Large language models (LLMs) have demonstrated superior performance compared to previous methods on various tasks, and often serve as the foundation models for many researches and services. However, the untrustworthy third-party LLMs may covertly introduce vulnerabilities for downstream tasks. In this paper, we explore the vulnerability of LLMs through the lens of backdoor attacks. Different from existing backdoor attacks against LLMs, ours scatters multiple trigger keys in different prompt components. Such a Composite Backdoor Attack (CBA) is shown to be stealthier than implanting the same multiple trigger keys in only a single component. CBA ensures that the backdoor is activated only when all trigger keys appear. Our experiments demonstrate that CBA is effective in both natural language processing (NLP) and multimodal tasks. For instance, with 3% poisoning samples against the LLaMA-7B model on the Emotion dataset, our attack achieves a 100% Attack Success Rate (ASR) with a False Triggered Rate (FTR) below 2.06% and negligible model accuracy degradation. Our work highlights the necessity of increased security research on the trustworthiness of foundation LLMs.

Existing literature that integrates CLIP into federated learning (FL) largely ignores the inherent group unfairness within CLIP and its ethical implications on FL applications. Furthermore, such CLIP bias may be amplified in FL, due to the unique issue of data heterogeneity across clients. However, in identity-sensitive FL applications, model fairness (i.e., group fairness) is imperative for model development. Therefore, this work explores a critical question ignored by the existing literature: how can we build a fair FL framework using biased pre-trained VLMs (e.g., CLIP)? To address this problem, we propose a fairness-aware adaptation framework tailored for VLM (e.g., CLIP) in the context of FL, named Fair Federated Deep Visiual Prompting or FF-DVP. As implied by its name, trains a fair FL model with fairness-aware deep visual prompting (DVP). Moreover, incorporates modality-fused classification heads to learn client-specific knowledge and fairness constraints. These modules explicitly addresses a unique bias in FL, namely the bias triggered by data heterogeneity. We show that can be readily extended to prevailing parameter-efficient fine-tuning methods (e.g., adapter or LoRA) for debiasing. To the best of our knowledge, is the first to leverage biased VLMs for building fair FL frameworks. Extensive results on human face attribute recognition (FAR) applications suggest that effectively improves model fairness and training convergence, outperforming state-of-the-art baselines.

The rapid advancement of Large Language Models (LLMs) in the realm of mathematical reasoning necessitates comprehensive evaluations to gauge progress and inspire future directions. Existing assessments predominantly focus on problem-solving from the examinee perspective, overlooking a dual perspective of examiner regarding error identification and correction.From the examiner perspective, we define four evaluation tasks for error identification and correction along with a new dataset with annotated error types and steps. We also design diverse prompts to thoroughly evaluate eleven representative LLMs. Our principal findings indicate that GPT-4 outperforms all models, while open-source model LLaMA-2-7B demonstrates comparable abilities to closed-source models GPT-3.5 and Gemini Pro.Notably, calculation error proves the most challenging error type. Moreover, prompting LLMs with the error types can improve the average correction accuracy by 47.9%. These results reveal potential directions for developing the mathematical reasoning abilities of LLMs.Our code and dataset is available on https://github.com/LittleCirc1e/EIC.

This paper studies the relationship between the surface form of a mathematical problem and its solvability by large language models. We find that subtle alterations in the surface form can significantly impact the answer distribution and the solve rate, exposing the language model’s lack of robustness and sensitivity to the surface form in reasoning through complex problems. To improve mathematical reasoning performance, we propose Self-Consistency-over-Paraphrases (SCoP), which diversifies reasoning paths from specific surface forms of the problem. We evaluate our approach on four mathematics reasoning benchmarks over three large language models and show that SCoP improves mathematical reasoning performance over vanilla self-consistency, particularly for problems initially deemed unsolvable. Finally, we provide additional experiments and discussion regarding problem difficulty and surface forms, including cross-model difficulty agreement and paraphrasing transferability, and Variance of Variations (VOV) for language model evaluation.

