论文检索

Arthur Scalercio, Maria Finatto, Aline Paes

Automatic text simplification focuses on transforming texts into a more comprehensible version without sacrificing their precision. However, automatic methods usually require (paired) datasets that can be rather scarce in languages other than English. This paper presents a new approach to automatic sentence simplification that leverages paraphrases, context, and linguistic attributes to overcome the absence of paired texts in Portuguese.We frame the simplification problem as a textual style transfer task and learn a style representation using the sentences around the target sentence in the document and its linguistic attributes. Moreover, unlike most unsupervised approaches that require style-labeled training data, we fine-tune strong pre-trained models using sentence-level paraphrases instead of annotated data. Our experiments show that our model achieves remarkable results, surpassing the current state-of-the-art (BART+ACCESS) while competitively matching a Large Language Model.

Ansh Arora, Xuanli He, Maximilian Mozes, Srinibas Swain, Mark Dras, Qiongkai Xu

The democratization of pre-trained language models through open-source initiatives has rapidly advanced innovation and expanded access to cutting-edge technologies. However, this openness also brings significant security risks, including backdoor attacks, where hidden malicious behaviors are triggered by specific inputs, compromising natural language processing (NLP) system integrity and reliability. This paper suggests that merging a backdoored model with other homogeneous models can significantly remediate backdoor vulnerabilities even if such models are not entirely secure. In our experiments, we verify our hypothesis on various models (BERT-Base, RoBERTa-Large, Llama2-7B, and Mistral-7B) and datasets (SST-2, OLID, AG News, and QNLI). Compared to multiple advanced defensive approaches, our method offers an effective and efficient inference-stage defense against backdoor attacks on classification and instruction-tuned tasks without additional resources or specific knowledge. Our approach consistently outperforms recent advanced baselines, leading to an average of about 75% reduction in the attack success rate. Since model merging has been an established approach for improving model performance, the extra advantage it provides regarding defense can be seen as a cost-free bonus.

Beyza Ermis, Luiza Pozzobon, Sara Hooker, Patrick Lewis

To date, toxicity mitigation in language models has almost entirely been focused on single-language settings. As language models embrace multilingual capabilities, it’s crucial our safety measures keep pace. Recognizing this research gap, our approach expands the scope of conventional toxicity mitigation to address the complexities presented by multiple languages. In the absence of sufficient annotated datasets across languages, we employ translated data to evaluate and enhance our mitigation techniques. We also compare finetuning mitigation approaches against retrieval-augmented techniques under both static and continual toxicity mitigation scenarios. This allows us to examine the effects of translation quality and the cross-lingual transfer on toxicity mitigation. We also explore how model size and data quantity affect the success of these mitigation efforts. Covering nine languages, our study represents a broad array of linguistic families and levels of resource availability, ranging from high to mid-resource languages. Through comprehensive experiments, we provide insights into the complexities of multilingual toxicity mitigation, offering valuable insights and paving the way for future research in this increasingly important field.

Bowen Zhao, Zander Brumbaugh, Yizhong Wang, Hannaneh Hajishirzi, Noah Smith

Language models (LMs) are trained on web text originating from many points in time and, in general, without any explicit temporal grounding. This work investigates the temporal chaos of pretrained LMs and explores various methods to align their internal knowledge to a target time, which we call “temporal alignment.” To do this, we first automatically construct a dataset containing 20K time-sensitive questions and their answers for each year from 2000 to 2023. Based on this dataset, we empirically show that pretrained LMs (e.g., LLaMa2), despite having a recent pretraining cutoff (e.g., 2022), mostly answer questions using earlier knowledge (e.g., in 2019). We then develop several methods, from prompting to finetuning, to align LMs to use their most recent knowledge when answering questions, and investigate various factors in this alignment. Our experiments demonstrate that aligning LLaMa2 to the year 2022 can enhance its performance by up to 62% according to that year’s answers. This improvement occurs even without explicitly mentioning time information, indicating the possibility of aligning models’ internal sense of time after pretraining. Finally, we find that alignment to a historical time is also possible, with up to 2.8\times the performance of the unaligned LM in 2010 if finetuning models to that year. These findings hint at the sophistication of LMs’ internal knowledge organization and the necessity of tuning them properly.

