt flexibly to the changes in user queries and the invocation chain of tools. It makes malicious commands more likely to be detected by LLM and leads to attack failure. In this paper, we propose AutoCMD, a dynamic attack comment generation approach for information theft attacks in LLM tool-learning systems. Inspired by the concept of mimicking the familiar, AutoCMD is capable of inferring the information utilized by upstream tools in the toolchain through learning on open-source systems and reinforcement with target system examples, thereby generating more targeted commands for information theft. The evaluation results show that AutoCMD outperforms the baselines with +13.2% ASR_{Theft}, and can be generalized to new tool-learning systems to expose their information leakage risks. We also design four defense methods to effectively protect tool-learning systems from the attack.

Requirements linking is the core of software system maintenance and evolution, and it is critical to assuring software quality. In practice, however, the requirements links are frequently absent or incorrectly labeled, and reconstructing such ties is time-consuming and error-prone. Numerous learning-based approaches have been put forth to address the problem. However, these approaches will lose effectiveness for the Cold-Start projects with few labeled samples. To this end, we propose RADIATION, an adversarial-based domain adaptation approach for cross-domain requirements linking. Generally, RADIATION firstly adopts an IDF-based Masking strategy to filter the domain-specific features. Then it pre-trains a linking model in the source domain with sufficient labeled samples and adapts the model to target domains using a distance-enhanced adversarial technique without using any labeled target samples. Evaluation on five public datasets shows that RADIATION could achieve 66.4% precision, 89.2% recall, and significantly outperform state-of-the-art baselines by 13.4% -42.9% F1. In addition, the designed components, i.e., IDF-based Masking and Distance-enhanced Loss, could significantly improve performance.

For the wearable biomedical signal detection, both high accuracy and low-power consumption are critical requirements. Various works have employed the neural network to improve the detecting accuracy and develop the biomedical processor. However, the biomedical processor with neural network engine contains massive data movements and large data buffers. One solution is the computing-in memory (CIM) architecture, which locates more data near the computing engine to reduce data movements. In traditional CIM-based solution, the detecting accuracy and power consumption is difficult to be optimized simultaneously, where the accuracy should be satisfied for the detection. To date, the time-domain computing engine have been developed to employ both digital and time domain computation. In this work, we present a high-precision time-domain engine to perform 8-bit multiplication and addition operation for the biomedical signal detection. With the high precision time-domain engine, we develop a CIM-based neural-network processor, namely TDPRO, to perform the detection of arrhythmia. In addition, we develop TD-zero-jumping (TDJ) and idle-shutdown (ISD) techniques according to signal features and data mapping strategy, further optimizing the power consumption. Based on our evaluation, the TD-based 8-bit mulitply-accumulation operation is robust, without declining the accuracy of biomedical signal detection. We design a ECG processor with the proposed TDPRO architecture, which obtains 98.60% high accuracy and 75.7% power saving compared to the recent the state-of-the-art study.

Natural language (NL) documentation is the bridge between software managers and testers, and NL test cases are prevalent in system-level testing and other quality assurance activities. Due to reasons such as requirements redundancy, parallel testing, tester turn-over within long evolving history, there are inevitably lots of redundant test cases, which significantly increase the cost. Previous redundancy detection approaches typically treat the textual descriptions as a whole to compare their similarity and suffer from low precision. Our observation reveals that a test case can have explicit test-oriented entities, such as tested function Components, Constraints, etc; and there are also specific relations between these entities. This inspires us with a potential opportunity for accurate redundancy detection. In this paper, we first define five test-oriented entity categories and four associated relation categories, and re-formulate the NL test case redundancy detection problem as the comparison of detailed testing content guided by the test-oriented entities and relations. Following that, we propose Tscope, a fine-grained approach for redundant NL test case detection by dissecting test cases into atomic test tuple(s) with the entities restricted by associated relations. To serve as the test case dissection, Tscope designs a context-aware model for the automatic entity and relation extraction. Evaluation on 3,467 test cases from ten projects shows Tscope could achieve 91.8% precision, 74.8% recall and 82.4% F1, significantly outperforming state-of-the-art approaches and commonly-used classifiers. This new formulation of the NL test case redundant detection problem can motivate the follow-up studies in further improving this task and other related tasks involving NL descriptions.