The proliferation of online misinformation has posed significant threats to public interest. While numerous online users actively participate in the combat against misinformation, many of such responses can be characterized by the lack of politeness and supporting facts. As a solution, text generation approaches are proposed to automatically produce counter-misinformation responses. Nevertheless, existing methods are often trained end-to-end without leveraging external knowledge, resulting in subpar text quality and excessively repetitive responses. In this paper, we propose retrieval augmented response generation for online misinformation (RARG), which collects supporting evidence from scientific sources and generates counter-misinformation responses based on the evidences. In particular, our RARG consists of two stages: (1) evidence collection, where we design a retrieval pipeline to retrieve and rerank evidence documents using a database comprising over 1M academic articles; (2) response generation, in which we align large language models (LLMs) to generate evidence-based responses via reinforcement learning from human feedback (RLHF). We propose a reward function to maximize the utilization of the retrieved evidence while maintaining the quality of the generated text, which yields polite and factual responses that clearly refutes misinformation. To demonstrate the effectiveness of our method, we study the case of COVID-19 and perform extensive experiments with both in- and cross-domain datasets, where RARG consistently outperforms baselines by generating high-quality counter-misinformation responses.

Although large language models (LLMs) have achieved significant success, their vulnerability to adversarial perturbations, including recent jailbreak attacks, has raised considerable concerns. However, the increasing size of these models and their limited access make improving their robustness a challenging task. Among various defense strategies, randomized smoothing has shown great potential for LLMs, as it does not require full access to the model’s parameters or fine-tuning via adversarial training. However, randomized smoothing involves adding noise to the input before model prediction, and the final model’s robustness largely depends on the model’s performance on these noise-corrupted data. Its effectiveness is often limited by the model’s sub-optimal performance on noisy data. To address this issue, we propose to leverage the multitasking nature of LLMs to first denoise the noisy inputs and then to make predictions based on these denoised versions. We call this procedure self-denoised smoothing. Unlike previous denoised smoothing techniques in computer vision, which require training a separate model to enhance the robustness of LLMs, our method offers significantly better efficiency and flexibility. Our experimental results indicate that our method surpasses existing methods in both empirical and certified robustness in defending against adversarial attacks for both downstream tasks and human alignments (i.e., jailbreak attacks). Our code is publicly available at https://github.com/UCSB-NLP-Chang/SelfDenoise.

When adapting Large Language Models for Recommendation (LLMRec), it is crucial to integrate collaborative information. Existing methods achieve this by learning collaborative embeddings in LLMs’ latent space from scratch or by mapping from external models. However, they fail to represent the information in a text-like format, which may not align optimally with LLMs. To bridge this gap, we introduce BinLLM, a novel LLMRec method that seamlessly integrates collaborative information through text-like encoding. BinLLM converts collaborative embeddings from external models into binary sequences — a specific text format that LLMs can understand and operate on directly, facilitating the direct usage of collaborative information in text-like format by LLMs. Additionally, BinLLM provides options to compress the binary sequence using dot-decimal notation to avoid excessively long lengths. Extensive experiments validate that BinLLM introduces collaborative information in a manner better aligned with LLMs, resulting in enhanced performance. We release our code at https://github.com/zyang1580/BinLLM.

Some prior work has shown that LLMs perform well in NLG evaluation for different tasks. However, we discover that LLMs seem to confuse different evaluation criteria, which reduces their reliability. For further verification, we first consider avoiding issues of inconsistent conceptualization and vague expression in existing NLG quality criteria themselves. So we summarize a clear hierarchical classification system for 11 common aspects with corresponding different criteria from previous studies involved. Inspired by behavioral testing, we elaborately design 18 types of aspect-targeted perturbation attacks for fine-grained analysis of the evaluation behaviors of different LLMs. We also conduct human annotations beyond the guidance of the classification system to validate the impact of the perturbations. Our experimental results reveal confusion issues inherent in LLMs, as well as other noteworthy phenomena, and necessitate further research and improvements for LLM-based evaluation.

