Fei Wang

Mingkai Zheng, Xiu Su, Shan You et al. · cambridge

We investigate the potential of GPT-4~\cite{gpt4} to perform Neural Architecture Search (NAS) -- the task of designing effective neural architectures. Our proposed approach, \textbf{G}PT-4 \textbf{E}nhanced \textbf{N}eural arch\textbf{I}tect\textbf{U}re \textbf{S}earch (GENIUS), leverages the generative capabilities of GPT-4 as a black-box optimiser to quickly navigate the architecture search space, pinpoint promising candidates, and iteratively refine these candidates to improve performance. We assess GENIUS across several benchmarks, comparing it with existing state-of-the-art NAS techniques to illustrate its effectiveness. Rather than targeting state-of-the-art performance, our objective is to highlight GPT-4's potential to assist research on a challenging technical problem through a simple prompting scheme that requires relatively limited domain expertise\footnote{Code available at \href{https://github.com/mingkai-zheng/GENIUS}{https://github.com/mingkai-zheng/GENIUS}.}. More broadly, we believe our preliminary results point to future research that harnesses general purpose language models for diverse optimisation tasks. We also highlight important limitations to our study, and note implications for AI safety.

Nan Xu, Fei Wang, Ben Zhou et al.

While large language models (LLMs) have demonstrated increasing power, they have also given rise to a wide range of harmful behaviors. As representatives, jailbreak attacks can provoke harmful or unethical responses from LLMs, even after safety alignment. In this paper, we investigate a novel category of jailbreak attacks specifically designed to target the cognitive structure and processes of LLMs. Specifically, we analyze the safety vulnerability of LLMs in the face of (1) multilingual cognitive overload, (2) veiled expression, and (3) effect-to-cause reasoning. Different from previous jailbreak attacks, our proposed cognitive overload is a black-box attack with no need for knowledge of model architecture or access to model weights. Experiments conducted on AdvBench and MasterKey reveal that various LLMs, including both popular open-source model Llama 2 and the proprietary model ChatGPT, can be compromised through cognitive overload. Motivated by cognitive psychology work on managing cognitive load, we further investigate defending cognitive overload attack from two perspectives. Empirical studies show that our cognitive overload from three perspectives can jailbreak all studied LLMs successfully, while existing defense strategies can hardly mitigate the caused malicious uses effectively.

Fei Wang, Kaiqiang Song, Hongming Zhang et al. · tencent-ai

Abstractive summarization models typically learn to capture the salient information from scratch implicitly. Recent literature adds extractive summaries as guidance for abstractive summarization models to provide hints of salient content and achieves better performance. However, extractive summaries as guidance could be over strict, leading to information loss or noisy signals. Furthermore, it cannot easily adapt to documents with various abstractiveness. As the number and allocation of salience content pieces vary, it is hard to find a fixed threshold deciding which content should be included in the guidance. In this paper, we propose a novel summarization approach with a flexible and reliable salience guidance, namely SEASON (SaliencE Allocation as Guidance for Abstractive SummarizatiON). SEASON utilizes the allocation of salience expectation to guide abstractive summarization and adapts well to articles in different abstractiveness. Automatic and human evaluations on two benchmark datasets show that the proposed method is effective and reliable. Empirical results on more than one million news articles demonstrate a natural fifteen-fifty salience split for news article sentences, providing a useful insight for composing news articles.

Qin Liu, Wenjie Mo, Terry Tong et al. · harvard

The advancement of Large Language Models (LLMs) has significantly impacted various domains, including Web search, healthcare, and software development. However, as these models scale, they become more vulnerable to cybersecurity risks, particularly backdoor attacks. By exploiting the potent memorization capacity of LLMs, adversaries can easily inject backdoors into LLMs by manipulating a small portion of training data, leading to malicious behaviors in downstream applications whenever the hidden backdoor is activated by the pre-defined triggers. Moreover, emerging learning paradigms like instruction tuning and reinforcement learning from human feedback (RLHF) exacerbate these risks as they rely heavily on crowdsourced data and human feedback, which are not fully controlled. In this paper, we present a comprehensive survey of emerging backdoor threats to LLMs that appear during LLM development or inference, and cover recent advancement in both defense and detection strategies for mitigating backdoor threats to LLMs. We also outline key challenges in addressing these threats, highlighting areas for future research.