Sarkar Snigdha Sarathi Das, Chirag Shah, Mengting Wan, Jennifer Neville, Longqi Yang, Reid Andersen, Georg Buscher, Tara Safavi

Traditional Dialogue State Tracking (DST) has focused on tracking preferences and intents in conversations centered around specific tasks (e.g. booking services). These conventional systems assume a relatively restricted conversation flow in which each turn gradually offers new information. However, advancements in Large Language Models (LLMs) have ushered in more versatile open-domain chat systems in which extended dialogue sessions encompassing numerous tasks and topics are common—in turn requiring new conversational tracking tools in order to successfully orchestrate such systems. Addressing these challenges, we introduce a novel approach combining dialogue segmentation and state tracking within open-domain dialogues, tailored for zero-shot applications appropriate to a true open-domain dialogue system. Our proposed method S3-DST employs a unique structured prompting technique and *Pre-Analytical Recollection*, a novel grounding mechanism we designed for improving long context tracking. Tested on proprietary anonymized open-domain dialogue datasets as well as publicly available DST and segmentation datasets, S3-DST consistently outperforms the state-of-the-art, showcasing its effectiveness and adaptability state tracking in the next wave of LLM-based chat systems. We also release S3-DST annotations with GPT-4 on a curated subset of LMSYS-Chat-1M to be used as a testbed to fuel research in this direction.

Cheng-Yu Hsieh, Yung-Sung Chuang, Chun-Liang Li, Zifeng Wang, Long Le, Abhishek Kumar, James Glass, Alexander Ratner, Chen-Yu Lee, Ranjay Krishna 等

Large language models (LLMs), even when specifically trained to process long input contexts, struggle to capture relevant information located in the middle of their input. This phenomenon has been known as the lost-in-the-middle problem. In this work, we make three contributions. First, we set out to understand the factors that cause this phenomenon. In doing so, we establish a connection between lost-in-the-middle to LLMs’ intrinsic attention bias: LLMs exhibit an U-shaped attention bias where the tokens at the beginning and at the end of its input receive higher attention, regardless of their relevance. Second, we mitigate this positional bias through a calibration mechanism, found-in-the-middle, that allows the model to attend to contexts faithfully according to their relevance, even though when they are in the middle. Third, we show found-in-the-middle not only achieves better performance in locating relevant information within a long context, but also eventually leads to improved retrieval-augmented generation (RAG) performance across various tasks, outperforming existing methods by up to 10 percentage point. These findings open up future directions in understanding LLM attention bias and its potential consequences.

Eunseop Yoon, Hee Suk Yoon, SooHwan Eom, Gunsoo Han, Daniel Nam, Daejin Jo, Kyoung-Woon On, Mark Hasegawa-Johnson, Sungwoong Kim, Chang Yoo

Reinforcement Learning from Human Feedback (RLHF) leverages human preference data to train language models to align more closely with human essence. These human preference data, however, are labeled at the sequence level, creating a mismatch between sequence-level preference labels and tokens, which are autoregressively generated from the language model. Although several recent approaches have tried to provide token-level (i.e., dense) rewards for each individual token, these typically rely on predefined discrete reward values (e.g., positive: +1, negative: -1, neutral: 0), failing to account for varying degrees of preference inherent to each token. To address this limitation, we introduce TLCR (Token-Level Continuous Reward) for RLHF, which incorporates a discriminator trained to distinguish positive and negative tokens, and the confidence of the discriminator is used to assign continuous rewards to each token considering the context. Extensive experiments show that our proposed TLCR leads to consistent performance improvements over previous sequence-level or token-level discrete rewards on open-ended generation benchmarks.

