ACL 2024 Papers

Language agents have achieved considerable performance on various complex question-answering tasks by planning with external tools. Despite the incessant exploration in this field, existing language agent systems still struggle with costly, non-reproducible data reliance and face the challenge of compelling a single model for multiple functions. To this end, we introduce AutoAct, an automatic agent learning framework for QA that does not rely on large-scale annotated data and synthetic planning trajectories from closed-source models (e.g., GPT-4). Given limited data with a tool library, AutoAct first automatically synthesizes planning trajectories without any assistance from humans or strong closed-source models. Then, AutoAct leverages a division-of-labor strategy to automatically differentiate based on the target task information and synthesized trajectories, producing a sub-agent group to complete the task. We conduct comprehensive experiments with different LLMs, which demonstrates that AutoAct yields better or parallel performance compared to various strong baselines. Further analysis demonstrates the effectiveness of the division-of-labor strategy, with the trajectory quality generated by AutoAct generally outperforming that of others.

Model interpretability in toxicity detection greatly profits from token-level annotations. However, currently, such annotations are only available in English. We introduce a dataset annotated for offensive language detection sourced from a news forum, notable for its incorporation of the Austrian German dialect, comprising 4,562 user comments. In addition to binary offensiveness classification, we identify spans within each comment constituting vulgar language or representing targets of offensive statements. We evaluate fine-tuned Transformer models as well as large language models in a zero- and few-shot fashion. The results indicate that while fine-tuned models excel in detecting linguistic peculiarities such as vulgar dialect, large language models demonstrate superior performance in detecting offensiveness in AustroTox.

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.

Recent efforts to address hallucinations in Large Language Models (LLMs) have focused on attributed text generation, which supplements generated texts with citations of supporting sources for post-generation fact-checking and corrections. Yet, these citations often point to entire documents or paragraphs, burdening users with extensive verification work. In this paper, we introduce a locally-attributable text generation approach, prioritizing concise attributions. Our method, named “Attribute First, then Generate“, breaks down the conventional end-to-end generation process into three intuitive steps: content selection, sentence planning, and sequential sentence generation. By initially identifying relevant source segments (“select first“) and then conditioning the generation process on them (“then generate“), we ensure these segments also act as the output’s fine-grained attributions (“select“ becomes “attribute“). Tested on Multi-document Summarization and Long-form Question-answering, our method not only yields more concise citations than the baselines but also maintains - and in some cases enhances - both generation quality and attribution accuracy. Furthermore, it significantly reduces the time required for fact verification by human assessors.

The widespread use of Text-to-Image (T2I) models in content generation requires careful examination of their safety, including their robustness to adversarial attacks. Despite extensive research on adversarial attacks, the reasons for their effectiveness remain underexplored. This paper presents an empirical study on adversarial attacks against T2I models, focusing on analyzing factors associated with attack success rates (ASR). We introduce a new attack objective - entity swapping using adversarial suffixes and two gradient-based attack algorithms. Human and automatic evaluations reveal the asymmetric nature of ASRs on entity swap: for example, it is easier to replace “human” with “robot” in the prompt “a human dancing in the rain.” with an adversarial suffix, but the reverse replacement is significantly harder. We further propose probing metrics to establish indicative signals from the model’s beliefs to the adversarial ASR. We identify conditions that result in a success probability of 60% for adversarial attacks and others where this likelihood drops below 5%. The code and data are available at https://github.com/Patchwork53/AsymmetricAttack

This paper addresses the critical challenge of assessing the representativeness of news thumbnail images, which often serve as the first visual engagement for readers when an article is disseminated on social media. We focus on whether a news image represents the actors discussed in the news text. To serve the challenge, we introduce NewsTT, a manually annotated dataset of 1000 news thumbnail images and text pairs. We found that the pretrained vision and language models, such as BLIP-2, struggle with this task. Since news subjects frequently involve named entities or proper nouns, the pretrained models could have a limited capability to match news actors’ visual and textual appearances. We hypothesize that learning to contrast news text with its counterfactual, of which named entities are replaced, can enhance the cross-modal matching ability of vision and language models. We propose CFT-CLIP, a contrastive learning framework that updates vision and language bi-encoders according to the hypothesis. We found that our simple method can boost the performance for assessing news thumbnail representativeness, supporting our assumption. Code and data can be accessed at https://github.com/ssu-humane/news-images-acl24.