Bangzheng Li, Ben Zhou, Fei Wang et al.

Despite the recent advancement in large language models (LLMs) and their high performances across numerous benchmarks, recent research has unveiled that LLMs suffer from hallucinations and unfaithful reasoning. This work studies a specific type of hallucination induced by semantic associations. Specifically, we investigate to what extent LLMs take shortcuts from certain keyword/entity biases in the prompt instead of following the correct reasoning path. To quantify this phenomenon, we propose a novel probing method and benchmark called EureQA. We start from questions that LLMs will answer correctly with utmost certainty, and mask the important entity with evidence sentence recursively, asking models to find masked entities according to a chain of evidence before answering the question. During the construction of the evidence, we purposefully replace semantic clues (entities) that may lead to the correct answer with distractor clues (evidence) that will not directly lead to the correct answer but require a chain-like reasoning process. We evaluate if models can follow the correct reasoning chain instead of short-cutting through distractor clues. We find that existing LLMs lack the necessary capabilities to follow correct reasoning paths and resist the attempt of greedy shortcuts. We show that the distractor semantic associations often lead to model hallucination, which is strong evidence that questions the validity of current LLM reasoning.

Nan Xu, Fei Wang, Bangzheng Li et al.

Entity typing aims at predicting one or more words that describe the type(s) of a specific mention in a sentence. Due to shortcuts from surface patterns to annotated entity labels and biased training, existing entity typing models are subject to the problem of spurious correlations. To comprehensively investigate the faithfulness and reliability of entity typing methods, we first systematically define distinct kinds of model biases that are reflected mainly from spurious correlations. Particularly, we identify six types of existing model biases, including mention-context bias, lexical overlapping bias, named entity bias, pronoun bias, dependency bias, and overgeneralization bias. To mitigate model biases, we then introduce a counterfactual data augmentation method. By augmenting the original training set with their debiased counterparts, models are forced to fully comprehend sentences and discover the fundamental cues for entity typing, rather than relying on spurious correlations for shortcuts. Experimental results on the UFET dataset show our counterfactual data augmentation approach helps improve generalization of different entity typing models with consistently better performance on both the original and debiased test sets.

Qianqian Xie, Jiayu Zhou, Yifan Peng et al.

Automatic radiology report summarization is a crucial clinical task, whose key challenge is to maintain factual accuracy between produced summaries and ground truth radiology findings. Existing research adopts reinforcement learning to directly optimize factual consistency metrics such as CheXBert or RadGraph score. However, their decoding method using greedy search or beam search considers no factual consistency when picking the optimal candidate, leading to limited factual consistency improvement. To address it, we propose a novel second-stage summarizing approach FactReranker, the first attempt that learns to choose the best summary from all candidates based on their estimated factual consistency score. We propose to extract medical facts of the input medical report, its gold summary, and candidate summaries based on the RadGraph schema and design the fact-guided reranker to efficiently incorporate the extracted medical facts for selecting the optimal summary. We decompose the fact-guided reranker into the factual knowledge graph generation and the factual scorer, which allows the reranker to model the mapping between the medical facts of the input text and its gold summary, thus can select the optimal summary even the gold summary can't be observed during inference. We also present a fact-based ranking metric (RadMRR) for measuring the ability of the reranker on selecting factual consistent candidates. Experimental results on two benchmark datasets demonstrate the superiority of our method in generating summaries with higher factual consistency scores when compared with existing methods.

Shanmin Wang, Hui Shuai, Qingshan Liu et al.