The rapid propagation of misinformation poses substantial risks to public interest. To combat misinformation, large language models (LLMs) are adapted to automatically verify claim credibility. Nevertheless, existing methods heavily rely on the embedded knowledge within LLMs and / or black-box APIs for evidence collection, leading to subpar performance with smaller LLMs or upon unreliable context. In this paper, we propose retrieval augmented fact verification through the synthesis of contrasting arguments (RAFTS). Upon input claims, RAFTS starts with evidence retrieval, where we design a retrieval pipeline to collect and re-rank relevant documents from verifiable sources. Then, RAFTS forms contrastive arguments (i.e., supporting or refuting) conditioned on the retrieved evidence. In addition, RAFTS leverages an embedding model to identify informative demonstrations, followed by in-context prompting to generate the prediction and explanation. Our method effectively retrieves relevant documents as evidence and evaluates arguments from varying perspectives, incorporating nuanced information for fine-grained decision-making. Combined with informative in-context examples as prior, RAFTS achieves significant improvements to supervised and LLM baselines without complex prompts. We demonstrate the effectiveness of our method through extensive experiments, where RAFTS can outperform GPT-based methods with a significantly smaller 7B LLM.

With emerging topics (e.g., COVID-19) on social media as a source for the spreading misinformation, overcoming the distributional shifts between the original training domain (i.e., source domain) and such target domains remains a non-trivial task for misinformation detection. This presents an elusive challenge for early-stage misinformation detection, where a good amount of data and annotations from the target domain is not available for training. To address the data scarcity issue, we propose MetaAdapt, a meta learning based approach for domain adaptive few-shot misinformation detection. MetaAdapt leverages limited target examples to provide feedback and guide the knowledge transfer from the source to the target domain (i.e., learn to adapt). In particular, we train the initial model with multiple source tasks and compute their similarity scores to the meta task. Based on the similarity scores, we rescale the meta gradients to adaptively learn from the source tasks. As such, MetaAdapt can learn how to adapt the misinformation detection model and exploit the source data for improved performance in the target domain. To demonstrate the efficiency and effectiveness of our method, we perform extensive experiments to compare MetaAdapt with state-of-the-art baselines and large language models (LLMs) such as LLaMA, where MetaAdapt achieves better performance in domain adaptive few-shot misinformation detection with substantially reduced parameters on real-world datasets.

Prompt-based learning is vulnerable to backdoor attacks. Existing backdoor attacks against prompt-based models consider injecting backdoors into the entire embedding layers or word embedding vectors. Such attacks can be easily affected by retraining on downstream tasks and with different prompting strategies, limiting the transferability of backdoor attacks. In this work, we propose transferable backdoor attacks against prompt-based models, called NOTABLE, which is independent of downstream tasks and prompting strategies. Specifically, NOTABLE injects backdoors into the encoders of PLMs by utilizing an adaptive verbalizer to bind triggers to specific words (i.e., anchors). It activates the backdoor by pasting input with triggers to reach adversary-desired anchors, achieving independence from downstream tasks and prompting strategies. We conduct experiments on six NLP tasks, three popular models, and three prompting strategies. Empirical results show that NOTABLE achieves superior attack performance (i.e., attack success rate over 90% on all the datasets), and outperforms two state-of-the-art baselines. Evaluations on three defenses show the robustness of NOTABLE. Our code can be found at https://github.com/RU-System-Software-and-Security/Notable.

Temporal Logic (TL) can be used to rigorously specify complex high-level specification for systems in many engineering applications. The translation between natural language (NL) and TL has been under-explored due to the lack of dataset and generalizable model across different application domains. In this paper, we propose an accurate and generalizable transformation framework of English instructions from NL to TL, exploring the use of Large Language Models (LLMs) at multiple stages. Our contributions are twofold. First, we develop a framework to create a dataset of NL-TL pairs combining LLMs and human annotation. We publish a dataset with 23K NL-TL pairs. Then, we finetune T5 models on the lifted versions (i.e., the specific Atomic Propositions (AP) are hidden) of the NL and TL. The enhanced generalizability originates from two aspects: 1) Usage of lifted NL-TL characterizes common logical structures, without constraints of specific domains. 2) Application of LLMs in dataset creation largely enhances corpus richness. We test the generalization of trained models on five varied domains. To achieve full NL-TL transformation, we either combine the lifted model with AP recognition task or do the further finetuning on each specific domain. During the further finetuning, our model achieves higher accuracy (> 95%) using only <10% training data, compared with the baseline sequence to sequence (Seq2Seq) model.

