ACL 2024 Papers

Scientific document summarization aims to condense complex and long articles in both technical and plain-language terms to facilitate the accessibility and dissemination of scientific findings. Existing datasets suffer from a deficiency in source heterogeneity, as their data predominantly stem from a single common resource, hindering effective model training and generalizability. First, we introduce SciLay, a novel dataset that includes documents from multiple natural science journals with expert-authored technical and lay summaries. Second, we propose PrunePert, a new transformer-based model that incorporates a differentiable perturbed top-k encoder layer to prune irrelevant tokens in end-to-end learning. Experimental results show that our model achieves a nearly 2x speed-up compared to a state-of-the-art linear transformer, remaining comparable in effectiveness. Additional examinations underscore the importance of employing a training dataset that includes different sources to enhance the generalizability of the models. Code is available at https://github.com/disi-unibo-nlp/sci-lay.

The rapid advancement of large language models (LLMs) has led to a new era marked by the development of autonomous applications in real-world scenarios, which drives innovation in creating advanced web agents. Existing web agents typically only handle one input modality and are evaluated only in simplified web simulators or static web snapshots, greatly limiting their applicability in real-world scenarios. To bridge this gap, we introduce WebVoyager, an innovative Large Multimodal Model (LMM) powered web agent that can complete user instructions end-to-end by interacting with real-world websites. Moreover, we establish a new benchmark by compiling real-world tasks from 15 popular websites and introduce an automatic evaluation protocol leveraging multimodal understanding abilities of GPT-4V to evaluate open-ended web agents. We show that WebVoyager achieves a 59.1% task success rate on our benchmark, significantly surpassing the performance of both GPT-4 (All Tools) and the WebVoyager (text-only) setups, underscoring the exceptional capability of WebVoyager. The proposed automatic evaluation metric achieves 85.3% agreement with human judgment, indicating its effectiveness in providing reliable and accurate assessments of web agents.

Enhancing the attribution in large language models (LLMs) is a crucial task. One feasible approach is to enable LLMs to cite external sources that support their generations. However, existing datasets and evaluation methods in this domain still exhibit notable limitations. In this work, we formulate the task of attributed query-focused summarization (AQFS) and present WebCiteS, a Chinese dataset featuring 7k human-annotated summaries with citations. WebCiteS derives from real-world user queries and web search results, offering a valuable resource for model training and evaluation. Prior works in attribution evaluation do not differentiate between groundedness errors and citation errors. They also fall short in automatically verifying sentences that draw partial support from multiple sources. We tackle these issues by developing detailed metrics and enabling the automatic evaluator to decompose the sentences into sub-claims for fine-grained verification. Our comprehensive evaluation of both open-source and proprietary models on WebCiteS highlights the challenge LLMs face in correctly citing sources, underscoring the necessity for further improvement. The dataset and code will be open-sourced to facilitate further research in this crucial field.

Recent work demonstrates that, after instruction tuning, Code Large Language Models (Code LLMs) can obtain impressive capabilities to address a wide range of code-related tasks. However, current instruction tuning methods for Code LLMs mainly focus on the traditional code generation task, resulting in poor performance in complex multi-task scenarios. In this paper, we concentrate on multiple code-related tasks and present WaveCoder, a series of Code LLMs trained with Widespread And Versatile Enhanced instruction data. To enable the models to tackle complex code-related tasks, we propose a method to stably generate diverse, high-quality instruction data from open source code dataset in multi-task scenarios and obtain CodeOcean, a dataset comprising 19,915 instruction instances across 4 code-related tasks, which is aimed at improving the generalization ability of Code LLM. Our experiments demonstrate that WaveCoder models significantly outperform other open-source models in terms of the generalization ability across different code-related tasks. Moreover, WaveCoder-Ultra-6.7B presents the state-of-the-art generalization abilities on a wide range of code-related tasks.