Social bias is shaped by the accumulation of social perceptions towards targets across various demographic identities. To fully understand such social bias in large language models (LLMs), it is essential to consider the composite of social perceptions from diverse perspectives among identities. Previous studies have either evaluated biases in LLMs by indirectly assessing the presence of sentiments towards demographic identities in the generated text or measuring the degree of alignment with given stereotypes. These methods have limitations in directly quantifying social biases at the level of distinct perspectives among identities. In this paper, we aim to investigate how social perceptions from various viewpoints contribute to the development of social bias in LLMs. To this end, we propose a novel strategy to intuitively quantify these social perceptions and suggest metrics that can evaluate the social biases within LLMs by aggregating diverse social perceptions. The experimental results show the quantitative demonstration of the social attitude in LLMs by examining social perception. The analysis we conducted shows that our proposed metrics capture the multi-dimensional aspects of social bias, enabling a fine-grained and comprehensive investigation of bias in LLMs.

We observe that current large language models often waver in their judgments when faced with follow-up questions, even if the original judgment was correct. This wavering presents a significant challenge for generating reliable responses and building user trust. To comprehensively assess this issue, we introduce a Follow-up Questioning Mechanism along with two metrics to quantify this inconsistency, confirming its widespread presence in current large language models. Furthermore, to mitigate this issue, we explore various prompting strategies for closed-source models, and develop a training-based framework Unwavering-FQ that teaches large language models to maintain their originally correct judgments through synthesized high-quality preference data. Our experimental results confirm the effectiveness of our framework and its ability to enhance the general capabilities of large language models.

Multiple-choice question answering (MCQA) is often used to evaluate large language models (LLMs). To see if MCQA assesses LLMs as intended, we probe if LLMs can perform MCQA with choices-only prompts, where models must select the correct answer only from the choices. In three MCQA datasets and four LLMs, this prompt bests a majority baseline in 11/12 cases, with up to 0.33 accuracy gain. To help explain this behavior, we conduct an in-depth, black-box analysis on memorization, choice dynamics, and question inference. Our key findings are threefold. First, we find no evidence that the choices-only accuracy stems from memorization alone. Second, priors over individual choices do not fully explain choices-only accuracy, hinting that LLMs use the group dynamics of choices. Third, LLMs have some ability to infer a relevant question from choices, and surprisingly can sometimes even match the original question. We hope to motivate the use of stronger baselines in MCQA benchmarks, the design of robust MCQA datasets, and further efforts to explain LLM decision-making.

Safety is critical to the usage of large language models (LLMs). Multiple techniques such as data filtering and supervised fine-tuning have been developed to strengthen LLM safety. However, currently known techniques presume that corpora used for safety alignment of LLMs are solely interpreted by semantics. This assumption, however, does not hold in real-world applications, which leads to severe vulnerabilities in LLMs. For example, users of forums often use ASCII art, a form of text-based art, to convey image information. In this paper, we propose a novel ASCII art-based jailbreak attack and introduce a comprehensive benchmark Vision-in-Text Challenge (ViTC) to evaluate the capabilities of LLMs in recognizing prompts that cannot be solely interpreted by semantics. We show that five SOTA LLMs (GPT-3.5, GPT-4, Gemini, Claude, and Llama2) struggle to recognize prompts provided in the form of ASCII art. Based on this observation, we develop the jailbreak attack ArtPrompt, which leverages the poor performance of LLMs in recognizing ASCII art to bypass safety measures and elicit undesired behaviors from LLMs. ArtPrompt only requires black-box access to the victim LLMs, making it a practical attack. We evaluate ArtPrompt on five SOTA LLMs, and show that ArtPrompt can effectively and efficiently induce undesired behaviors from all five LLMs.

