ACL 2024 Papers — Page 6
Annual Meeting of the Association for Computational Linguistics · 940 papers
M^3AV: A Multimodal, Multigenre, and Multipurpose Audio-Visual Academic Lecture Dataset
Zhe Chen (Shanghai JiaoTong University), Yanfeng Wang (Shanghai JiaoTong University)
TransformerLarge Language ModelVideoTextMultimodalityBenchmarkRetrieval-Augmented GenerationAudio
🎯 What it does: Constructed a multimodal, multi-genre, cross-purpose academic lecture video dataset M3AV, providing high-quality manually annotated speech transcripts, slide OCR, and paper text.
M^3CoT: A Novel Benchmark for Multi-Domain Multi-step Multi-modal Chain-of-Thought
Qiguang Chen (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
ClassificationImage TranslationLarge Language ModelVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Developed the multi-domain, multi-step, and multi-modal chained reasoning benchmark M CoT 3, and conducted systematic evaluations of various VLLMs;
M4GT-Bench: Evaluation Benchmark for Black-Box Machine-Generated Text Detection
Yuxia Wang (Mohamed bin Zayed University of Artificial Intelligence), Preslav Nakov (Mohamed bin Zayed University of Artificial Intelligence)
ClassificationAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed and implemented M4GT-Bench—a multilingual, multi-domain, multi-generator black-box machine-generated text detection benchmark covering three tasks: binary classification, human-machine text hybrid detection, and multi-generator attribution.
M4LE: A Multi-Ability Multi-Range Multi-Task Multi-Domain Long-Context Evaluation Benchmark for Large Language Models
Wai-Chung Kwan (Chinese University of Hong Kong), Kam-Fai Wong (Huawei Noah's Ark Lab)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed a multidimensional (multi-capability, multi-scope, multi-task, multi-domain) long-text evaluation benchmark, M4LE, which automatically converts short-text tasks into long-text scenarios.
Machine Unlearning of Pre-trained Large Language Models
Jin Yao (University of Virginia), Xiang Yue (Carnegie Mellon University)
Safty and PrivacyComputational EfficiencyAdversarial AttackData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a unified framework for machine unlearning and implemented and evaluated seven unlearning methods on pre-trained large language models (LLMs).
MAGE: Machine-generated Text Detection in the Wild
Yafu Li (Zhejiang University), Yue Zhang (Westlake University)
Anomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Constructed a large-scale multi-domain and multi-model benchmark for detecting machine-generated text (MAGE), and systematically evaluated the performance of various detection methods in real-world scenarios on this benchmark.
Make-A-Voice: Revisiting Voice Large Language Models as Scalable Multilingual and Multitask Learners
Rongjie Huang (Zhejiang University), Dong Yu (Tencent AI Lab)
GenerationTransformerLarge Language ModelPrompt EngineeringTextMultimodalityAudio
🎯 What it does: Built a multi-modal, scalable voice large language model called Make-A-Voice, supporting multi-task (TTS, VC, SVS, SVC) and multilingual zero-shot generation.
Making Long-Context Language Models Better Multi-Hop Reasoners
Yanyang Li (Chinese University of Hong Kong), Liwei Wang (Chinese University of Hong Kong)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the 'Reasoning with Attributions' method, enabling long-context LMs to provide the source of each step during multi-hop reasoning through chain citation (CoC) or citation (CoQ), thereby improving the handling of noisy contexts and reasoning accuracy.
MAP’s not dead yet: Uncovering true language model modes by conditioning away degeneracy
Davis Yoshida (Toyota Technological Institute at Chicago), Kevin Gimpel (Toyota Technological Institute at Chicago)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Investigated the problem of degenerate outputs in natural language generation models under MAP (maximum a posteriori) decoding, proving that even when models perfectly fit the training distribution, low-entropy noise can still lead to modal degradation, and proposed attribute conditional search (e.g., length conditioning) to identify high-quality conditional modes, further validating its effectiveness through exact search and approximate conditional search.