We release a multi-accent dataset and propose speech-programming and gradient reversal classifier to improve the generalization.Abstract: Speech-to-SQL (S2SQL) aims to convert spoken questions into SQL queries given relational databases, which has been traditionally implemented in a cascaded manner while facing the following challenges: 1) model training is faced with the major issue of data scarcity, where limited parallel data is available; and 2) the systems should be robust enough to handle diverse out-of-domain speech samples that differ from the source data. In this work, we propose the direct generalizable speech-to-SQL parsing model Wav2SQL which avoids error compounding across cascaded systems. Specifically, 1) to accelerate speech-driven SQL parsing research in the community, we release a large-scale and multi-accent dataset MASpider; 2) leveraging the recent progress in the large-scale pre-training, we show that it alleviates the data scarcity issue and allow for direct speech-to-SQL parsing; and 3) we include the speech re-programming and gradient reversal classifier techniques to reduce acoustic variance and learned style-agnostic representation, improving generalization to unseen out-of-domain custom data. Experimental results demonstrate that Wav2SQL avoids error compounding and achieves state-of-the-art results by up to 4.7% accuracy improvement over the baseline.

Thousands of the world’s languages are in danger of extinction—a tremendous threat to cultural identities and human language diversity. Interlinear Glossed Text (IGT) is a form of linguistic annotation that can support documentation and resource creation for these languages’ communities. IGT typically consists of (1) transcriptions, (2) morphological segmentation, (3) glosses, and (4) free translations to a majority language. We propose Wav2Gloss: a task in which these four annotation components are extracted automatically from speech, and introduce the first dataset to this end, Fieldwork: a corpus of speech with all these annotations, derived from the work of field linguists, covering 37 languages, with standard formatting, and train/dev/test splits. We provide various baselines to lay the groundwork for future research on IGT generation from speech, such as end-to-end versus cascaded, monolingual versus multilingual, and single-task versus multi-task approaches.

To mitigate the potential misuse of large language models (LLMs), recent research has developed watermarking algorithms, which restrict the generation process to leave an invisible trace for watermark detection. Due to the two-stage nature of the task, most studies evaluate the generation and detection separately, thereby presenting a challenge in unbiased, thorough, and applicable evaluations. In this paper, we introduce WaterBench, the first comprehensive benchmark for LLM watermarks, in which we design three crucial factors: (1) For benchmarking procedure, to ensure an apples-to-apples comparison, we first adjust each watermarking method’s hyper-parameter to reach the same watermarking strength, then jointly evaluate their generation and detection performance. (2) For task selection, we diversify the input and output length to form a five-category taxonomy, covering 9 tasks. (3) For evaluation metric, we adopt the GPT4-Judge for automatically evaluating the decline of instruction-following abilities after watermarking. We evaluate 4 open-source watermarks on 2 LLMs under 2 watermarking strengths and observe the common struggles for current methods on maintaining the generation quality. The code and data are available at https://github.com/THU-KEG/WaterBench.

Text watermarking has emerged as a pivotal technique for identifying machine-generated text. However, existing methods often rely on arbitrary vocabulary partitioning during decoding to embed watermarks, which compromises the availability of suitable tokens and significantly degrades the quality of responses. This study assesses the impact of watermarking on different capabilities of large language models (LLMs) from a cognitive science lens. Our finding highlights a significant disparity; knowledge recall and logical reasoning are more adversely affected than language generation. These results suggest a more profound effect of watermarking on LLMs than previously understood. To address these challenges, we introduce Watermarking with Mutual Exclusion (WatME), a novel approach leveraging linguistic prior knowledge of inherent lexical redundancy in LLM vocabularies to seamlessly integrate watermarks. Specifically, WatME dynamically optimizes token usage during the decoding process by applying a mutually exclusive rule to the identified lexical redundancies. This strategy effectively prevents the unavailability of appropriate tokens and preserves the expressive power of LLMs. We provide both theoretical analysis and empirical evidence showing that WatME effectively preserves the diverse capabilities of LLMs while ensuring watermark detectability.