Fine-grained control over large language models (LLMs) remains a significant challenge, hindering their adaptability to diverse user needs. While Reinforcement Learning from Human Feedback (RLHF) shows promise in aligning LLMs, its reliance on scalar rewards often limits its ability to capture diverse user preferences in real-world applications. To address this limitation, we introduce the Directional Preference Alignment (DPA) framework. Unlike the scalar-reward RLHF, DPA incorporates multi-objective reward modeling to represent diverse preference profiles. Additionally, DPA models user preferences as directions (i.e., unit vectors) in the reward space to achieve user-dependent preference control. Our method involves training a multi-objective reward model and then fine-tuning the LLM with a preference-conditioned variant of Rejection Sampling Finetuning (RSF), an RLHF method adopted by Llama 2. This method enjoys a better performance trade-off across various reward objectives. In comparison with the scalar-reward RLHF, DPA offers users intuitive control over LLM generation: they can arithmetically specify their desired trade-offs (e.g., more helpfulness with less verbosity). We also validate the effectiveness of DPA with real-world alignment experiments on Mistral-7B. Our method provides straightforward arithmetic control over the trade-off between helpfulness and verbosity while maintaining competitive performance with strong baselines such as Direct Preference Optimization (DPO).

This paper introduces a novel approach to analyzing the forward-looking statements in equity research reports by integrating argument mining with sentiment analysis. Recognizing the limitations of traditional models in capturing the nuances of future-oriented analysis, we propose a refined categorization of argument units into claims, premises, and scenarios, coupled with a unique sentiment analysis framework. Furthermore, we incorporate a temporal dimension to categorize the anticipated impact duration of market events. To facilitate this study, we present the Equity Argument Mining and Sentiment Analysis (Equity-AMSA) dataset. Our research investigates the extent to which detailed domain-specific annotations can be provided, the necessity of fine-grained human annotations in the era of large language models, and whether our proposed framework can improve performance in downstream tasks over traditional methods. Experimental results reveal the significance of manual annotations, especially for scenario identification and sentiment analysis. The study concludes that our annotation scheme and dataset contribute to a deeper understanding of forward-looking statements in equity research reports.

Event linking connects event mentions in text with relevant nodes in a knowledge base (KB). Prior research in event linking has mainly borrowed methods from entity linking, overlooking the distinct features of events. Compared to the extensively explored entity linking task, events have more complex structures and can be more effectively distinguished by examining their associated arguments. Moreover, the information-rich nature of events leads to the scarcity of event KBs. This emphasizes the need for event linking models to identify and classify event mentions not in the KB as “out-of-KB,” an area that has received limited attention. In this work, we tackle these challenges by introducing an argument-aware approach. First, we improve event linking models by augmenting input text with tagged event argument information, facilitating the recognition of key information about event mentions. Subsequently, to help the model handle “out-of-KB” scenarios, we synthesize out-of-KB training examples from in-KB instances through controlled manipulation of event arguments. Our experiment across two test datasets showed significant enhancements in both in-KB and out-of-KB scenarios, with a notable 22% improvement in out-of-KB evaluations.

Recent research on sequence labelling has been exploring different strategies to mitigate the lack of manually annotated data for the large majority of the world languages. Among others, the most successful approaches have been based on (i) the crosslingual transfer capabilities of multilingual pre-trained language models (model-transfer), (ii) data translation and label projection (data-transfer) and (iii), prompt-based learning by reusing the mask objective to exploit the few-shot capabilities of pre-trained language models (few-shot). Previous work seems to conclude that model-transfer outperform data-transfer methods and that few-shot techniques based on prompting are superior to updating the model’s weights via fine-tuning. In this paper we empirically demonstrate that, for Argument Mining, a sequence labelling task which requires the detection of long and complex discourse structures, previous insights on crosslingual transfer or few-shot learning do not apply. Contrary to previous work, we show that for Argument Mining data-transfer obtains better results than model-transfer and that fine-tuning outperforms few-shot methods. Regarding the former, the domain of the dataset used for data-transfer seems to be a deciding factor, while, for few-shot, the type of task (length and complexity of the sequence spans) and sampling method proves to be crucial.