XiaoQi Han, Ru Li, Xiaoli Li, Jiye Liang, Zifang Zhang, Jeff Pan

The task of model editing becomes popular for correcting inaccurate or outdated parametric knowledge in Large Language Models (LLMs). However, there are major limitations of state of the art (SOTA) model editing methods, including the excessive memorization issue caused by the direct editing methods, as well as the error propagation and knowledge conflict issues from the memory enhancement methods, resulting in hindering models’ *portability*, e.g., the ability to transfer the new knowledge to related one-hop or multi-hop content. To address these issues, we propose the InstructEd method, the idea of which is to insert soft instructions into the attention module so as to facilitate interactions between instructions and questions and to understand and utilize new facts. Our main findings are: (i) InstructEd has achieved SOTA performance on three datasets for one-hop/multi-hop evaluation with LLaMAs and GPT2, achieving 10% (5%) improvement in one-hop (multi-hop) model editing.(ii) Different from earlier methods on editing parameters in FFN, we show that editing attention can also help. (iii) Model editing is highly related to retrieval augmented methods, which can help improve the locality of model editing while slightly decrease the editing performance with hops.

Justin Lovelace, Varsha Kishore, Yiwei Chen, Kilian Weinberger

Current language models demonstrate remarkable proficiency in text generation. However, for many applications it is desirable to control attributes, such as sentiment, or toxicity, of the generated language—ideally tailored towards each specific use case and target audience. For auto-regressive language models, existing guidance methods are prone to decoding errors that cascade during generation and degrade performance. In contrast, text diffusion models can easily be guided with, for example, a simple linear sentiment classifier—however they do suffer from significantly higher perplexity than auto-regressive alternatives. In this paper we use a guided diffusion model to produce a latent proposal that steers an auto-regressive language model to generate text with desired properties. Our model inherits the unmatched fluency of the auto-regressive approach and the plug-and-play flexibility of diffusion. We show that it outperforms previous plug-and-play guidance methods across a wide range of benchmark data sets. Further, controlling a new attribute in our framework is reduced to training a single logistic regression classifier.

Yifei Li, Xiang Yue, Zeyi Liao, Huan Sun

Modern generative search engines enhance the reliability of large language model (LLM) responses by providing cited evidence. However, evaluating the answer’s attribution, i.e., whether every claim within the generated responses is fully supported by its cited evidence, remains an open problem. This verification, traditionally dependent on costly human evaluation, underscores the urgent need for automatic attribution evaluation methods. To bridge the gap in the absence of standardized benchmarks for these methods, we present AttributionBench, a comprehensive benchmark compiled from various existing attribution datasets. Our extensive experiments on AttributionBench reveal the challenges of automatic attribution evaluation, even for state-of-the-art LLMs. Specifically, our findings show that even a fine-tuned GPT-3.5 only achieves around 80% macro-F1 under a binary classification formulation. A detailed analysis of more than 300 error cases indicates that a majority of failures stem from the model’s inability to process nuanced information, and the discrepancy between the information the model has access to and that human annotators do.

Yunye Gong, Robik Shrestha, Jared Claypoole, Michael Cogswell, Arijit Ray, Christopher Kanan, Ajay Divakaran

We propose a novel VQA dataset, BloomVQA, to facilitate comprehensive evaluation of large vision-language models on comprehension tasks. Unlike current benchmarks that often focus on fact-based memorization and simple reasoning tasks without theoretical grounding, we collect multiple-choice samples based on picture stories that reflect different levels of comprehension, as laid out in Bloom’s Taxonomy, a classic framework for learning assessment widely adopted in education research. Our data maps to a novel hierarchical graph representation which enables automatic data augmentation and novel measures characterizing model consistency. We perform graded evaluation and reliability analysis on recent multi-modal models. In comparison to low-level tasks, we observe decreased performance on tasks requiring advanced comprehension and cognitive skills with up to 38.0% drop in VQA accuracy. In comparison to earlier models, GPT-4V demonstrates improved accuracy over all comprehension levels and also shows a tendency of bypassing visual inputs especially for higher-level tasks. Current models also show consistency patterns misaligned with human comprehension in various scenarios, demonstrating the need for improvement based on theoretically-grounded criteria. The dataset can be accessed at https://huggingface.co/datasets/ygong/BloomVQA.