As the scale of Large Language Models (LLMs) increases, it is necessary to compress the models to reduce the substantial demand on computational resources. Network pruning significantly reduces the model size by converting the weight matrix from dense to sparse data format. Current methodologies advocate for one-shot pruning to avoid the expense of retraining, ensuring the maintenance of model performance under conditions of 50%-60% unstructured pruning. Nevertheless, matrices characterized by this level of sparsity could not be treated as sparse matrices, because the indices would incur significant costs. To mitigate this problem, NVIDIA introduced the 2:4 structured sparsity. However, we observe a notable decline in model performance when adopting 2:4 structured sparsity due to group constraints. In this paper, we introduce the Weight Recover Prune (WRP) approach. By recovering a minimal set of critical weights, WRP aims to enhance model performance while maintaining the efficiency of the compression. Our evaluation of the WRP method on the LLAMA2 and OPT models shows that it outperforms other 2:4 pattern one-shot pruning methods. Meanwhile, WRP can guarantee that the size of the pruned model is about 60% of the dense model. Our code is available at: https://github.com/TanZhendong/WRP.

Embedding as a Service (EaaS) has become a widely adopted solution, which offers feature extraction capabilities for addressing various downstream tasks in Natural Language Processing (NLP). Prior studies have shown that EaaS can be prone to model extraction attacks; nevertheless, this concern could be mitigated by adding backdoor watermarks to the text embeddings and subsequently verifying the attack models post-publication. Through the analysis of the recent watermarking strategy for EaaS, EmbMarker, we design a novel CSE (Clustering, Selection, Elimination) attack that removes the backdoor watermark while maintaining the high utility of embeddings, indicating that the previous watermarking approach can be breached. In response to this new threat, we propose a new protocol to make the removal of watermarks more challenging by incorporating multiple possible watermark directions. Our defense approach, WARDEN, notably increases the stealthiness of watermarks and has been empirically shown to be effective against CSE attack.

Security practitioners maintain vulnerability reports (e.g., GitHub Advisory) to help developers mitigate security risks. An important task for these databases is automatically extracting structured information mentioned in the report, e.g., the affected software packages, to accelerate the defense of the vulnerability ecosystem.However, it is challenging for existing work on affected package identification to achieve high precision. One reason is that all existing work focuses on relatively smaller models, thus they cannot harness the knowledge and semantic capabilities of large language models.To address this limitation, we propose VulLibGen, the first method to use LLM for affected package identification. In contrast to existing work, VulLibGen proposes the novel idea to directly generate the affected package. To improve the precision, VulLibGen employs supervised fine-tuning (SFT), retrieval augmented generation (RAG) and a local search algorithm. The local search algorithm is a novel post-processing algorithm we introduce for reducing the hallucination of the generated packages. Our evaluation results show that VulLibGen has an average precision of 0.806 for identifying vulnerable packages in the four most popular ecosystems in GitHub Advisory (Java, JS, Python, Go) while the best average precision in previous work is 0.721. Additionally, VulLibGen has high value to security practice: we submitted 60 <vulnerability, affected package> pairs to GitHub Advisory (covers four ecosystems) and 34 of them have been accepted and merged.

We introduce VoiceCraft, a token infilling neural codec language model, that achieves state-of-the-art performance on both speech editing and zero-shot text-to-speech (TTS) on audiobooks, internet videos, and podcasts. VoiceCraft employs a Transformer decoder architecture and introduces a token rearrangement procedure that combines causal masking and delayed stacking to enable generation within an existing sequence. On speech editing tasks, VoiceCraft produces edited speech that is nearly indistinguishable from unedited recordings in terms of naturalness, as evaluated by humans; for zero-shot TTS, our model outperforms prior SotA models including VALL-E and the popular commercial model XTTS v2. Crucially, the models are evaluated on challenging and realistic datasets, that consist of diverse accents, speaking styles, recording conditions, and background noise and music, and our model performs consistently well compared to other models and real recordings. In particular, for speech editing evaluation, we introduce a high quality, challenging, and realistic dataset named . We encourage readers to listen to the demos at https://jasonppy.github.io/VoiceCraft_web. Data, code, and model weights are available at https://github.com/jasonppy/VoiceCraft

