ACL 2024 论文列表

With the advancement of large language models (LLMs) and the expansion of their context windows, existing long-context benchmarks fall short in effectively evaluating the models’ comprehension and reasoning abilities in extended texts. Moreover, conventional benchmarks relying on F1 metrics often inaccurately score responses: they may undervalue correct answers that differ from the reference responses and overvalue incorrect ones that resemble the reference texts. In response to these limitations, we introduce Marathon, a novel evaluation benchmark that adopts a multiple-choice question format. It is specifically designed to overcome the constraints of previous benchmarks and provide a rapid, precise, and unbiased appraisal of the long-context comprehension skills of large language models. We conducted comprehensive evaluations on the Marathon benchmark with a range of state-of-the-art LLMs and assessed the effectiveness of various optimization strategies tailored for long-context generation. We anticipate that the Marathon benchmark and its associated leaderboard will enable a more precise and equitable evaluation of LLMs’ capabilities in understanding and reasoning over extended contexts.

Embodied agents equipped with GPT as their brain have exhibited extraordinary decision-making and generalization abilities across various tasks. However, existing zero-shot agents for vision-and-language navigation (VLN) only prompt the GPT-4 to select potential locations within localized environments, without constructing an effective “global-view” for the agent to understand the overall environment. In this work, we present a novel **map**-guided **GPT**-based agent, dubbed **MapGPT**, which introduces an online linguistic-formed map to encourage the global exploration. Specifically, we build an online map and incorporate it into the prompts that include node information and topological relationships, to help GPT understand the spatial environment. Benefiting from this design, we further propose an adaptive planning mechanism to assist the agent in performing multi-step path planning based on a map, systematically exploring multiple candidate nodes or sub-goals step by step. Extensive experiments demonstrate that our MapGPT is applicable to both GPT-4 and GPT-4V, achieving state-of-the-art zero-shot performance on the R2R and REVERIE simultaneously (~10% and ~12% improvements in SR), and showcasing the newly emergent global thinking and path planning abilities of the GPT.

Code synthesis, which requires a deep understanding of complex natural language (NL) problem descriptions, generation of code instructions for complex algorithms and data structures, and the successful execution of comprehensive unit tests, presents a significant challenge. Thus, while large language models (LLMs) demonstrate impressive proficiency in natural language processing (NLP), their performance in code generation tasks remains limited. In this paper, we introduce a new approach to code generation tasks leveraging the multi-agent prompting that uniquely replicates the full cycle of program synthesis as observed in human developers. Our framework, MapCoder, consists of four LLM agents specifically designed to emulate the stages of this cycle: recalling relevant examples, planning, code generation, and debugging. After conducting thorough experiments, with multiple LLMs ablations and analyses across eight challenging competitive problem-solving and program synthesis benchmarks—MapCoder showcases remarkable code generation capabilities, achieving their new state-of-the-art (pass@1) results—(HumanEval 93.9%, MBPP 83.1%, APPS 22.0%, CodeContests 28.5%, and xCodeEval 45.3%). Moreover, our method consistently delivers superior performance across various programming languages and varying problem difficulties. We open-source our framework at https://github.com/Md-Ashraful-Pramanik/MapCoder.

Recent advancements in long-context modeling have enhanced language models (LMs) for complex tasks across multiple NLP applications. Despite this progress, we find that these models struggle with multi-hop reasoning and exhibit decreased performance in the presence of noisy contexts. In this paper, we introduce Reasoning with Attributions, a novel approach that prompts LMs to supply attributions for each assertion during their reasoning. We validate our approach through experiments on three multi-hop datasets, employing both proprietary and open-source models, and demonstrate its efficacy and resilience. Furthermore, we explore methods to augment reasoning capabilities via fine-tuning and offer an attribution-annotated dataset and a specialized training strategy. Our fine-tuned model achieves competitive performance on multi-hop reasoning benchmarks, closely paralleling proprietary LMs such as ChatGPT and Claude-instant.

Large language models (LLMs) have shown great potential to empower various domains and are often customized by fine-tuning for the requirements of different applications. However, the powerful learning ability of LLMs not only enables them to learn new tasks but also makes them vulnerable to learning undesired behaviors, such as harmfulness and hallucination, as the fine-tuning data often implicitly or explicitly contains such content. Can we fine-tune LLMs on harmful data without learning harmful behaviors? This paper proposes a controllable training framework to make undesired behaviors unlearnable during the fine-tuning process. Specifically, we introduce security vectors to control the model’s behavior and make it consistent with the undesired behavior. Security vectors are activated during fine-tuning, the consistent behavior makes the model believe that such behavior has already been learned and there is no need for further optimization, while inconsistent data can still be learned. After fine-tuning, security vectors are deactivated to restore the LLM’s normal behavior. Our experiments show that the security vectors can prevent LLM from learning harmful and hallucination behavior while preserving the ability to learn other information.