We propose DiffCSE, an unsupervised contrastive learning framework for learning sentence embeddings. DiffCSE learns sentence embeddings that are sensitive to the difference between the original sentence and an edited sentence, where the edited sentence is obtained by stochastically masking out the original sentence and then sampling from a masked language model. We show that DiffSCE is an instance of equivariant contrastive learning, which generalizes contrastive learning and learns representations that are insensitive to certain types of augmentations and sensitive to other “harmful” types of augmentations. Our experiments show that DiffCSE achieves state-of-the-art results among unsupervised sentence representation learning methods, outperforming unsupervised SimCSE by 2.3 absolute points on semantic textual similarity tasks.

Based on an exponentially increasing number of academic articles, discovering and citing comprehensive and appropriate resources have become non-trivial tasks. Conventional citation recommendation methods suffer from severe information losses. For example, they do not consider the section header of the paper that the author is writing and for which they need to find a citation, the relatedness between the words in the local context (the text span that describes a citation), or the importance of each word from the local context. These shortcomings make such methods insufficient for recommending adequate citations to academic manuscripts. In this study, we propose a novel embedding-based neural network called dual attention model for citation recommendation (DACR) to recommend citations during manuscript preparation. Our method adapts the embedding of three semantic pieces of information: words in the local context, structural contexts,1 and the section on which the author is working. A neural network model is designed to maximize the similarity between the embedding of the three inputs (local context words, section headers, and structural contexts) and the target citation appearing in the context. The core of the neural network model comprises self-attention and additive attention; the former aims to capture the relatedness between the contextual words and structural context, and the latter aims to learn their importance. Recommendation experiments on real-world datasets demonstrate the effectiveness of the proposed approach. To seek explainability on DACR, particularly the two attention mechanisms, the learned weights from them are investigated to determine how the attention mechanisms interpret “relatedness” and “importance” through the learned weights. In addition, qualitative analyses were conducted to testify that DACR could find necessary citations that were not noticed by the authors in the past due to the limitations of the keyword-based searching.

Text-to-SQL systems offers natural language interfaces to databases, which can automatically generates SQL queries given natural language questions. On the WikiSQL benchmark, state-of- the-art text-to-SQL systems typically take a slot-filling approach by building several specialized models for each type of slot. Despite being effective, such modularized systems are complex and also fall short in jointly learning for different slots. To solve these problems, this paper proposes a novel approach that formulates the task as a question answering problem, where different slots are predicted by a unified machine reading comprehension (MRC) model. For this purpose, we use a BERT-based MRC model, which can also benefit from intermediate training on other MRC datasets. The proposed method can achieve competitive results on WikiSQL, suggesting it being a promising direction for text-to-SQL.

Based on an exponentially increasing number of academic articles, discovering and citing comprehensive and appropriate resources has become a non-trivial task. Conventional citation recommender methods suffer from severe information loss. For example, they do not consider the section of the paper that the user is writing and for which they need to find a citation, the relatedness between the words in the local context (the text span that describes a citation), or the importance on each word from the local context. These shortcomings make such methods insufficient for recommending adequate citations to academic manuscripts. In this study, we propose a novel embedding-based neural network called “dual attention model for citation recommendation (DACR)” to recommend citations during manuscript preparation. Our method adapts embedding of three semantic information: words in the local context, structural contexts, and the section on which a user is working. A neural network model is designed to maximize the similarity between the embedding of the three input (local context words, section and structural contexts) and the target citation appearing in the context. The core of the neural network model is composed of self-attention and additive attention, where the former aims to capture the relatedness between the contextual words and structural context, and the latter aims to learn the importance of them. The experiments on real-world datasets demonstrate the effectiveness of the proposed approach.

Recently, pre-trained language models such as BERT have shown state-of-the-art accuracies in text matching. When being applied to IR (or QA), the BERT-based matching models need to online calculate the representations and interactions for all query-candidate pairs. The high inference cost has prohibited the deployments of BERT-based matching models in many practical applications. To address this issue, we propose a novel BERT-based text matching model, in which the representations and the interactions are decoupled. Then, the representations of the candidates can be calculated and stored offline, and directly retrieved during the online matching phase. To conduct the interactions and generate final matching scores, a lightweight attention network is designed. Experiments based on several large scale text matching datasets show that the proposed model, called FASTMATCH, can achieve up to 100X speed-up to BERT and RoBERTa at the online matching phase, while keeping more up to 98.7% of the performance.