For dialogue systems, the utilization of multimodal dialogue responses, as opposed to relying solely on text-only responses, offers the capability to describe different concepts through various modalities. This enhances the effectiveness of communication and elevates the overall conversational experience. However, current methods for dialogue-to-image retrieval are constrained by the capabilities of the pre-trained vision language models (VLMs). They struggle to accurately extract key information from conversations and are unable to handle long-turn conversations. In this paper, we leverage the reasoning capabilities of large language models (LLMs) to predict the potential features that may be present in the images to be shared, based on the dialogue context. This approach allows us to obtain succinct and precise descriptors, thereby improving the performance of text-image retrieval. Experimental results shows that our method outperforms previous approaches significantly in terms of Recall@k.

Visualization recommendations, which aim to automatically match proper visual charts for specific data tables, can significantly simplify the data analysis process. Traditional approaches in this domain have primarily relied on rule-based or machine learning-based methodologies. These methods often demand extensive manual maintenance and yet fail to fully comprehend the tabular data, leading to unsatisfactory performance. Recently, Large Language Models (LLMs) have emerged as powerful tools, exhibiting strong reasoning capabilities. This advancement suggests their substantial promise in addressing visualization recommendation challenges. However, effectively harnessing LLMs to discern and rationalize patterns in tabular data, and consequently deduce the essential information for chart generation, remains an unresolved challenge. To this end, we introduce a novel Hierarchical Table Prompt-based reprogramming framework, named HTP. This framework aims to integrate multi-dimensional tabular data into LLMs through a strategically crafted prompt learning method while keeping the LLMs’ backbone and weights unaltered. The HTP framework uniquely incorporates a four-level prompt structure, encompassing general, instance, cluster, and column levels. This multi-level approach is engineered to provide a comprehensive understanding of both general distribution and multifaceted fine-grained features of tabular data, before inputting the tabular data into the frozen LLM. Our empirical studies confirm that the HTP framework achieves state-of-the-art performance, marking an advancement in the field of data visualization and analysis. The code and data will be made publicly available upon acceptance.

Autonomous agents capable of planning, reasoning, and executing actions on the web offer a promising avenue for automating computer tasks. However, the majority of existing benchmarks primarily focus on text-based agents, neglecting many natural tasks that require visual information to effectively solve. Given that most computer interfaces cater to human perception, visual information often augments textual data in ways that text-only models struggle to harness effectively. To bridge this gap, we introduce VisualWebArena, a benchmark designed to assess the performance of multimodal web agents on *realistic visually grounded tasks*. VisualWebArena comprises of a set of diverse and complex web-based tasks that evaluate various capabilities of autonomous multimodal agents. To perform on this benchmark, agents need to accurately process image-text inputs, interpret natural language instructions, and execute actions on websites to accomplish user-defined objectives. We conduct an extensive evaluation of state-of-the-art LLM-based autonomous agents, including several multimodal models. Through extensive quantitative and qualitative analysis, we identify several limitations of text-only LLM agents, and reveal gaps in the capabilities of state-of-the-art multimodal language agents. VisualWebArena provides a framework for evaluating multimodal autonomous language agents, and offers insights towards building stronger autonomous agents for the web.

In Large Visual Language Models (LVLMs), the efficacy of In-Context Learning (ICL) remains limited by challenges in cross-modal interactions and representation disparities. To overcome these challenges, we introduce a novel Visual In-Context Learning (VICL) method comprising Visual Demonstration Retrieval, Intent-Oriented Image Summarization, and Intent-Oriented Demonstration Composition. Our approach retrieves images via ”Retrieval & Rerank” paradigm, summarises images with task intent and task-specific visual parsing, and composes language-based demonstrations that reduce token count and alleviate cross-modal interaction problem. Experimental evaluations on five visual reasoning datasets demonstrate the effectiveness of our method. Moreover, our extensive experiments leverage information flow analysis to elucidate the effectiveness of our method, and investigate the impact of length and position of demonstrations for LVLM. The use of in-context unlearning further shows promise in resetting specific model knowledge without retraining.

