论文检索

Jiate Li, Meng Pang, Binghui Wang

Link prediction in dynamic graphs (LPDG) has been widely applied to real-world applications such as website recommendation, traffic flow prediction, organizational studies, etc. These models are usually kept local and secure, with only the interactive interface restrictively available to the public. Thus, the problem of the black-box evasion attack on the LPDG model, where model interactions and data perturbations are restricted, seems to be essential and meaningful in practice. In this paper, we propose the first practicable black-box evasion attack method that achieves effective attacks against the target LPDG model, within a limited amount of interactions and perturbations. To perform effective attacks under limited perturbations, we develop a graph sequential embedding model to find the desired state embedding of the dynamic graph sequences, under a deep reinforcement learning framework. To overcome the scarcity of interactions, we design a multi-environment training pipeline and train our agent for multiple instances, by sharing an aggregate interaction buffer. Finally, we evaluate our attack against three advanced LPDG models on three real-world graph datasets of different scales and compare its performance with related methods under the interaction and perturbation constraints. Experimental results show that our attack is both effective and practicable.

Jia Li, Lijie Hu, Jingfeng Zhang, Tianhang Zheng, Hua Zhang, Di Wang

In this paper, we address the limitations of existing text-to-image diffusion models in generating demographically fair results when given human-related descriptions. These models often struggle to disentangle the target language context from sociocultural biases, resulting in biased image generation. To overcome this challenge, we propose Fair Mapping, a flexible, model-agnostic, and lightweight approach that modifies a pre-trained text-to-image diffusion model by controlling the prompt to achieve fair image generation. One key advantage of our approach is its high efficiency. It only requires updating an additional linear network with few parameters at a low computational cost. By developing a linear network that maps conditioning embeddings into a debiased space, we enable the generation of relatively balanced demographic results based on the specified text condition. With comprehensive experiments on face image generation, we show that our method significantly improves image generation fairness with almost the same image quality compared to conventional diffusion models when prompted with descriptions related to humans. By effectively addressing the issue of implicit language bias, our method produces more fair and diverse image outputs.

Jinhyeok Jang, Jaehong Kim, Chan-Hyun Youn

In artificial intelligence (AI), many legal conflicts have arisen, especially concerning privacy and copyright associated with training data. When an AI model's training data incurs privacy concerns, it becomes imperative to develop a new model devoid of influences from such contentious data. However, retraining from scratch is often not viable due to the extensive data requirements and heavy computational costs. Machine unlearning presents a promising solution by enabling the selective erasure of specific knowledge from models. Despite its potential, many existing approaches in machine unlearning are based on scenarios that are either impractical or could lead to unintended degradation of model performance. We utilize the concept of weight prediction to approximate the less-learned weights based on observations about further training. By repetition of 1) finetuning on specific data and 2) weight prediction, our work gradually eliminates knowledge about the specific data. We verify its ability to eliminate side effects caused by problematic data and show its effectiveness across various architectures, datasets, and tasks.

Yihao Huang, Le Liang, Tianlin Li, Xiaojun Jia, Run Wang, Weikai Miao, Geguang Pu, Yang Liu

In recent years, Text-to-Image (T2I) models have garnered significant attention due to their remarkable advancements. However, security concerns have emerged due to their potential to generate inappropriate or Not-Safe-For-Work (NSFW) images. In this paper, inspired by the observation that texts with different semantics can lead to similar human perceptions, we propose an LLM-driven perception-guided jailbreak method, termed PGJ. It is a black-box jailbreak method that requires no specific T2I model (model-free) and generates highly natural attack prompts. Specifically, we propose identifying a safe phrase that is similar in human perception yet inconsistent in text semantics with the target unsafe word and using it as a substitution. The experiments conducted on six open-source models and commercial online services with thousands of prompts have verified the effectiveness of PGJ.