MapCoder: Multi-Agent Code Generation for Competitive Problem Solving
Md. Ashraful Islam (Bangladesh University of Engineering and Technology), Md Rizwan Parvez (Bangladesh University of Engineering and Technology)
GenerationAI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented the MapCoder framework, which simulates human programming loops by utilizing four LLM agents (retrieval, planning, coding, debugging) to automatically generate executable code for competitive programming problems.
MapGPT: Map-Guided Prompting with Adaptive Path Planning for Vision-and-Language Navigation
Jiaqi Chen (University of Hong Kong), Kwan-Yee Wong
TransformerLarge Language ModelPrompt EngineeringVision Language ModelVision-Language-Action ModelImageTextMultimodality
🎯 What it does: Proposes MapGPT, a zero-shot vision-language navigation (VLN) agent that achieves global exploration using online-generated textualized topological maps and adaptive multi-step path planning.
MAPO: Advancing Multilingual Reasoning through Multilingual-Alignment-as-Preference Optimization
Shuaijie She (Nanjing University), Jiajun Chen (Nanjing University)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: This paper proposes the MAPO framework, which aligns the reasoning process of non-main languages to the main language using a multilingual translation model, thereby enhancing the model's reasoning capabilities on low-resource languages.
Marathon: A Race Through the Realm of Long Context with Large Language Models
Lei Zhang (Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences), Min Yang (Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences)
TransformerPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the Marathon benchmark, which evaluates the understanding and reasoning capabilities of large language models in 2K–260K long contexts using multiple-choice questions, and conducts experiments on various long-context optimization methods.
MARS: Meaning-Aware Response Scoring for Uncertainty Estimation in Generative LLMs
Yavuz Faruk Bakman (University of Southern California), Salman Avestimehr (University of Southern California)
GenerationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelTextBiomedical Data
🎯 What it does: Proposed a new score function called MARS to replace traditional length-normalized scoring, better capturing sentence semantic importance in uncertainty estimation for generative LLMs.
MARVEL: Unlocking the Multi-Modal Capability of Dense Retrieval via Visual Module Plugin
Tianshuo Zhou (Northeastern University), Ge Yu (Northeastern University)
RetrievalTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the MARVEL model, integrating the CLIP visual module as a plugin into a pre-trained T5-ANCE dense retriever, and achieve multimodal retrieval by fine-tuning only the language model after contrastive pre-training of the visual module on image-caption pairs.
Masked Thought: Simply Masking Partial Reasoning Steps Can Improve Mathematical Reasoning Learning of Language Models
Changyu Chen (Renmin University of China), Yongbin Li (Alibaba Group)
TransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Proposed Masked Thought Fine-Tuning (MFT), enhancing large language models' mathematical reasoning ability by randomly masking partial steps in chain-of-thought reasoning.
MaskLID: Code-Switching Language Identification through Iterative Masking
Amir Hossein Kargaran (LMU Munich & Munich Center for Machine Learning), Hinrich Schuetze
ClassificationText
🎯 What it does: MaskLID detects multilingual (code-switching) components in sentences by iteratively masking the features of the dominant language in existing sentence-level language identification (LID) models.
Math-Shepherd: Verify and Reinforce LLMs Step-by-step without Human Annotations
Peiyi Wang (Peking University), Zhifang Sui (Peking University)
OptimizationReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: Propose a step-based mathematical reasoning reward model called MATH-SHEPHERD, which scores each step of the mathematical solutions generated by LLMs and is used for verification and reinforcement learning.
MathGenie: Generating Synthetic Data with Question Back-translation for Enhancing Mathematical Reasoning of LLMs
Zimu Lu (Chinese University of Hong Kong), Hongsheng Li (CPII under InnoHK)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the MathGenie framework, which first generates diverse and reliable math problems through iterative solution augmentation and problem reverse translation. It then uses a verification-based filtering scheme to generate credible code-integrated solutions. Subsequently, multiple open-source LLMs are fine-tuned on this synthetic data to obtain the MathGenieLM series of models.