In this paper, we propose a new Multimodal Representation Learning (MRL) method for Multimodal Sentiment Analysis (MSA), which facilitates the adaptive interaction between modalities through Cooperative Sentiment Agents, named Co-SA. Co-SA comprises two critical components: the Sentiment Agents Establishment (SAE) phase and the Sentiment Agents Cooperation (SAC) phase. During the SAE phase, each sentiment agent deals with an unimodal signal and highlights explicit dynamic sentiment variations within the modality via the Modality-Sentiment Disentanglement (MSD) and Deep Phase Space Reconstruction (DPSR) modules. Subsequently, in the SAC phase, Co-SA meticulously designs task-specific interaction mechanisms for sentiment agents so that coordinating multimodal signals to learn the joint representation. Specifically, Co-SA equips an independent policy model for each sentiment agent that captures significant properties within the modality. These policies are optimized mutually through the unified reward adaptive to downstream tasks. Benefitting from the rewarding mechanism, Co-SA transcends the limitation of pre-defined fusion modes and adaptively captures unimodal properties for MRL in the multimodal interaction setting. To demonstrate the effectiveness of Co-SA, we apply it to address Multimodal Sentiment Analysis (MSA) and Multimodal Emotion Recognition (MER) tasks. Our comprehensive experimental results demonstrate that Co-SA excels at discovering diverse cross-modal features, encompassing both common and complementary aspects. The code can be available at https://github.com/smwanghhh/Co-SA.

Xingbo Wang, Samantha L. Huey, Rui Sheng et al.

The explosion of scientific literature has made the efficient and accurate extraction of structured data a critical component for advancing scientific knowledge and supporting evidence-based decision-making. However, existing tools often struggle to extract and structure multimodal, varied, and inconsistent information across documents into standardized formats. We introduce SciDaSynth, a novel interactive system powered by large language models (LLMs) that automatically generates structured data tables according to users' queries by integrating information from diverse sources, including text, tables, and figures. Furthermore, SciDaSynth supports efficient table data validation and refinement, featuring multi-faceted visual summaries and semantic grouping capabilities to resolve cross-document data inconsistencies. A within-subjects study with nutrition and NLP researchers demonstrates SciDaSynth's effectiveness in producing high-quality structured data more efficiently than baseline methods. We discuss design implications for human-AI collaborative systems supporting data extraction tasks. The system code is available at https://github.com/xingbow/SciDaEx

Jiashu Xu, Fei Wang, Mingyu Derek Ma et al.

The exorbitant cost of training Large language models (LLMs) from scratch makes it essential to fingerprint the models to protect intellectual property via ownership authentication and to ensure downstream users and developers comply with their license terms (e.g. restricting commercial use). In this study, we present a pilot study on LLM fingerprinting as a form of very lightweight instruction tuning. Model publisher specifies a confidential private key and implants it as an instruction backdoor that causes the LLM to generate specific text when the key is present. Results on 11 popularly-used LLMs showed that this approach is lightweight and does not affect the normal behavior of the model. It also prevents publisher overclaim, maintains robustness against fingerprint guessing and parameter-efficient training, and supports multi-stage fingerprinting akin to MIT License. Code is available in https://cnut1648.github.io/Model-Fingerprint/.

Fei Wang, James Y. Huang, Tianyi Yan et al.

Natural language understanding (NLU) models often suffer from unintended dataset biases. Among bias mitigation methods, ensemble-based debiasing methods, especially product-of-experts (PoE), have stood out for their impressive empirical success. However, previous ensemble-based debiasing methods typically apply debiasing on top-level logits without directly addressing biased attention patterns. Attention serves as the main media of feature interaction and aggregation in PLMs and plays a crucial role in providing robust prediction. In this paper, we propose REsidual Attention Debiasing (READ), an end-to-end debiasing method that mitigates unintended biases from attention. Experiments on three NLU tasks show that READ significantly improves the performance of BERT-based models on OOD data with shortcuts removed, including +12.9% accuracy on HANS, +11.0% accuracy on FEVER-Symmetric, and +2.7% F1 on PAWS. Detailed analyses demonstrate the crucial role of unbiased attention in robust NLU models and that READ effectively mitigates biases in attention. Code is available at https://github.com/luka-group/READ.

Fei Wang, Wenjie Mo, Yiwei Wang et al.

Entity bias widely affects pretrained (large) language models, causing them to rely on (biased) parametric knowledge to make unfaithful predictions. Although causality-inspired methods have shown great potential to mitigate entity bias, it is hard to precisely estimate the parameters of underlying causal models in practice. The rise of black-box LLMs also makes the situation even worse, because of their inaccessible parameters and uncalibrated logits. To address these problems, we propose a specific structured causal model (SCM) whose parameters are comparatively easier to estimate. Building upon this SCM, we propose causal intervention techniques to mitigate entity bias for both white-box and black-box settings. The proposed causal intervention perturbs the original entity with neighboring entities. This intervention reduces specific biasing information pertaining to the original entity while still preserving sufficient semantic information from similar entities. Under the white-box setting, our training-time intervention improves OOD performance of PLMs on relation extraction (RE) and machine reading comprehension (MRC) by 5.7 points and by 9.1 points, respectively. Under the black-box setting, our in-context intervention effectively reduces the entity-based knowledge conflicts of GPT-3.5, achieving up to 20.5 points of improvement of exact match accuracy on MRC and up to 17.6 points of reduction in memorization ratio on RE. Our code is available at https://github.com/luka-group/Causal-View-of-Entity-Bias.