Large language models (LLMs) have successfully served as a general-purpose interface across multiple tasks and languages, while the adaptation of voice LLMs is mostly designed for specific purposes (either single-task or monolingual), where the advantages of LLMs especially for low-resource language processing and zero-shot task generalization are less exploited in the audio community. To bridge the gap, we introduce Make-A-Voice as a multi-modal voice LLM and conduct a comprehensive study on its capability to deal with multiple tasks/languages. When trained on ~200K hours of 6-language data for 4 voice generation applications, Make-A-Voice emerges notable advantages: 1) as scalable learners to improve performance with end-to-end local and global multiscale transformers; and 2) as multitask learners by adjusting prompts to share common knowledge across modalities (speech/singing) and present in-context learning abilities by generalizing to unseen tasks not explicitly train on; 3) as multilingual learners to alleviate data scarcity of low-resource languages by including rich-resource language training data. Experimental results demonstrate that Make-A-Voice exhibits superior audio quality and style similarity compared with competitive baseline models in monolingual/cross-lingual voice generation. Audio samples are available at https://M-Voice.github.io

Machine-Generated Text (MGT) detection aims to identify a piece of text as machine or human written. Prior work has primarily formulated MGT detection as a binary classification task over an entire document, with limited work exploring cases where only part of a document is machine generated. This paper provides the first in-depth study of MGT that localizes the portions of a document that were machine generated. Thus, if a bad actor were to change a key portion of a news article to spread misinformation, whole document MGT detection may fail since the vast majority is human written, but our approach can succeed due to its granular approach. A key challenge in our MGT localization task is that short spans of text, *e.g.*, a single sentence, provides little information indicating if it is machine generated due to its short length. To address this, we leverage contextual information, where we predict whether multiple sentences are machine or human written at once. This enables our approach to identify changes in style or content to boost performance. A gain of 4-13% mean Average Precision (mAP) over prior work demonstrates the effectiveness of approach on five diverse datasets: GoodNews, VisualNews, WikiText, Essay, and WP. We release our implementation at https://github.com/Zhongping-Zhang/MGT_Localization.

This study investigates the concept of the ‘right to be forgotten’ within the context of large language models (LLMs). We explore machine unlearning as a pivotal solution, with a focus on pre-trained models–a notably under-researched area. Our research delineates a comprehensive framework for machine unlearning in pre-trained LLMs, encompassing a critical analysis of seven diverse unlearning methods. Through rigorous evaluation using curated datasets from arXiv, books, and GitHub, we establish a robust benchmark for unlearning performance, demonstrating that these methods are over 105 times more computationally efficient than retraining. Our results show that integrating gradient ascent with gradient descent on in-distribution data improves hyperparameter robustness. We also provide detailed guidelines for efficient hyperparameter tuning in the unlearning process. Our findings advance the discourse on ethical AI practices, offering substantive insights into the mechanics of machine unlearning for pre-trained LLMs and underscoring the potential for responsible AI development.

A major consideration in multilingual language modeling is how to best represent languages with diverse vocabularies and scripts.Although contemporary text encoding methods cover most of the world’s writing systems, they exhibit bias towards the high-resource languages of the Global West. As a result, texts of underrepresented languages tend to be segmented into long sequences of linguistically meaningless units. To address the disparities, we introduce a new paradigm that encodes the same information with segments of consistent size across diverse languages. Our encoding convention (MYTE) is based on morphemes, as their inventories are more balanced across languages than characters, which are used in previous methods. We show that MYTE produces shorter encodings for all 99 analyzed languages, with the most notable improvements for non-European languages and non-Latin scripts. This, in turn, improves multilingual LM performance and diminishes the perplexity gap throughout diverse languages.