MAVEN-ARG: Completing the Puzzle of All-in-One Event Understanding Dataset with Event Argument Annotation
Xiaozhi Wang (Tsinghua University), Juanzi Li (Tencent Inc)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Proposed and constructed the MAVEN-ARG dataset, achieving a full-process event understanding dataset from event detection to event argument extraction and event relation extraction;
Maverick: Efficient and Accurate Coreference Resolution Defying Recent Trends
Giuliano Martinelli (Sapienza University of Rome), Roberto Navigli (Sapienza University of Rome)
Computational EfficiencyRepresentation LearningTransformerText
🎯 What it does: Propose the Maverick framework, achieving efficient and accurate coreference resolution using a lightweight Encoder-only model;
MC^2: Towards Transparent and Culturally-Aware NLP for Minority Languages in China
Chen Zhang (Peking University), Yansong Feng (Peking University)
ClassificationRecognitionRepresentation LearningData-Centric LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Built and publicly released the largest, quality-focused open corpus MC² for four minority languages in China (Tibetan, Uyghur, Kazakh, Mongolian), and conducted continuous pre-training on it to obtain two usable models: MC² XLMR-large and MC² Llama-13B.
mCoT: Multilingual Instruction Tuning for Reasoning Consistency in Language Models
Huiyuan Lai (Center for Language and Cognition University of Groningen), Malvina Nissim (Center for Language and Cognition University of Groningen)
TransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Proposed and implemented a multi-lingual Chain-of-Thought (mCoT) instruction fine-tuning framework to improve the consistency and accuracy of large language models in multi-lingual mathematical reasoning tasks.
Measuring Meaning Composition in the Human Brain with Composition Scores from Large Language Models
Changjiang Gao (Nanjing University), Shujian Huang (Nanjing University)
TransformerLarge Language ModelTextBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposed a Composition Score based on Transformer FFN to quantify the combination of meanings in sentence understanding.
Measuring Political Bias in Large Language Models: What Is Said and How It Is Said
Yejin Bang (Hong Kong University of Science and Technology), Pascale Fung (Hong Kong University of Science and Technology)
Explainability and InterpretabilityLarge Language ModelText
🎯 What it does: This paper proposes a two-layer framework to measure political bias in large language models (LLMs) when generating content on political issues. It evaluates the model's stance on various topics and further decomposes the framework bias into content and style dimensions, quantified through semantic similarity, frame dimensions/entity frequency, and sentiment polarity.
Media Framing: A typology and Survey of Computational Approaches Across Disciplines
Yulia Otmakhova (University of Melbourne), Lea Frermann (University of Melbourne)
TextReview/Survey Paper
🎯 What it does: Conduct an interdisciplinary review of the concept of framing in media, propose a typology of frames, evaluate the coverage of existing NLP methods across different levels (semantic, cognitive, communicative) of framing, and highlight the fragmentation and theoretical gaps in current research;
MEFT: Memory-Efficient Fine-Tuning through Sparse Adapter
Jitai Hao (Shandong University), Zhaochun Ren (Leiden University)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: Proposed a MEFT (Memory-Efficient Fine-Tuning) framework based on sparse adapters, leveraging FFN activation sparsity and MoE structures. It places large-scale trainable parameters on the CPU and dynamically loads only a small number of parameters related to the input to the GPU during forward/backward propagation, achieving efficient fine-tuning of large-scale adapters.
MELA: Multilingual Evaluation of Linguistic Acceptability
Ziyin Zhang (Shanghai Jiao Tong University), Hai Hu (Shanghai Jiao Tong University)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes MELA—a multilingual acceptability judgment benchmark containing 46k sentences across 10 languages—and evaluates multiple large language models, conducts cross-lingual transfer experiments, and probes syntactic knowledge based on this benchmark.
MELoRA: Mini-Ensemble Low-Rank Adapters for Parameter-Efficient Fine-Tuning
Pengjie Ren (Shandong University), Jiahuan Pei (Centrum Wiskunde & Informatica)
Computational EfficiencyTextBenchmark
🎯 What it does: Propose a parameter-efficient fine-tuning method called MELoRA, which enhances model expressiveness by parallel stacking of multiple mini LoRA modules.