Yiwei Wang, Bryan Hooi, Fei Wang et al.

Relation extraction (RE) aims to extract the relations between entity names from the textual context. In principle, textual context determines the ground-truth relation and the RE models should be able to correctly identify the relations reflected by the textual context. However, existing work has found that the RE models memorize the entity name patterns to make RE predictions while ignoring the textual context. This motivates us to raise the question: ``are RE models robust to the entity replacements?'' In this work, we operate the random and type-constrained entity replacements over the RE instances in TACRED and evaluate the state-of-the-art RE models under the entity replacements. We observe the 30\% - 50\% F1 score drops on the state-of-the-art RE models under entity replacements. These results suggest that we need more efforts to develop effective RE models robust to entity replacements. We release the source code at https://github.com/wangywUST/RobustRE.

Fei Wang, Kexuan Sun, Jay Pujara et al.

Tables provide valuable knowledge that can be used to verify textual statements. While a number of works have considered table-based fact verification, direct alignments of tabular data with tokens in textual statements are rarely available. Moreover, training a generalized fact verification model requires abundant labeled training data. In this paper, we propose a novel system to address these problems. Inspired by counterfactual causality, our system identifies token-level salience in the statement with probing-based salience estimation. Salience estimation allows enhanced learning of fact verification from two perspectives. From one perspective, our system conducts masked salient token prediction to enhance the model for alignment and reasoning between the table and the statement. From the other perspective, our system applies salience-aware data augmentation to generate a more diverse set of training instances by replacing non-salient terms. Experimental results on TabFact show the effective improvement by the proposed salience-aware learning techniques, leading to the new SOTA performance on the benchmark. Our code is publicly available at https://github.com/luka-group/Salience-aware-Learning .

Terry Tong, Fei Wang, Zhe Zhao et al.

This paper proposes a novel backdoor threat attacking the LLM-as-a-Judge evaluation regime, where the adversary controls both the candidate and evaluator model. The backdoored evaluator victimizes benign users by unfairly assigning inflated scores to adversary. A trivial single token backdoor poisoning 1% of the evaluator training data triples the adversary's score with respect to their legitimate score. We systematically categorize levels of data access corresponding to three real-world settings, (1) web poisoning, (2) malicious annotator, and (3) weight poisoning. These regimes reflect a weak to strong escalation of data access that highly correlates with attack severity. Under the weakest assumptions - web poisoning (1), the adversary still induces a 20% score inflation. Likewise, in the (3) weight poisoning regime, the stronger assumptions enable the adversary to inflate their scores from 1.5/5 to 4.9/5. The backdoor threat generalizes across different evaluator architectures, trigger designs, evaluation tasks, and poisoning rates. By poisoning 10% of the evaluator training data, we control toxicity judges (Guardrails) to misclassify toxic prompts as non-toxic 89% of the time, and document reranker judges in RAG to rank the poisoned document first 97% of the time. LLM-as-a-Judge is uniquely positioned at the intersection of ethics and technology, where social implications of mislead model selection and evaluation constrain the available defensive tools. Amidst these challenges, model merging emerges as a principled tool to offset the backdoor, reducing ASR to near 0% whilst maintaining SOTA performance. Model merging's low computational cost and convenient integration into the current LLM Judge training pipeline position it as a promising avenue for backdoor mitigation in the LLM-as-a-Judge setting.

Qin Liu, Fei Wang, Nan Xu et al.