Jiliang Hu, Jiajia Li, Ziyi Pan, Chong Chen, Zuchao Li, Ping Wang, Lefei Zhang

Recently, there have been significant advancements in music generation. However, existing models primarily focus on creating modern pop songs, making it challenging to produce ancient music with distinct rhythms and styles, such as ancient Chinese SongCi. In this paper, we introduce SongSong, the first music generation model capable of restoring Chinese SongCi to our knowledge. Our model first predicts the melody from the input SongCi, then separately generates the singing voice and accompaniment based on that melody, and finally combines all elements to create the final piece of music. Additionally, to address the lack of ancient music datasets, we create OpenSongSong, a comprehensive dataset of ancient Chinese SongCi music, featuring 29.9 hours of compositions by various renowned SongCi music masters. To assess SongSong's proficiency in performing SongCi, we randomly select 85 SongCi sentences that were not part of the training set for evaluation against SongSong and music generation platforms such as Suno and SkyMusic. The subjective and objective outcomes indicate that our proposed model achieves leading performance in generating high-quality SongCi music.

Jiaming He, Wenbo Jiang, Guanyu Hou, Wenshu Fan, Rui Zhang, Hongwei Li

Mainstream backdoor attacks on large language models (LLMs) typically set a fixed trigger in the input instance and specific responses for triggered queries. However, the fixed trigger setting (e.g., unusual words) may be easily detected by human detection, limiting the effectiveness and practicality in real-world scenarios. To enhance the stealthiness of backdoor activation, we present a new poisoning paradigm against LLMs triggered by specifying generation conditions, which are commonly adopted strategies by users during model inference. The poisoned model performs normally for output under normal/other generation conditions, while becomes harmful for output under target generation conditions. To achieve this objective, we introduce BrieFool, an efficient attack framework. It leverages the characteristics of generation conditions by efficient instruction sampling and poisoning data generation, thereby influencing the behavior of LLMs under target conditions. Our attack can be generally divided into two types with different targets: Safety unalignment attack and Ability degradation attack. Our extensive experiments demonstrate that BrieFool is effective across safety domains and ability domains, achieving higher success rates than baseline methods, with 94.3% on GPT-3.5-turbo.

Odd Erik Gundersen, Odd Cappelen, Martin Mølnå, Nicklas Grimstad Nilsen

A reproducibility crisis has been reported in science, but the extent to which it affects AI research is not yet fully understood. Therefore, we performed a systematic replication study including 30 highly cited AI studies relying on original materials when available. In the end, eight articles were rejected because they required access to data or hardware that was practically impossible to acquire as part of the project. Six articles were successfully reproduced, while five were partially reproduced. In total, 50% of the articles included was reproduced to some extent. The availability of code and data correlate strongly with reproducibility, as 86% of articles that shared code and data were fully or partly reproduced, while this was true for 33% of articles that shared only data. The quality of the data documentation correlates with successful replication. Poorly documented or miss-specified data will probably result in unsuccessful replication. Surprisingly, the quality of the code documentation does not correlate with successful replication. Whether the code is poorly documented, partially missing, or not versioned is not important for successful replication, as long as the code is shared. This study emphasizes the effectiveness of open science and the importance of properly documenting data work.

Shuyang Dong, Meiyi Ma, Josephine Lamp, Sebastian Elbaum, Matthew B. Dwyer, Lu Feng

There is a growing trend toward AI systems interacting with humans to revolutionize a range of application domains such as healthcare and transportation. However, unsafe human-machine interaction can lead to catastrophic failures. We propose a novel approach that predicts future states by accounting for the uncertainty of human interaction, monitors whether predictions satisfy or violate safety requirements, and adapts control actions based on the predictive monitoring results. Specifically, we develop a new quantitative predictive monitor based on Signal Temporal Logic with Uncertainty (STL-U) to compute a robustness degree interval, which indicates the extent to which a sequence of uncertain predictions satisfies or violates an STL-U requirement. We also develop a new loss function to guide the uncertainty calibration of Bayesian deep learning and a new adaptive control method, both of which leverage STL-U quantitative predictive monitoring results. We apply the proposed approach to two case studies: Type 1 Diabetes management and semi-autonomous driving. Experiments show that the proposed approach improves safety and effectiveness in both case studies.