Instruction-tuned Large Language Models (LLMs) excel at many tasks and will even explain their reasoning, so-called self-explanations. However, convincing and wrong self-explanations can lead to unsupported confidence in LLMs, thus increasing risk. Therefore, it’s important to measure if self-explanations truly reflect the model’s behavior. Such a measure is called interpretability-faithfulness and is challenging to perform since the ground truth is inaccessible, and many LLMs only have an inference API. To address this, we propose employing self-consistency checks to measure faithfulness. For example, if an LLM says a set of words is important for making a prediction, then it should not be able to make its prediction without these words. While self-consistency checks are a common approach to faithfulness, they have not previously been successfully applied to LLM self-explanations for counterfactual, feature attribution, and redaction explanations. Our results demonstrate that faithfulness is explanation, model, and task-dependent, showing self-explanations should not be trusted in general. For example, with sentiment classification, counterfactuals are more faithful for Llama2, feature attribution for Mistral, and redaction for Falcon 40B.

Although large language models (LLMs) acquire extensive world knowledge and some reasoning abilities, their proficiency in generating humorous sentences remains a challenge. Previous research has demonstrated that the humor generation capabilities of ChatGPT are confined to producing merely 25 unique jokes. In this work, we concentrate on endowing LLMs with the ability of generating puns, a particular category of humor by preference learning method. We propose a multi-stage curriculum preference learning framework to optimize both pun structure preferences and humor preferences. Specifically, we improve the Direct Preference Optimization (DPO) algorithm to address the challenge of multi-objective alignment problem. Besides, to facilitate further advancement in this field, we collect a Chinese Pun (ChinesePun) dataset, containing 2.1k puns and corresponding annotations. Experimental results on both Chinese and English benchmark datasets demonstrate that our method significantly outperforms all the baseline models.

Making inferences in text comprehension to understand the meaning is essential in language processing. This work studies the entailment verification (EV) problem of complex, multi-sentence premises requiring a system to make multiple inferences implicitly. Modern applications of EV in detecting inconsistent model-generated rationales require complex multi-hop reasoning. However, current textual inference datasets mostly contain short-sentence premises that partially focus on this. To address this, we compile an EV benchmark that includes datasets from three NLP domains (NLI, contextual QA, and rationales) containing multi-sentence premises. On benchmarking humans and LLMs, we find that LLMs are better than humans in multi-hop reasoning across extended contexts, while humans perform better in simple deductive reasoning tasks. We also finetune a Flan-T5 model for EV using two training objectives to obtain a strong open-source model that outperforms GPT-3.5 and rivals GPT-4. Finally, we use our finetuned model to filter out inconsistent model-generated rationales in self-consistency decoding, resulting in a 6% accuracy improvement on average across three MCQ datasets.

Recent scholarship on reasoning in LLMs has supplied evidence of impressive performance and flexible adaptation to machine generated or human critique. Nonmonotonic reasoning, crucial to human cognition for navigating the real world, remains a challenging, yet understudied task. In this work, we study nonmonotonic reasoning capabilities of seven state-of-the-art LLMs in one abstract and one commonsense reasoning task featuring generics, such as ‘Birds fly’, and exceptions, ‘Penguins don’t fly’ (see Fig. 1). While LLMs exhibit reasoning patterns in accordance with human nonmonotonic reasoning abilities, they fail to maintain stable beliefs on truth conditions of generics at the addition of supporting examples (‘Owls fly’) or unrelated information (‘Lions have manes’).Our findings highlight pitfalls in attributing human reasoning behaviours to LLMs as long as consistent reasoning remains elusive.

Quantitative reasoning is a critical skill to analyze data, yet the assessment of such ability remains limited. To address this gap, we introduce the Quantitative Reasoning with Data (QRData) benchmark, aiming to evaluate Large Language Models’ capability in statistical and causal reasoning with real-world data. The benchmark comprises a carefully constructed dataset of 411 questions accompanied by data sheets from textbooks, online learning materials, and academic papers. To compare models’ quantitative reasoning abilities on data and text, we enrich the benchmark with an auxiliary set of 290 text-only questions, namely QRText. We evaluate natural language reasoning, program-based reasoning, and agent reasoning methods including Chain-of-Thought, Program-of-Thoughts, ReAct, and code interpreter assistants on diverse models. The strongest model GPT-4 achieves an accuracy of 58%, which has much room for improvement. Among open-source models, Deepseek-coder-instruct, a code LLM pretrained on 2T tokens, gets the highest accuracy of 37%. Analysis reveals that models encounter difficulties in data analysis and causal reasoning, and struggle in using causal knowledge and provided data simultaneously. Code and data are in https://github.com/xxxiaol/QRData.