Adjusting the outdated behaviors of large langugae models (LLMs) after deployment remains a significant challenge. It motivates the model editing research, which is however mainly explored in a restricted task form with triple-based edit requests. Recent works have initiated a transition to a more practical and unified editing task that takes free-form text as edit requests. However, there are gaps in nuanced benchmark designs and re-evaluation of existing methods. To bridge the gaps, we introduce a multi-level benchmark for free text model editing (MULFE). The benchmark categorizes probe queries into three levels of generalization, ranging from basic literal memory to deeper understanding and reasoning. Based on the benchmark, we conduct extensive experiments across various base models, edit sizes, and editing methods, including adaptations of mainstream locate-and-edit and hypernetwork methods. The results highlight the inconsistent behaviors of edited models on different generalization levels. Higher-level generalization remains a significant challenge. Based on the findings, we propose SIDE, a simple yet effective method based on in-context distillation to enhance the generalization performance. The benchmark dataset and evaluation scripts are publicly available at http://github.com/wchrepo/mulfe.

Detecting critical moments, such as emotional outbursts or changes in decisions during conversations, is crucial for understanding shifts in human behavior and their consequences. Our work introduces a novel problem setting focusing on these moments as turning points (TPs), accompanied by a meticulously curated, high-consensus, human-annotated multi-modal dataset. We provide precise timestamps, descriptions, and visual-textual evidence high-lighting changes in emotions, behaviors, perspectives, and decisions at these turning points. We also propose a framework, TPMaven, utilizing state-of-the-art vision-language models to construct a narrative from the videos and large language models to classify and detect turning points in our multi-modal dataset. Evaluation results show that TPMaven achieves an F1-score of 0.88 in classification and 0.61 in detection, with additional explanations aligning with human expectations.

The advent of Large Language Models (LLMs) has drastically enhanced dialogue systems. However, comprehensively evaluating the dialogue abilities of LLMs remains a challenge. Previous benchmarks have primarily focused on single-turn dialogues or provided coarse-grained and incomplete assessments of multi-turn dialogues, overlooking the complexity and fine-grained nuances of real-life dialogues. To address this issue, we introduce MT-Bench-101, specifically designed to evaluate the fine-grained abilities of LLMs in multi-turn dialogues. By conducting a detailed analysis of real multi-turn dialogue data, we construct a three-tier hierarchical ability taxonomy comprising 4208 turns across 1388 multi-turn dialogues in 13 distinct tasks. We then evaluate 21 popular LLMs based on MT-Bench-101, conducting comprehensive analyses from both ability and task perspectives and observing differing trends in LLMs performance across dialogue turns within various tasks. Further analysis indicates that neither utilizing common alignment techniques nor chat-specific designs has led to obvious enhancements in the multi-turn abilities of LLMs. Extensive case studies suggest that our designed tasks accurately assess the corresponding multi-turn abilities. The data and code are available at https://github.com/mtbench101/mt-bench-101.

Sign language understanding has made significant strides; however, there is still no viable solution for generating sign sequences directlyfrom entire spoken content, e.g., text or speech. In this paper, we propose a unified framework for continuous sign language production, easing communication between sign and non-sign language users. In particular, a sequence diffusion model, utilizing embeddings extracted from text or speech, is crafted to generate sign predictions step by step. Moreover, by creating a joint embedding space for text, audio, and sign, we bind these modalities and leverage the semantic consistency among them to provide informative feedback for the model training. This embedding-consistency learning strategy minimizes the reliance on sign triplets and ensures continuous model refinement, evenwith a missing audio modality. Experiments on How2Sign and PHOENIX14T datasets demonstrate that our model achieves competitive performance in sign language production.

Syntactic parsing remains a critical tool for relation extraction and information extraction, especially in resource-scarce languages where LLMs are lacking. Yet in morphologically rich languages (MRLs), where parsers need to identify multiple lexical units in each token, existing systems suffer in latency and setup complexity. Some use a pipeline to peel away the layers: first segmentation, then morphology tagging, and then syntax parsing; however, errors in earlier layers are then propagated forward. Others use a joint architecture to evaluate all permutations at once; while this improves accuracy, it is notoriously slow. In contrast, and taking Hebrew as a test case, we present a new “flipped pipeline”: decisions are made directly on the whole-token units by expert classifiers, each one dedicated to one specific task. The classifier predictions are independent of one another, and only at the end do we synthesize their predictions. This blazingly fast approach requires only a single huggingface call, without the need for recourse to lexicons or linguistic resources. When trained on the same training set used in previous studies, our model achieves near-SOTA performance on a wide array of Hebrew NLP tasks. Furthermore, when trained on a newly enlarged training corpus, our model achieves a new SOTA for Hebrew POS tagging and dependency parsing. We release this new SOTA model to the community. Because our architecture does not rely on any language-specific resources, it can serve as a model to develop similar parsers for other MRLs.