MemeGuard: An LLM and VLM-based Framework for Advancing Content Moderation via Meme Intervention
Prince Jha (Indian Institute of Technology Patna), Pushpak Bhattacharyya (Indian Institute of Technology Bombay)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: Built a framework named MemeGuard to detect and actively intervene in harmful meme content, comprising three modules: meme parsing, knowledge selection, and intervention generation.
Mementos: A Comprehensive Benchmark for Multimodal Large Language Model Reasoning over Image Sequences
Xiyao Wang (University of Maryland), Furong Huang (University of Maryland)
TransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: This paper proposes the Mementos benchmark to evaluate the capabilities of multimodal large language models (MLLMs) in image sequence reasoning.
MentalManip: A Dataset For Fine-grained Analysis of Mental Manipulation in Conversations
Yuxin Wang (Dartmouth College), Soroush Vosoughi (Dartmouth College)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper creates and publicly releases the MENTALMANIP dataset, which contains 4,000 multi-turn fictional dialogues, for fine-grained detection of psychological manipulation.
MERA: A Comprehensive LLM Evaluation in Russian
Alena Fenogenova (SaluteDevices), Sergey Markov
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed the MERA benchmark, covering 21 Russian instruction-based tasks (10 skills), and provided a unified experimental process, automatic evaluation, and public leaderboard;
Meta-Task Prompting Elicits Embeddings from Large Language Models
Yibin Lei (University of Amsterdam), Andrew Yates (University of Amsterdam)
Representation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose MetaEOL—a method for unsupervised generation of sentence embeddings directly from LLMs through meta-task prompting.
Meta-Tuning LLMs to Leverage Lexical Knowledge for Generalizable Language Style Understanding
Ruohao Guo (Georgia Institute of Technology), Alan Ritter (Georgia Institute of Technology)
ClassificationMeta LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Studied using style lexicons for meta-training to enable large language models to better identify unseen writing styles under zero-shot conditions.
Metaphor Understanding Challenge Dataset for LLMs
Xiaoyu Tong (University of Amsterdam), Ekaterina Shutova (University of Amsterdam)
ClassificationGenerationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Created a large-scale MUNCH dataset containing over 10k semantic replacements (appropriate) and 1.5k inappropriate metaphor sentences, and defined two evaluation tasks: sentence judgment and sentence generation;
MetaSumPerceiver: Multimodal Multi-Document Evidence Summarization for Fact-Checking
Ting-Chih Chen (Virginia Tech), Chris Thomas (Virginia Tech)
ClassificationGenerationRetrievalExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposed and implemented the MetaSumPerceiver (MSP) model for generating fact-checking-specific summaries from multimodal, multi-document inputs (text, images); simultaneously constructed the new Multi-News-Fact-Checking dataset.
MIDGARD: Self-Consistency Using Minimum Description Length for Structured Commonsense Reasoning
Inderjeet Nair (University of Michigan), Lu Wang (University of Michigan)
TextGraphChain-of-Thought
🎯 What it does: This study proposes a method called MIDGARD for generating reasoning graphs from natural language input to perform structured common-sense reasoning. The method improves reasoning accuracy by employing a self-consistency strategy that samples multiple reasoning chains and performs majority voting.
MindMap: Knowledge Graph Prompting Sparks Graph of Thoughts in Large Language Models
Yilin Wen (University of Illinois Urbana Champaign), Jimeng Sun (University of Illinois Urbana Champaign)
Explainability and InterpretabilityLarge Language ModelPrompt EngineeringGraphBiomedical DataRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the MindMap framework, which utilizes knowledge graphs (KG) as an external knowledge source, enabling large language models (LLM) to understand graph structures and generate answers along with their 'mind maps' through schematic reasoning, thereby enhancing transparency and accuracy.
MinPrompt: Graph-based Minimal Prompt Data Augmentation for Few-shot Question Answering
Xiusi Chen (University of California, Los Angeles), Wei Wang (University of California, Los Angeles)
Data SynthesisData-Centric LearningGraph Neural NetworkPrompt EngineeringText
🎯 What it does: Proposes MINPROMPT, an unsupervised data augmentation framework based on sentence graphs and minimum dominating sets, which generates high-quality QA training pairs under the premise of few examples to enhance few-shot QA performance.