There has been an influx of biomedical domain-specific language models, showing language models pre-trained on biomedical text perform better on biomedical domain benchmarks than those trained on general domain text corpora such as Wikipedia and Books. Yet, most works do not study the factors affecting each domain language application deeply. Additionally, the study of model size on domain-specific models has been mostly missing. We empirically study and evaluate several factors that can affect performance on domain language applications, such as the sub-word vocabulary set, model size, pre-training corpus, and domain transfer. We show consistent improvements on benchmarks with our larger BioMegatron model trained on a larger domain corpus, contributing to our understanding of domain language model applications. We demonstrate noticeable improvements over the previous state-of-the-art (SOTA) on standard biomedical NLP benchmarks of question answering, named entity recognition, and relation extraction. Code and checkpoints to reproduce our experiments are available at [github.com/NVIDIA/NeMo].

On the WikiSQL benchmark, state-of-the-art text-to-SQL systems typically take a slot- filling approach by building several dedicated models for each type of slots. Such modularized systems are not only complex but also of limited capacity for capturing inter-dependencies among SQL clauses. To solve these problems, this paper proposes a novel extraction-linking approach, where a unified extractor recognizes all types of slot mentions appearing in the question sentence before a linker maps the recognized columns to the table schema to generate executable SQL queries. Trained with automatically generated annotations, the proposed method achieves the first place on the WikiSQL benchmark.

Selective rationalization has become a common mechanism to ensure that predictive models reveal how they use any available features. The selection may be soft or hard, and identifies a subset of input features relevant for prediction. The setup can be viewed as a co-operate game between the selector (aka rationale generator) and the predictor making use of only the selected features. The co-operative setting may, however, be compromised for two reasons. First, the generator typically has no direct access to the outcome it aims to justify, resulting in poor performance. Second, there’s typically no control exerted on the information left outside the selection. We revise the overall co-operative framework to address these challenges. We introduce an introspective model which explicitly predicts and incorporates the outcome into the selection process. Moreover, we explicitly control the rationale complement via an adversary so as not to leave any useful information out of the selection. We show that the two complementary mechanisms maintain both high predictive accuracy and lead to comprehensive rationales.

Representation learning is a key issue for most Natural Language Processing (NLP) tasks. Most existing representation models either learn little structure information or just rely on pre-defined structures, leading to degradation of performance and generalization capability. This paper focuses on learning both local semantic and global structure representations for text classification. In detail, we propose a novel Sandwich Neural Network (SNN) to learn semantic and structure representations automatically without relying on parsers. More importantly, semantic and structure information contribute unequally to the text representation at corpus and instance level. To solve the fusion problem, we propose two strategies: Adaptive Learning Sandwich Neural Network (AL-SNN) and Self-Attention Sandwich Neural Network (SA-SNN). The former learns the weights at corpus level, and the latter further combines attention mechanism to assign the weights at instance level. Experimental results demonstrate that our approach achieves competitive performance on several text classification tasks, including sentiment analysis, question type classification and subjectivity classification. Specifically, the accuracies are MR (82.1%), SST-5 (50.4%), TREC (96%) and SUBJ (93.9%).

Although many sentiment lexicons in different languages exist, most are not comprehensive. In a recent sentiment analysis application, we used a large Chinese sentiment lexicon and found that it missed a large number of sentiment words in social media. This prompted us to make a new attempt to study sentiment lexicon expansion. This paper first poses the problem as a PU learning problem, which is a new formulation. It then proposes a new PU learning method suitable for our problem using a neural network. The results are enhanced further with a new dictionary-based technique and a novel polarity classification technique. Experimental results show that the proposed approach outperforms baseline methods greatly.

Neural models have recently shown big improvements in the performance of phrase-based machine translation. Recurrent language models, in particular, have been a great success due to their ability to model arbitrary long context. In this work, we integrate global semantic information extracted from large encyclopedic sources into neural network language models. We integrate semantic word classes extracted from Wikipedia and sentence level topic information into a recurrent neural network-based language model. The new resulting models exhibit great potential in alleviating data sparsity problems with the additional knowledge provided. This approach of integrating global information is not restricted to language modeling but can also be easily applied to any model that profits from context or further data resources, e.g. neural machine translation. Using this model has improved rescoring quality of a state-of-the-art phrase-based translation system by 0.84 BLEU points. We performed experiments on two language pairs.