Large Language Models (LLMs) have demonstrated great potential for assisting developers in their daily development. However, most research focuses on generating correct code, how to use LLMs to generate personalized code has seldom been investigated. To bridge this gap, we proposed MPCoder (Multi-user Personalized Code Generator) to generate personalized code for multiple users. To better learn coding style features, we utilize explicit coding style residual learning to capture the syntax code style standards and implicit style learning to capture the semantic code style conventions. We train a multi-user style adapter to better differentiate the implicit feature representations of different users through contrastive learning, ultimately enabling personalized code generation for multiple users. We further propose a novel evaluation metric for estimating similarities between codes of different coding styles. The experimental results show the effectiveness of our approach for this novel task.

Dialogue discourse parsing (DDP) aims to capture the relations between utterances in the dialogue. In everyday real-world scenarios, dialogues are typically multi-modal and cover open-domain topics. However, most existing widely used benchmark datasets for DDP contain only textual modality and are domain-specific. This makes it challenging to accurately and comprehensively understand the dialogue without multi-modal clues, and prevents them from capturing the discourse structures of the more prevalent daily conversations. This paper proposes MODDP, the first multi-modal Chinese discourse parsing dataset derived from open-domain daily dialogues, consisting 864 dialogues and 18,114 utterances, accompanied by 12.7 hours of video clips. We present a simple yet effective benchmark approach for multi-modal DDP. Through extensive experiments, we present several benchmark results based on MODDP. The significant improvement in performance from introducing multi-modalities into the original textual unimodal DDP model demonstrates the necessity of integrating multi-modalities into DDP.

The rapid proliferation of online content necessitates effective summarization methods, among which dynamic aspect-based summarization stands out. Unlike its traditional counterpart, which assumes a fixed set of known aspects, this approach adapts to the varied aspects of the input text. We introduce a novel multi-objective learning framework employing a Longformer-Encoder-Decoder for this task. The framework optimizes aspect number prediction, minimizes disparity between generated and reference summaries for each aspect, and maximizes dissimilarity across aspect-specific summaries. Extensive experiments show our method significantly outperforms baselines on three diverse datasets, largely due to the effective alignment of generated and reference aspect counts without sacrificing single-aspect summarization quality.

Theory of Mind (ToM), the ability to understand people’s mental states, is an essential ingredient for developing machines with human-level social intelligence. Recent machine learning models, particularly large language models, seem to show some aspects of ToM understanding. However, existing ToM benchmarks use unimodal datasets – either video or text. Human ToM, on the other hand, is more than video or text understanding. People can flexibly reason about another person’s mind based on conceptual representations (e.g., goals, beliefs, plans) extracted from any available data. To address this, we introduce a multimodal Theory of Mind question answering (MMToM-QA) benchmark. MMToM-QA comprehensively evaluates machine ToM both on multimodal data and on different kinds of unimodal data about a person’s activity in a household environment. To engineer multimodal ToM capacity, we propose a novel method, BIP-ALM (Bayesian Inverse Planning Accelerated by Language Models). BIP-ALM extracts unified representations from multimodal data and utilizes language models for scalable Bayesian inverse planning. We conducted a systematic comparison of human performance, BIP-ALM, and state-of-the-art models, including GPT-4. The experiments demonstrate that large language models and large multimodal models still lack robust ToM capacity. BIP-ALM, on the other hand, shows promising results, by leveraging the power of both model-based mental inference and language models.

Social media platforms are hubs for multimodal information exchange, encompassing text, images, and videos, making it challenging for machines to comprehend the information or emotions associated with interactions in online spaces. Multimodal Large Language Models (MLLMs) have emerged as a promising solution to address these challenges, yet struggle with accurately interpreting human emotions and complex contents like misinformation. This paper introduces MM-Soc, a comprehensive benchmark designed to evaluate MLLMs’ understanding of multimodal social media content. MM-Soc compiles prominent multimodal datasets and incorporates a novel large-scale YouTube tagging dataset, targeting a range of tasks from misinformation detection, hate speech detection, and social context generation. Through our exhaustive evaluation on ten size-variants of four open-source MLLMs, we have identified significant performance disparities, highlighting the need for advancements in models’ social understanding capabilities. Our analysis reveals that, in a zero-shot setting, various types of MLLMs generally exhibit difficulties in handling social media tasks. However, MLLMs demonstrate performance improvements post fine-tuning, suggesting potential pathways for improvement.