Mirror: Multiple-perspective Self-Reflection Method for Knowledge-rich Reasoning
Hanqi Yan (King's College London), Yulan He (King's College London)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed a multi-perspective self-reflection method called Mirror, which helps large language models self-improve in knowledge-intensive reasoning tasks;
Missci: Reconstructing Fallacies in Misrepresented Science
Max Glockner (TU Darmstadt), Iryna Gurevych (TU Darmstadt)
ClassificationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the MISSCI dataset and task, utilizing large language models to automatically reconstruct and classify implicit logical fallacy reasoning in miscommunicated scientific claims;
Mission: Impossible Language Models
Julie Kallini (Stanford University), Christopher Potts (Stanford University)
Explainability and InterpretabilityRepresentation LearningData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: This paper constructs multiple synthetic unlearnable languages based on English, training the GPT-2 model to evaluate the learnability of these languages.
MIST: Mutual Information Maximization for Short Text Clustering
Krissanee Kamthawee (Vidyasirimedhi Institute of Science and Technology), Sarana Nutanong (Vidyasirimedhi Institute of Science and Technology)
Representation LearningTransformerContrastive LearningText
🎯 What it does: Propose the MIST framework, which enhances the representation quality of short text clustering by simultaneously maximizing sequence-level and word-level mutual information during the representation learning phase, and employs k-means clustering and KL divergence in the clustering phase.
Mitigating Biases for Instruction-following Language Models via Bias Neurons Elimination
Nakyeong Yang (Seoul National University), Kyomin Jung (Seoul National University)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerText
🎯 What it does: Proposes the concept of 'biased neurons' and proves their existence, designing a bias elimination method called CRISPR based on interpretability attribution to mitigate social and linguistic biases in language models under zero-shot instruction following scenarios.
Mitigating Catastrophic Forgetting in Large Language Models with Self-Synthesized Rehearsal
Jianheng Huang (Xiamen University), Jinsong Su (Xiamen University)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Propose a self-synthesized replay (SSR) framework that generates synthetic instances using LLMs and replays them to alleviate catastrophic forgetting in large language models.
MM-SAP: A Comprehensive Benchmark for Assessing Self-Awareness of Multimodal Large Language Models in Perception
Yuhao Wang (Shanghai Jiao Tong University), Yu Wang (Shanghai Artificial Intelligence Laboratory)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes the MM-SAP benchmark to evaluate the self-awareness capability of multimodal large language models (MLLMs) in visual perception.
MMToM-QA: Multimodal Theory of Mind Question Answering
Chuanyang Jin (New York University), Tianmin Shu (Johns Hopkins University)
TransformerLarge Language ModelSupervised Fine-TuningVision-Language-Action ModelVideoTextMultimodalityBenchmark
🎯 What it does: Constructed the first multi-modal Theory of Mind (ToM) question-answering benchmark, MMToM-QA, and proposed the BIP-ALM (Bayesian Inverse Planning Accelerated by Language Models) model to infer agents' goals and beliefs from joint video and textual information.
Mobile-Bench: An Evaluation Benchmark for LLM-based Mobile Agents
Shihan Deng (University of Electronic Science and Technology of China), Shuo Shang (University of Electronic Science and Technology of China)
TransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Constructed Mobile-Bench, a mobile LLM agent evaluation platform supporting dual interaction modes of UI and API, and provided an evaluation dataset containing 832 tasks (SAST, SAMT, MAMT) and 103 callable APIs, further proposing the CheckPoint evaluation metric for fine-grained inspection of agent execution processes.
MobileSpeech: A Fast and High-Fidelity Framework for Mobile Zero-Shot Text-to-Speech
Shengpeng Ji (Zhejiang University), Zhou Zhao (Zhejiang University)
GenerationComputational EfficiencyTransformerPrompt EngineeringTextAudio
🎯 What it does: Developed MobileSpeech, a zero-shot text-to-speech (TTS) system deployable on mobile devices, leveraging parallel mask generation and fine-grained speaker prompting to achieve low-latency, high-quality synthesis.
