Eric Hanchen Jiang, Weixuan Ou, Run Liu, Shengyuan Pang, Guancheng Wan, Ranjie Duan, Wei Dong, Kai-Wei Chang, XiaoFeng Wang, Ying Nian Wu, Xinfeng Li. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Multimodal large language models (MLLMs) have achieved significant advancements in multimodal understanding, reasoning, and interaction. However, they still suffer from hallucination, where the generated text often deviates from the factual content of the input image. To mitigate this issue, prior studies have primarily employed direct preference optimization (DPO) for human preference alignment. However, these approaches treat all textual words equally, neglecting the varying significance of individual words in grounding text generation to image content. This limitation hinders fine-grained semantic alignment and consequently constrains their effectiveness in hallucination suppression. To address this limitation, we propose a vision-guided lexical DPO method, called VGL-DPO. Specifically, we quantify the significance of words in positive preference data based on their relevance to the visual input and dynamically assign different weights to different words during training. This facilitates more precise optimization by emphasizing critical words that contribute to factual grounding. Additionally, we leverage the importance differences between high-significance words in positive and negative preference data to adaptively adjust the weight of the negative preference loss. This dynamic reweighting mechanism further refines the model’s ability to suppress hallucinated content while reinforcing factual accuracy. Extensive experiments across various models demonstrate that our method outperforms existing state-of-the-art methods in reducing hallucination and enhancing factual accuracy.

Ruoxi Cheng, Yizhong Ding, Shuirong Cao, Ranjie Duan, Xiaoshuang Jia, Shaowei Yuan, Simeng Qin, Zhiqiang Wang, Xiaojun Jia. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Multimodal Large Language Models (MLLMs) have achieved impressive performance and have been put into practical use in commercial applications, but they still have potential safety mechanism vulnerabilities. Jailbreak attacks are red teaming methods that aim to bypass safety mechanisms and discover MLLMs' potential risks. Existing MLLMs' jailbreak methods often bypass the model's safety mechanism through complex optimization methods or carefully designed image and text prompts. Despite achieving some progress, they have a low attack success rate on commercial closed-source MLLMs. Unlike previous research, we empirically find that there exists a Shuffle Inconsistency between MLLMs' comprehension ability and safety ability for the shuffled harmful instruction. That is, from the perspective of comprehension ability, MLLMs can understand the shuffled harmful text-image instructions well. However, they can be easily bypassed by the shuffled harmful instructions from the perspective of safety ability, leading to harmful responses. Then we innovatively propose a text-image jailbreak attack named SI-Attack. Specifically, to fully utilize the Shuffle Inconsistency and overcome the shuffle randomness, we apply a query-based black-box optimization method to select the most harmful shuffled inputs based on the feedback of the toxic judge model. A series of experiments show that SI-Attack can improve the attack's performance on three benchmarks. In particular, SI-Attack can obviously improve the attack success rate for commercial MLLMs such as GPT-4o or Claude-3.5-Sonnet.

Ensuring the safety and harmlessness of Large Language Models (LLMs) has become equally critical as their performance in applications. However, existing safety alignment methods typically suffer from safety-performance trade-offs and susceptibility to jailbreak attacks, primarily due to their reliance on direct refusals for malicious queries. In this paper, we propose STAIR, a novel framework that integrates SafeTy Alignment with Itrospective Reasoning. We enable LLMs to identify safety risks through step-by-step analysis by self-improving chain-of-thought (CoT) reasoning with safety awareness. STAIR first equips the model with a structured reasoning capability and then advances safety alignment via iterative preference optimization on step-level reasoning data generated using our newly proposed Safety-Informed Monte Carlo Tree Search (SI-MCTS). Specifically, we design a theoretically grounded reward for outcome evaluation to seek balance between helpfulness and safety. We further train a process reward model on this data to guide test-time searches for improved responses. Extensive experiments show that STAIR effectively mitigates harmful outputs while better preserving helpfulness, compared to instinctive alignment strategies. With test-time scaling, STAIR achieves a safety performance comparable to Claude-3.5 against popular jailbreak attacks. We have open-sourced our code, datasets and models at https://github.com/thu-ml/STAIR.