Qi Guo, Leiyu Wang, Yidong Wang, Wei Ye, Shikun Zhang

Although large language models (LLMs) have demonstrated impressive few-shot learning capabilities via in-context learning (ICL), ICL performance is known to be highly sensitive to the order of examples provided. To identify appropriate orders, recent studies propose heuristic methods to evaluate order performance using a set of unlabeled data. However, the requirement of in-domain data limits their utility in real-world scenarios where additional annotated data is challenging to acquire. Additionally, these dataset-based approaches are prone to being sub-optimal for a lack of consideration for individual differences. To address the problems, we first analyze the properties of performant example orders at both corpus level and instance level. Based on the analysis we propose **DEmO** to adaptively identify performant example order for each instance without extra data. DEmO works by filtering out a subset of orders featuring label fairness, then selecting the most influential order for each test instance. The employment of a content-free metric makes DEmO independent of in-domain data. Extensive experiments indicate the superiority of DEmO over a wide range of strong baselines. Further analysis validates the generalizability across various settings.

David Mueller, Mark Dredze, Nicholas Andrews

Instruction-tuning trains a language model on hundreds of tasks jointly to improve a model’s ability to learn in-context;however, the mechanisms that drive in-context learning are poorly understood and, as a result, the role of instruction-tuning on in-context generalization is poorly understood as well.In this work, we study the impact of instruction-tuning on multi-task transfer: how well a model’s parameters adapt to an unseen task via fine-tuning.We find that instruction-tuning negatively impacts a model’s transfer to unseen tasks, and that model transfer and in-context generalization are highly correlated, suggesting that this catastrophic forgetting may impact in-context learning.We study methods to improve model transfer, finding that multi-task training—how well the training tasks are optimized—can significantly impact ICL generalization; additionally, we find that continual training on unsupervised pre-training data can mitigate forgetting and improve ICL generalization as well.Finally, we demonstrate that, early into training, the impact of instruction-tuning on model transfer to tasks impacts in-context generalization on that task.Overall, we provide significant evidence that multi-task transfer is deeply connected to a model’s ability to learn a task in-context.

Xuxin Cheng, Zhihong Zhu, Xianwei Zhuang, Zhanpeng Chen, Zhiqi Huang, Yuexian Zou

As a crucial task in the task-oriented dialogue systems, spoken language understanding (SLU) has garnered increasing attention. However, errors from automatic speech recognition (ASR) often hinder the performance of understanding. To tackle this problem, we propose MoE-SLU, an ASR-Robust SLU framework based on the mixture-of-experts technique. Specifically, we first introduce three strategies to generate additional transcripts from clean transcripts. Then, we employ the mixture-of-experts technique to weigh the representations of the generated transcripts, ASR transcripts, and the corresponding clean manual transcripts. Additionally, we also regularize the weighted average of predictions and the predictions of ASR transcripts by minimizing the Jensen-Shannon Divergence (JSD) between these two output distributions. Experiment results on three benchmark SLU datasets demonstrate that our MoE-SLU achieves state-of-the-art performance. Further model analysis also verifies the superiority of our method.

Divyanshu Aggarwal, Ashutosh Sathe, Ishaan Watts, Sunayana Sitaram

Parameter efficient finetuning has emerged as a viable solution for improving the performance of Large Language Models without requiring massive resources and compute. Prior work on multilingual evaluation has shown that there is a large gap between the performance of LLMs on English and other languages. Further, there is also a large gap between the performance of smaller open-source models and larger LLMs. Finetuning can be an effective way to bridge this gap and make language models more equitable. In this work, we finetune the Llama-2 and Mistral models on two synthetic multilingual instruction tuning datasets to determine its effect on model performance on six downstream tasks covering forty one languages in all. Additionally, we experiment with various parameters, such as rank for low-rank adaptation and values of quantisation to determine their effects on downstream performance and find that higher rank and higher quantisation values benefit low-resource languages. We find that parameter efficient finetuning of smaller open-source models sometimes bridges the gap between the performance of these models and the larger ones, however, English performance can take a hit. We also find that finetuning sometimes improves performance on low-resource languages, while degrading performance on high-resource languages.