Tianyu Zong, Bingkang Shi, Hongzhu Yi, Jungang Xu

Unsupervised sentence embedding representation has become a hot research topic in natural language processing. As a tensor, sentence embedding has two critical properties: direction and norm. Existing works have been limited to constraining only the orientation of the samples' representations while ignoring the features of their module lengths. To address this issue, we propose a new training objective that optimizes the training of unsupervised contrastive learning by constraining the module length features between positive samples. We combine the training objective of Tensor's Norm Constraints with ensemble learning to propose a new Sentence Embedding representation framework, TNCSE. We evaluate seven semantic text similarity tasks, and the results show that TNCSE and derived models are the current state-of-the-art approach; in addition, we conduct extensive zero-shot evaluations, and the results show that TNCSE outperforms other baselines.

Wenwen Zhuang, Xin Huang, Xiantao Zhang, Jin Zeng

Multimodal Large Language Models (MLLMs) excel in solving text-based mathematical problems, but they struggle with mathematical diagrams since they are primarily trained on natural scene images. For humans, visual aids generally enhance problem-solving, but MLLMs perform worse as information shifts from textual to visual modality. This decline is mainly due to their shortcomings in aligning images and text. To tackle aforementioned challenges, we propose Math-PUMA, a methodology focused on Progressive Upward Multimodal Alignment. This approach is designed to improve the mathematical reasoning skills of MLLMs through a three-stage training process, with the second stage being the critical alignment stage. We first enhance the language model's mathematical reasoning capabilities with extensive set of textual mathematical problems. We then construct a multimodal dataset with varying degrees of textual and visual information, creating data pairs by presenting each problem in at least two forms. By leveraging the Kullback-Leibler (KL) divergence of next-token prediction distributions to align visual and textual modalities, consistent problem-solving abilities are ensured. Finally, we utilize multimodal instruction tuning for MLLMs with high-quality multimodal data. Experimental results on multiple mathematical reasoning benchmarks demonstrate that the MLLMs trained with Math-PUMA surpass most open-source MLLMs. Our approach effectively narrows the performance gap for problems presented in different modalities.

Richard Zhuang, Akshat Gupta, Richard Yang, Aniket Rahane, Zhengyu Li, Gopala Anumanchipalli

We introduce PokerBench - a benchmark for evaluating the poker-playing abilities of large language models (LLMs). As LLMs excel in traditional NLP tasks, their application to complex, strategic games like poker poses a new challenge. Poker, an incomplete information game, demands a multitude of skills such as mathematics, reasoning, planning, strategy, and a deep understanding of game theory and human psychology. This makes Poker the ideal next frontier for large language models. PokerBench consists of a comprehensive compilation of 11,000 most important scenarios, split between pre-flop and post-flop play, developed in collaboration with trained poker players. We evaluate prominent models including GPT-4, ChatGPT 3.5, and various Llama and Gemma series models, finding that all state-of-the-art LLMs underperform in playing optimal poker. However, after fine-tuning, these models show marked improvements. We validate PokerBench by having models with different scores compete with each other, demonstrating that higher scores on PokerBench leads to higher win rates in actual poker games. Through gameplay between our fine-tuned model and GPT-4, we also identify limitations of simple supervised fine-tuning for learning optimal playing strategy, suggesting the need for more advanced methodologies for effectively training language models to excel in games. PokerBench thus presents a unique benchmark for a quick and reliable evaluation of the poker-playing ability of LLMs as well as a comprehensive benchmark to study the progress of LLMs in complex game-playing scenarios.