Performance of large language models (LLMs) may vary with different prompts or instructions of even the same task. One commonly recognized factor for this phenomenon is the model's familiarity with the given prompt or instruction, which is typically estimated by its perplexity. However, finding the prompt with the lowest perplexity is challenging, given the enormous space of possible prompting phrases. In this paper, we propose monotonic paraphrasing (MonoPara), an end-to-end decoding strategy that paraphrases given prompts or instructions into their lower perplexity counterparts based on an ensemble of a paraphrase LM for prompt (or instruction) rewriting, and a target LM (i.e. the prompt or instruction executor) that constrains the generation for lower perplexity. The ensemble decoding process can efficiently paraphrase the original prompt without altering its semantic meaning, while monotonically decreasing the perplexity of each generation as calculated by the target LM. We explore in detail both greedy and search-based decoding as two alternative decoding schemes of MonoPara. Notably, MonoPara does not require any training and can monotonically lower the perplexity of the paraphrased prompt or instruction, leading to improved performance of zero-shot LM prompting as evaluated on a wide selection of tasks. In addition, MonoPara is also shown to effectively improve LMs' generalization on perturbed and unseen task instructions.

Lihui Liu, Blaine Hill, Boxin Du et al.

Conversational question answering (convQA) over knowledge graphs (KGs) involves answering multi-turn natural language questions about information contained in a KG. State-of-the-art methods of ConvQA often struggle with inexplicit question-answer pairs. These inputs are easy for human beings to understand given a conversation history, but hard for a machine to interpret, which can degrade ConvQA performance. To address this problem, we propose a reinforcement learning (RL) based model, CornNet, which utilizes question reformulations generated by large language models (LLMs) to improve ConvQA performance. CornNet adopts a teacher-student architecture where a teacher model learns question representations using human writing reformulations, and a student model to mimic the teacher model's output via reformulations generated by LLMs. The learned question representation is then used by an RL model to locate the correct answer in a KG. Extensive experimental results show that CornNet outperforms state-of-the-art convQA models.

Hadi Askari, Shivanshu Gupta, Fei Wang et al.

Pretrained Large Language Models (LLMs) achieve strong performance across a wide range of tasks, yet exhibit substantial variability in the various layers' training quality with respect to specific downstream applications, limiting their downstream performance. It is therefore critical to estimate layer-wise training quality in a manner that accounts for both model architecture and training data. However, existing approaches predominantly rely on model-centric heuristics (such as spectral statistics, outlier detection, or uniform allocation) while overlooking the influence of data. To address these limitations, we propose LayerIF, a data-driven framework that leverages Influence Functions to quantify the training quality of individual layers in a principled and task-sensitive manner. By isolating each layer's gradients and measuring the sensitivity of the validation loss to training examples by computing layer-wise influences, we derive data-driven estimates of layer importance. Notably, our method produces task-specific layer importance estimates for the same LLM, revealing how layers specialize for different test-time evaluation tasks. We demonstrate the utility of our scores by leveraging them for two downstream applications: (a) expert allocation in LoRA-MoE architectures and (b) layer-wise sparsity distribution for LLM pruning. Experiments across multiple LLM architectures demonstrate that our model-agnostic, influence-guided allocation leads to consistent gains in task performance.

Tanay Dixit, Fei Wang, Muhao Chen

Improving factual consistency of abstractive summarization has been a widely studied topic. However, most of the prior works on training factuality-aware models have ignored the negative effect it has on summary quality. We propose EFACTSUM (i.e., Effective Factual Summarization), a candidate summary generation and ranking technique to improve summary factuality without sacrificing summary quality. We show that using a contrastive learning framework with our refined candidate summaries leads to significant gains on both factuality and similarity-based metrics. Specifically, we propose a ranking strategy in which we effectively combine two metrics, thereby preventing any conflict during training. Models trained using our approach show up to 6 points of absolute improvement over the base model with respect to FactCC on XSUM and 11 points on CNN/DM, without negatively affecting either similarity-based metrics or absractiveness.

Qin Liu, Fei Wang, Chaowei Xiao et al.