Jianyu Liu, Hangyu Guo, Ranjie Duan, Xingyuan Bu, Yancheng He, Shilong Li, Hui Huang, Jiaheng Liu, Yucheng Wang, Chenchen Jing, Xingwei Qu, Xiao Zhang, Pei Wang, Yanan Wu, Jihao Gu, Yangguang Li, Jianke Zhu. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Retrieval-augmented generation (RAG) systems enhance large language models (LLMs) with external knowledge but are vulnerable to corpus poisoning and contamination attacks, which can compromise output integrity. Existing defenses often apply aggressive filtering, leading to unnecessary loss of valuable information and reduced reliability in generation. To address this problem, we propose a two-stage semantic filtering and conflict-free framework for trustworthy RAG. In the first stage, we perform a joint filter with semantic and cluster-based filtering which is guided by the Entity-intent-relation extractor (EIRE). EIRE extracts entities, latent objectives, and entity relations from both the user query and filtered documents, scores their semantic relevance, and selectively adds valuable documents into the clean retrieval database. In the second stage, we proposed an EIRE-guided conflict-aware filtering module, which analyzes semantic consistency between the query, candidate answers, and retrieved knowledge before final answer generation, filtering out internal and external contradictions that could mislead the model. Through this two-stage process, SeCon-RAG effectively preserves useful knowledge while mitigating conflict contamination, achieving significant improvements in both generation robustness and output trustworthiness. Extensive experiments across various LLMs and datasets demonstrate that the proposed SeCon-RAG markedly outperforms state-of-the-art defense methods.

Adversarial Training (AT) has been widely proved to be an effective method to improve the adversarial robustness against adversarial examples for Deep Neural Networks (DNNs). As a variant of AT, Adversarial Robustness Distillation (ARD) has demonstrated its superior performance in improving the robustness of small student models with the guidance of large teacher models. However, both AT and ARD encounter the robust fairness problem: these models exhibit strong robustness when facing part of classes (easy class), but weak robustness when facing others (hard class). In this paper, we give an in-depth analysis of the potential factors and argue that the smoothness degree of samples' soft labels for different classes (i.e., hard class or easy class) will affect the robust fairness of DNNs from both empirical observation and theoretical analysis. Based on the above finding, we propose an Anti-Bias Soft Label Distillation (ABSLD) method to mitigate the adversarial robust fairness problem within the framework of Knowledge Distillation (KD). Specifically, ABSLD adaptively reduces the student's error risk gap between different classes to achieve fairness by adjusting the class-wise smoothness degree of samples' soft labels during the training process, and the smoothness degree of soft labels is controlled by assigning different temperatures in KD to different classes. Extensive experiments demonstrate that ABSLD outperforms state-of-the-art AT, ARD, and robust fairness methods in the comprehensive metric (Normalized Standard Deviation) of robustness and fairness.

Fast Adversarial Training (FAT) not only improves the model robustness but also reduces the training cost of standard adversarial training. However, fast adversarial training often suffers from Catastrophic Overfitting (CO), which results in poor robustness performance. Catastrophic Overfitting describes the phenomenon of a sudden and significant decrease in robust accuracy during the training of fast adversarial training. Many effective techniques have been developed to prevent Catastrophic Overfitting and improve the model robustness from different perspectives. However, these techniques adopt inconsistent training settings and require different training costs, i.e., training time and memory costs, leading to unfair comparisons. In this paper, we conduct a comprehensive study of over 10 fast adversarial training methods in terms of adversarial robustness and training costs. We revisit the effectiveness and efficiency of fast adversarial training techniques in preventing Catastrophic Overfitting from the perspective of model local nonlinearity and propose an effective Lipschitz regularization method for fast adversarial training. Furthermore, we explore the effect of data augmentation and weight averaging in fast adversarial training and propose a simple yet effective auto weight averaging method to improve robustness further. By assembling these techniques, we propose an effective FGSM-based fast adversarial training method equipped with Lipschitz regularization and Auto Weight averaging, abbreviated as FGSM-LAW. Experimental evaluations on four benchmark databases demonstrate the superiority of our method over state-of-the-art fast adversarial training methods and the advanced standard adversarial training methods.