Yutao Zhu, Zhaoheng Huang, Zhicheng Dou, Ji-Rong Wen

Retrieval-augmented generation (RAG) is a promising way to improve large language models (LLMs) for generating more factual, accurate, and up-to-date content. Existing methods either optimize prompts to guide LLMs in leveraging retrieved information or directly fine-tune LLMs to adapt to RAG scenarios. Although fine-tuning can yield better performance, it often compromises the LLMs' general generation capabilities by modifying their parameters. This limitation poses challenges in practical applications, especially when LLMs are already deployed, as parameter adjustments may affect their original functionality. To address this, we propose a novel method that involves learning scalable and pluggable virtual tokens for RAG. By maintaining the LLMs' original parameters and fine-tuning only the embeddings of these pluggable tokens, our approach not only enhances LLMs' performance but also preserves their general generation capabilities. Furthermore, we design several training strategies to improve the scalability, flexibility, and generalizability of our method. Comprehensive experiments across 12 question-answering tasks demonstrate the superiority of our approach.

Runchuan Zhu, Zhipeng Ma, Jiang Wu, Junyuan Gao, Jiaqi Wang, Dahua Lin, Conghui He

Refusal-Aware Instruction Tuning (RAIT) enables Large Language Models (LLMs) to refuse to answer unknown questions. By modifying responses of unknown questions in the training data to refusal responses such as ''I don't know", RAIT enhances the reliability of LLMs and reduces their hallucination. Generally, RAIT modifies training samples based on the correctness of the initial LLM's response. However, this crude approach can cause LLMs to excessively refuse answering questions they could have correctly answered, the problem we call over-refusal. In this paper, we explore two primary causes of over-refusal: Static conflict occurs when similar samples within the LLM’s feature space receive differing supervision signals (original vs. modified ''I don't know"). Dynamic conflict arises as the LLM's evolving knowledge during SFT enables it to answer previously unanswerable questions, but the now-answerable training samples still retain the original ''I don't know" supervision signals from the initial LLM state, leading to inconsistencies. These conflicts cause the trained LLM to misclassify known questions as unknown, resulting in over-refusal. To address this issue, we introduce Certainty Represented Knowledge Flow for Refusal-Aware Instructions Tuning (CRaFT). CRaFT centers on two main contributions: First, we additionally incorporate response certainty to selectively filter and modify data, reducing static conflicts. Second, we implement preliminary rehearsal training to characterize changes in the LLM's knowledge state, which helps mitigate dynamic conflicts during the fine-tuning process. We conducted extensive experiments on open-ended question answering and multiple-choice question task. Experiment results show that CRaFT can improve LLM's overall performance during the RAIT process.

Qiming Zhu, Jialun Cao, Yaojie Lu, Hongyu Lin, Xianpei Han, Le Sun, Shing-Chi Cheung

Code benchmarks such as HumanEval are widely adopted to evaluate the capabilities of Large Language Models (LLMs), providing insights into their strengths and weaknesses. However, current benchmarks primarily exercise LLMs' capability on common coding tasks (e.g., bubble sort, greatest common divisor), leaving domain-specific coding tasks (e.g., computation, system, cryptography) unexplored. To fill this gap, we propose a multi-domain code benchmark, DOMAINEVAL, designed to evaluate LLMs' coding capabilities thoroughly. Our pipeline works in a fully automated manner, enabling a push-button construction from code repositories into formatted subjects under study. Interesting findings are observed by evaluating 12 representative LLMs against DOMAINEVAL. We notice that LLMs are generally good at computation tasks while falling short on cryptography and system coding tasks. The performance gap can be as much as 68.94% (80.94% - 12.0%) in some LLMs. We also observe that generating more samples can increase the overall performance of LLMs, while the domain bias may even increase. The contributions of this study include a code generation benchmark dataset DOMAINEVAL, encompassing six popular domains, a fully automated pipeline for constructing code benchmarks, and an identification of the limitations of LLMs in code generation tasks based on their performance on DOMAINEVAL, providing directions for future research improvements.