Adil Soubki, John Murzaku, Arash Yousefi Jordehi, Peter Zeng, Magdalena Markowska, Seyed Abolghasem Mirroshandel, Owen Rambow

Evaluating the theory of mind (ToM) capabilities of language models (LMs) has recently received a great deal of attention. However, many existing benchmarks rely on synthetic data, which risks misaligning the resulting experiments with human behavior. We introduce the first ToM dataset based on naturally occurring spoken dialogs, Common-ToM, and show that LMs struggle to demonstrate ToM. We then show that integrating a simple, explicit representation of beliefs improves LM performance on Common-ToM.

Zixia Jia, Mengmeng Wang, Baichen Tong, Song-Chun Zhu, Zilong Zheng

Recent advances in Large Language Models (LLMs) have shown inspiring achievements in constructing autonomous agents that rely onlanguage descriptions as inputs. However, it remains unclear how well LLMs can function as few-shot or zero-shot embodied agents in dynamic interactive environments. To address this gap, we introduce LangSuit·E, a versatile and simulation-free testbed featuring 6 representative embodied tasks in textual embodied worlds. Compared with previous LLM-based testbeds, LangSuit·E (i) offers adaptability to diverse environments without multiple simulation engines, (ii) evaluates agents’ capacity to develop “internalized world knowledge” with embodied observations, and (iii) allows easy customization of communication and action strategies. To address the embodiment challenge, we devise a novel chain-of-thought (CoT) schema, EmMem, which summarizes embodied states w.r.t. history information. Comprehensive benchmark results illustrate challenges and insights of embodied planning. LangSuit·E represents a significant step toward building embodied generalists in the context of language models.

Noah Wang, Z.Y. Peng, Haoran Que, Jiaheng Liu, Wangchunshu Zhou, Yuhan Wu, Hongcheng Guo, Ruitong Gan, Zehao Ni, Jian Yang 等

The advent of Large Language Models (LLMs) has paved the way for complex tasks such as role-playing, which enhances user interactions by enabling models to imitate various characters. However, the closed-source nature of state-of-the-art LLMs and their general-purpose training limit role-playing optimization. In this paper, we introduce RoleLLM, a framework to benchmark, elicit, and enhance role-playing abilities in LLMs. RoleLLM comprises four stages: (1) Role Profile Construction for 100 roles; (2) Context-Based Instruction Generation (Context-Instruct) for role-specific knowledge extraction; (3) Role Prompting using GPT (RoleGPT) for speaking style imitation; and (4) Role-Conditioned Instruction Tuning (RoCIT) for fine-tuning open-source models along with role customization. By Context-Instruct and RoleGPT, we create RoleBench, the first systematic and fine-grained character-level benchmark dataset for role-playing with 168,093 samples. Moreover, RoCIT on RoleBench yields RoleLLaMA (English) and RoleGLM (Chinese), significantly enhancing role-playing abilities and even achieving comparable results with RoleGPT (using GPT-4).

Jaap Jumelet, Willem Zuidema, Arabella Sinclair

We explore which linguistic factors—at the sentence and token level—play an important role in influencing language model predictions, and investigate whether these are reflective of results found in humans and human corpora (Gries and Kootstra, 2017). We make use of the structural priming paradigm—where recent exposure to a structure facilitates processing of the same structure—to investigate where priming effects manifest, and what factors predict them. We find these effects can be explained via the inverse frequency effect found in human priming, where rarer elements within a prime increase priming effects, as well as lexical dependence between prime and target. Our results provide an important piece in the puzzle of understanding how properties within their context affect structural prediction in language models.

Rongwu Xu, Zehan Qi, Wei Xu

Large language models (LLMs) showcase impressive reasoning capabilities when coupled with Chain-of-Thought (CoT) prompting. However, the robustness of this approach warrants further investigation. In this paper, we introduce a novel scenario termed preemptive answers, where the LLM obtains an answer before engaging in reasoning. This situation can arise inadvertently or induced by malicious users by prompt injection attacks. Experiments reveal that preemptive answers significantly impair the model’s reasoning capability across various CoT methods and a broad spectrum of datasets. To bolster the robustness of reasoning, we propose two measures aimed at mitigating this issue to some extent.