Language models are often at risk of diverse backdoor attacks, especially data poisoning. Thus, it is important to investigate defense solutions for addressing them. Existing backdoor defense methods mainly focus on backdoor attacks with explicit triggers, leaving a universal defense against various backdoor attacks with diverse triggers largely unexplored. In this paper, we propose an end-to-end ensemble-based backdoor defense framework, DPoE (Denoised Product-of-Experts), which is inspired by the shortcut nature of backdoor attacks, to defend various backdoor attacks. DPoE consists of two models: a shallow model that captures the backdoor shortcuts and a main model that is prevented from learning the backdoor shortcuts. To address the label flip caused by backdoor attackers, DPoE incorporates a denoising design. Experiments on SST-2 dataset show that DPoE significantly improves the defense performance against various types of backdoor triggers including word-level, sentence-level, and syntactic triggers. Furthermore, DPoE is also effective under a more challenging but practical setting that mixes multiple types of trigger.

Jian Liang, Fangrui Lv, Di Liu et al.

In this paper, we focus on learning effective entity matching models over multi-source large-scale data. For real applications, we relax typical assumptions that data distributions/spaces, or entity identities are shared between sources, and propose a Relaxed Multi-source Large-scale Entity-matching (RMLE) problem. Challenges of the problem include 1) how to align large-scale entities between sources to share information and 2) how to mitigate negative transfer from joint learning multi-source data. What's worse, one practical issue is the entanglement between both challenges. Specifically, incorrect alignments may increase negative transfer; while mitigating negative transfer for one source may result in poorly learned representations for other sources and then decrease alignment accuracy. To handle the entangled challenges, we point out that the key is to optimize information sharing first based on Pareto front optimization, by showing that information sharing significantly influences the Pareto front which depicts lower bounds of negative transfer. Consequently, we proposed an Incentive Compatible Pareto Alignment (ICPA) method to first optimize cross-source alignments based on Pareto front optimization, then mitigate negative transfer constrained on the optimized alignments. This mechanism renders each source can learn based on its true preference without worrying about deteriorating representations of other sources. Specifically, the Pareto front optimization encourages minimizing lower bounds of negative transfer, which optimizes whether and which to align. Comprehensive empirical evaluation results on four large-scale datasets are provided to demonstrate the effectiveness and superiority of ICPA. Online A/B test results at a search advertising platform also demonstrate the effectiveness of ICPA in production environments.

Sendong Zhao, Fei Wang

With the rapid development of precision medicine, a large amount of health data (such as electronic health records, gene sequencing, medical images, etc.) has been produced. It encourages more and more interest in data-driven insight discovery from these data. It is a reasonable way to verify the derived insights in biomedical literature. However, manual verification is inefficient and not scalable. Therefore, an intelligent technique is necessary to solve this problem. In this paper, we propose a task of biomedical evidence generation, which is very novel and different from existing NLP tasks. Furthermore, we developed a biomedical evidence generation engine for this task with the pipeline of three components which are a literature retrieval module, a skeleton information identification module, and a text summarization module.

Wei Wu, Fei Wang, Arianna Yuan et al.

In this paper, we present an accurate and extensible approach for the coreference resolution task. We formulate the problem as a span prediction task, like in machine reading comprehension (MRC): A query is generated for each candidate mention using its surrounding context, and a span prediction module is employed to extract the text spans of the coreferences within the document using the generated query. This formulation comes with the following key advantages: (1) The span prediction strategy provides the flexibility of retrieving mentions left out at the mention proposal stage; (2) In the MRC framework, encoding the mention and its context explicitly in a query makes it possible to have a deep and thorough examination of cues embedded in the context of coreferent mentions; and (3) A plethora of existing MRC datasets can be used for data augmentation to improve the model's generalization capability. Experiments demonstrate significant performance boost over previous models, with 87.5 (+2.5) F1 score on the GAP benchmark and 83.1 (+3.5) F1 score on the CoNLL-2012 benchmark.

Sendong Zhao, Ting Liu, Sicheng Zhao et al.

State-of-the-art studies have demonstrated the superiority of joint modelling over pipeline implementation for medical named entity recognition and normalization due to the mutual benefits between the two processes. To exploit these benefits in a more sophisticated way, we propose a novel deep neural multi-task learning framework with explicit feedback strategies to jointly model recognition and normalization. On one hand, our method benefits from the general representations of both tasks provided by multi-task learning. On the other hand, our method successfully converts hierarchical tasks into a parallel multi-task setting while maintaining the mutual supports between tasks. Both of these aspects improve the model performance. Experimental results demonstrate that our method performs significantly better than state-of-the-art approaches on two publicly available medical literature datasets.