Mengna Zhu, Kaisheng Zeng, Mao Wang, Kaiming Xiao, Lei Hou, Hongbin Huang, Juanzi Li

In real life, many dynamic events, such as major disasters and large-scale sports events, evolve continuously over time. Obtaining an overview of these events can help people quickly understand the situation and respond more effectively. This is challenging because the key information of the event is often scattered across multiple documents, involving complex event knowledge understanding and reasoning, which is under-explored in previous work. Therefore, we proposed the Event-Centric Multi-Document Summarization task, which aims to generate concise and comprehensive summaries of a given event based on multiple related news documents. Based on this, we constructed the EventSum dataset, which was constructed using Baidu Baike entries and underwent extensive human annotation, to facilitate relevant research. It is the first large-scale Chinese multi-document summarization dataset, containing 5,100 events and a total of 57,984 news documents, with an average of 11.4 input news documents and 13,471 characters per event. To ensure data quality and mitigate potential data leakage, we adopted a multi-stage annotation approach for manually labeling the test set. Given the complexity of event-related information, existing metrics struggle to comprehensively assess the quality of generated summaries. We designed specific metrics including Event Recall, Argument Recall, Causal Recall, and Temporal Recall along with corresponding calculation methods for evaluation. We conducted comprehensive experiments on EventSum to evaluate the performance of advanced long-context Large Language Models (LLMs) on this task. Our experimental results indicate that: 1) The event-centric multi-document summarization task remains challenging for existing long-context LLMs; 2) The recall metrics we designed are crucial for evaluating the comprehensiveness of the summary information.

Jia Zhu, Hanghui Guo, Weijie Shi, Zhangze Chen, Pasquale De Meo

The Dynamic Retrieval Augmented Generation (RAG) paradigm actively decides when and what to retrieve during the text generation process of Large Language Models (LLMs). However, current dynamic RAG methods fall short in both aspects: identifying the optimal moment to activate the retrieval module and crafting the appropriate query once retrieval is triggered. To overcome these limitations, we introduce an approach, namely, RaDIO, Real-Time Hallucination Detection with Contextual Index Optimized query formulation for dynamic RAG. The approach is specifically designed to make decisions on when and what to retrieve based on the LLM’s real-time information needs during the text generation process. We evaluate RaDIO along with existing methods comprehensively over several knowledge-intensive generation datasets. Experimental results show that RaDIO achieves superior performance on all tasks, demonstrating the effectiveness of our work.

Zhenghao Zhou, Yongjie Liu, Chen Cao

Human feedback in generative systems is a highly active frontier of research that aims to improve the quality of generated content and align it with subjective preferences. Existing efforts predominantly focus on text-only large language models (LLMs) or text-based image generation, while cross-modal generation between audio and text remains largely unexplored. Moreover, there is currently no open-source preference dataset to support the deployment of alignment algorithms in this domain. In this work, we take audio speech translation (AST) and audio captioning (AAC) tasks as examples to explore how to enhance the performance of mainstream audio-based text generation models with limited human annotation. Specifically, we propose an novel framework named IPO that includes a model adversarial sampling concept--human annotators act as referees to determine model outcomes, using these results as pseudo-labels for the corresponding beam search hypotheses. Given these imbalance win-loss results, IPO effectively enable the two models to update interactively to win the next round of adversarial sampling. We conduct both subjective and objective evaluations to demonstrate the alignment benefits of IPO and its enhancement on model perception and generation capacities. On both AAC and AST, a few hundreds of annotations significantly enhance the weak model, and the strong model can also be encouraged to achieve new state-of-the-art results in terms of different objective metrics. Additionally, we show the extensibility of IPO by applying it to the reverse task of text-to-speech generation, improving the robustness of system on unseen reference speaker.