Recent advances on Vision Transformer (ViT) and its improved variants have shown that self-attention-based networks surpass traditional Convolutional Neural Networks (CNNs) in most vision tasks. However, existing ViTs focus on the standard accuracy and computation cost, lacking the investigation of the intrinsic influence on model robustness and generalization. In this work, we conduct systematic evaluation on components of ViTs in terms of their impact on robustness to adversarial examples, common corruptions and distribution shifts. We find some components can be harmful to robustness. By leveraging robust components as building blocks of ViTs, we propose Robust Vision Transformer (RVT), which is a new vision transformer and has superior performance with strong robustness. Inspired by the findings during the evaluation, we further propose two new plug-and-play techniques called position-aware attention scaling and patch-wise augmentation to augment our RVT, which we abbreviate as RVT*. The experimental results of RVT on ImageNet and six robustness benchmarks demonstrate its advanced robustness and generalization ability compared with previous ViTs and state-of-the-art CNNs. Furthermore, RVT-S* achieves Top-1 rank on multiple robustness leaderboards including ImageNet-C, ImageNet-Sketch and ImageNet-R.

Adversarial Training (AT), which is commonly accepted as one of the most effective approaches defending against adversarial examples, can largely harm the standard performance, thus has limited usefulness on industrial-scale production and applications. Surprisingly, this phenomenon is totally opposite in Natural Language Processing (NLP) task, where AT can even benefit for generalization. We notice the merit of AT in NLP tasks could derive from the discrete and symbolic input space. For borrowing the advantage from NLP-style AT, we propose Discrete Adversarial Training (DAT). DAT leverages VQGAN to reform the image data to discrete text-like inputs, i.e. visual words. Then it minimizes the maximal risk on such discrete images with symbolic adversarial perturbations. We further give an explanation from the perspective of distribution to demonstrate the effectiveness of DAT. As a plug-and-play technique for enhancing the visual representation, DAT achieves significant improvement on multiple tasks including image classification, object detection and self-supervised learning. Especially, the model pre-trained with Masked Auto-Encoding (MAE) and fine-tuned by our DAT without extra data can get 31.40 mCE on ImageNet-C and 32.77% top-1 accuracy on Stylized-ImageNet, building the new state-of-the-art. The code will be available at https://github.com/alibaba/easyrobust.

Though it is well known that the performance of deep neural networks (DNNs) degrades under certain light conditions, there exists no study on the threats of light beams emitted from some physical source as adversarial attacker on DNNs in a real-world scenario. In this work, we show by simply using a laser beam that DNNs are easily fooled. To this end, we propose a novel attack method called Adversarial Laser Beam (AdvLB), which enables manipulation of laser beam’s physical parameters to perform adversarial attack. Experiments demonstrate the effectiveness of our proposed approach in both digital- and physical-settings. We further empirically analyze the evaluation results and reveal that the proposed laser beam attack may lead to some interesting prediction errors of the state-of-the-art DNNs. We envisage that the proposed AdvLB method enriches the current family of adversarial attacks and builds the foundation for future robustness studies for light.

Human can easily recognize visual objects with lost information: even losing most details with only contour reserved, e.g. cartoon. However, in terms of visual perception of Deep Neural Networks (DNNs), the ability for recognizing abstract objects (visual objects with lost information) is still a challenge. In this work, we investigate this issue from an adversarial viewpoint: will the performance of DNNs decrease even for the images only losing a little information? Towards this end, we propose a novel adversarial attack, named AdvDrop, which crafts adversarial examples by dropping existing information of images. Previously, most adversarial attacks add extra disturbing information on clean images explicitly. Opposite to previous works, our proposed work explores the adversarial robustness of DNN models in a novel perspective by dropping imperceptible de-tails to craft adversarial examples. We demonstrate the effectiveness of AdvDrop by extensive experiments, and show that this new type of adversarial examples is more difficult to be defended by current defense systems.

Deep neural networks (DNNs) are known to be vulnerable to adversarial examples. Existing works have mostly focused on either digital adversarial examples created via small and imperceptible perturbations, or physical-world adversarial examples created with large and less realistic distortions that are easily identified by human observers. In this paper, we propose a novel approach, called Adversarial Camouflage (\emph{AdvCam}), to craft and camouflage physical-world adversarial examples into natural styles that appear legitimate to human observers. Specifically, \emph{AdvCam} transfers large adversarial perturbations into customized styles, which are then “hidden” on-target object or off-target background. Experimental evaluation shows that, in both digital and physical-world scenarios, adversarial examples crafted by \emph{AdvCam} are well camouflaged and highly stealthy, while remaining effective in fooling state-of-the-art DNN image classifiers. Hence, \emph{AdvCam} is a flexible approach that can help craft stealthy attacks to evaluate the robustness of DNNs.