Visual hallucination (VH) means that a multi-modal LLM (MLLM) imagines incorrect details about an image in visual question answering. Existing studies find VH instances only in existing image datasets, which results in biased understanding of MLLMs’ performance under VH due to limited diversity of such VH instances. In this work, we propose a tool called VHTest to generate a diverse set of VH instances. Specifically, VHTest finds some initial VH instances in existing image datasets (e.g., COCO), generates a text description for each VH mode, and uses a text-to-image generative model (e.g., DALL-E-3) to generate VH images based on the text descriptions. We collect a benchmark dataset with 1,200 VH instances in 8 VH modes using VHTest. We find that existing MLLMs such as GPT-4, LLaVA-1.5, and MiniGPT-v2 hallucinate for a large fraction of the instances in our benchmark. Moreover, we find that fine-tuning an MLLM using our benchmark dataset reduces its likelihood to hallucinate without sacrificing its performance on other benchmarks. Our benchmarks are publicly available: https://github.com/wenhuang2000/VHTest.

Despite vision-language models’ (VLMs) remarkable capabilities as versatile visual assistants, two substantial challenges persist within the existing VLM frameworks: (1) lacking task diversity in pretraining and visual instruction tuning, and (2) annotation error and bias in GPT-4 synthesized instruction tuning data. Both challenges lead to issues such as poor generalizability, hallucination, and catastrophic forgetting. To address these challenges, we construct Vision-Flan, the most diverse publicly available visual instruction tuning dataset to date, comprising 187 diverse tasks and 1,664,261 instances sourced from academic datasets, and each task is accompanied by an expert-written instruction. In addition, we propose a two-stage instruction tuning framework, in which VLMs are firstly finetuned on Vision-Flan and further tuned on GPT-4 synthesized data. We find this two-stage tuning framework significantly outperforms the traditional single-stage visual instruction tuning framework and achieves the state-of-the-art performance across a wide range of multi-modal evaluation benchmarks. Finally, we conduct in-depth analyses to understand visual instruction tuning and our findings reveal that: (1) GPT-4 synthesized data does not substantially enhance VLMs’ capabilities but rather modulates the model’s responses to human-preferred formats; (2) A minimal quantity (e.g., 1,000) of GPT-4 synthesized data can effectively align VLM responses with human-preference; (3) Visual instruction tuning mainly helps large-language models (LLMs) to understand visual features.

Despite the significant success of large vision-language models (LVLMs), some studies have revealed that LVLMs suffer from the hallucination problem, where the LVLMs’ response contains descriptions of non-existent objects. Although various benchmarks have been proposed to investigate this problem, they mostly focus on single-turn evaluation and overlook the hallucination raised by textual inputs. To investigate the hallucination problem of LVLMs when given long-term misleading textual history, we propose a novel visual dialogue hallucination evaluation benchmark VisDiaHalBench. The benchmark consists of samples with five-turn questions about an edited image and its original version. VisDiaHalBench differs from previous hallucination benchmarks in the following three points: 1) The questions and answers are unambiguously grounded by annotated scene graphs. 2) The images are uncommonly edited to inspect the visual model and common-object hallucination in LLMs. 3) The carefully designed dialogue refers a same object in different turns to assess the image consistency and influence of history for LVLMs. The detailed analysis of several state-of-the-art LVLMs across image consistency, visual understanding, history influence, and other dimensions reveals their substantial performance gap with single-turn VQA tasks. The benchmark is released in: https://github.com/qingxingcao/VisDiaHalBench