Wei Zhou, Wei Wei, Guibang Cao, Fei Wang

Model editing is a novel research topic in large language models (LLMs), aimed at efficiently handling various knowledge editing tasks. Since irrelevant knowledge is difficult to measure, existing editing methods often lack explicit ways to preserve it, especially for editing methods based on the fine-tuning paradigm. They generally control the locality performance of model editing by constraining the range of changes in model parameters. However, their performance improvements are not always ideal, and may even lead to a decrease in the editing reliability. In this paper, we try to explore effective editing locality control methods based on the relationship between the stored knowledge and the strongly associated model components. Based on the discovery of ``knowledge neurons'' and enough experimental results, we further explore the potential characteristics between knowledge and model components, confirm and point out: (1) only 1% neurons have significant contributions to specific knowledge storage, and (2) these targeted neurons often have a high overlap for knowledge with similar relational descriptions, which means that knowledge with similar relationships may be severely affected when these targeted neurons are modified. Based on these findings, we propose Targeted Neurons Fine-tuning with Data Augmentation (TNF-DA), which performs data augmentation based on the relational representation of edited knowledge to improve editing locality. By freezing most of the model parameters and only fine-tuning the highly contributing neurons corresponding to the edited knowledge, we obtain desirable results in terms of generalization and specificity compared with previous fine-tuning-based methods. Extensive experiments have demonstrated the superior editing performance achieved by our proposed method.

Jinfeng Zhou, Yongkang Huang, Bosi Wen, Guanqun Bi, Yuxuan Chen, Pei Ke, Zhuang Chen, Xiyao Xiao, Libiao Peng, Kuntian Tang 等

Character-based dialogue (aka role-playing) enables users to freely customize characters for interaction, which often relies on LLMs, raising the need to evaluate LLMs’ character customization capability. However, existing benchmarks fail to ensure a robust evaluation as they often only involve a single character category or evaluate limited dimensions. Moreover, the sparsity of character features in responses makes feature-focused generative evaluation both ineffective and inefficient. To address these issues, we propose CharacterBench, the largest bilingual generative benchmark, with 22,859 human-annotated samples covering 3,956 characters from 25 detailed character categories. We define 11 dimensions of 6 aspects, classified as sparse and dense dimensions based on whether character features evaluated by specific dimensions manifest in each response. We enable effective and efficient evaluation by crafting tailored queries for each dimension to induce characters’ responses related to specific dimensions. Further, we develop CharacterJudge model for cost-effective and stable evaluations. Experiments show its superiority over SOTA automatic judges (e.g., GPT-4) and our benchmark’s potential to optimize LLMs’ character customization.

Hao Zhou, Zhijun Wang, Shujian Huang, Xin Huang, Xue Han, Junlan Feng, Chao Deng, Weihua Luo, Jiajun Chen

Large Language Models (LLMs) are often English-centric due to the disproportionate distribution of languages in their pre-training data. Enhancing non-English language capabilities through post-pretraining often results in catastrophic forgetting of high-resource languages. Previous methods either achieve good expansion with severe forgetting or slight forgetting with poor expansion, indicating the challenge of balancing language expansion while preventing forgetting. In this paper, we propose a method called MoE-LPR (Mixture-of-Experts with Language Priors Routing) to alleviate this problem. MoE-LPR employs a two-stage training approach to enhance the multilingual capability. First, the model is post-pretrained into a Mixture-of-Experts(MoE) architecture by upcycling, where all the original parameters are frozen and new experts are added. In this stage, we focus improving the ability on expanded languages, without using any original language data. Then, the model reviews the knowledge of the original languages with replay data amounting to less than 1% of post-pretraining, where we incorporate language priors routing to better recover the abilities of the original languages. Evaluations on multiple benchmarks show that MoE-LPR outperforms other post-pretraining methods. Freezing original parameters preserves original language knowledge while adding new experts preserves the learning ability. Reviewing with LPR enables effective utilization of multilingual knowledge within the parameters. Additionally, the MoE architecture maintains the same inference overhead while increasing total model parameters. Extensive experiments demonstrate MoE-LPR’s effectiveness in improving expanded languages and preserving original language proficiency with superior scalability.