Modality-Aware Integration with Large Language Models for Knowledge-Based Visual Question Answering
Junnan Dong (Hong Kong Polytechnic University), Xiao Huang (Hong Kong Polytechnic University)
Graph Neural NetworkTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes a KVQA framework called MAIL based on multimodal knowledge graphs and large language models (LLMs), which enables tight interaction and reasoning among images, knowledge graphs, and LLMs.
Model Composition for Multimodal Large Language Models
Chi Chen (Tsinghua University), Yang Liu (Alibaba Group)
Large Language ModelImageVideoMultimodalityPoint CloudBenchmarkAudio
🎯 What it does: Propose achieving zero-training expansion for multimodal large language models through model composition, constructing a composite model compatible with various modalities;
Monotonic Representation of Numeric Attributes in Language Models
Benjamin Heinzerling (RIKEN), Kentaro Inui (MBZUAI)
Explainability and InterpretabilityRepresentation LearningTransformerTabular
🎯 What it does: Investigated how language models internally represent numerical attributes and found that low-dimensional linear subspaces can monotonically encode these attributes.
MoPS: Modular Story Premise Synthesis for Open-Ended Automatic Story Generation
Yan Ma (Fudan University), Pengfei Liu (Shanghai Jiao Tong University)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a modular story premise synthesis method called MoPS, which decomposes story premises into modules such as theme, setting, characters, and plot, generates candidate sets for each module, and then extracts module paths and synthesizes complete premises using an LLM.
More frequent verbs are associated with more diverse valency frames: Efficient principles at the lexicon-grammar interface
Siyu Tao (Saarland University), Michael Hahn (Saarland University)
Text
🎯 What it does: Investigated the relationship between verb frequency and the diversity of their valency frames, proposing and testing the hypothesis that more frequent verbs in cross-linguistic corpora have richer valency frames.
More Victories, Less Cooperation: Assessing Cicero’s Diplomacy Play
Wichayaporn Wongkamjan (University of Maryland), Jordan Lee Boyd-Graber (University of Maryland)
TransformerLarge Language ModelText
🎯 What it does: This paper systematically evaluates the performance of AI Cicero in the Diplomacy game by constructing Abstract Meaning Representation (AMR) annotations and an intent parser, and measures communication abilities such as persuasion, deception, and identification of the opponent's identity.
Moûsai: Efficient Text-to-Music Diffusion Models
Flavio Schneider (ETH Zürich), Bernhard Schölkopf (MPI for Intelligent Systems)
GenerationTransformerDiffusion modelTextMultimodalityAudio
🎯 What it does: Propose the Moûsai text-to-music diffusion model, which can generate high-quality 48kHz stereo music lasting multiple minutes based on text descriptions.
MPCoder: Multi-user Personalized Code Generator with Explicit and Implicit Style Representation Learning
Zhenlong Dai (Zhejiang University), Jingyuan Chen (Zhejiang University)
GenerationRepresentation LearningAI Code AssistantTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes a multi-user personalized code generation framework called MPCODER, which can generate code that aligns with individual developers' coding habits and personal styles;
MT-Bench-101: A Fine-Grained Benchmark for Evaluating Large Language Models in Multi-Turn Dialogues
Ge Bai (Alibaba Group), Wanli Ouyang (Alibaba Group)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposes MT-Bench-101, a three-tier hierarchical fine-grained capability classification covering 13 multi-turn dialogue tasks, containing 1,388 dialogues and 4,208 turns, evaluating the performance of 21 large language models in multi-turn dialogue.
MTP: A Dataset for Multi-Modal Turning Points in Casual Conversations
Gia-Bao Ho (VinUniversity), Wray Buntine (VinUniversity)
ClassificationRecognitionTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodalityAudio
🎯 What it does: Study and propose the task of identifying turning points (TP) in multimodal dialogues, create a high-quality human-consistent MTP dataset, and propose the TPMaven framework for TP classification and detection.
Muffin or Chihuahua? Challenging Multimodal Large Language Models with Multipanel VQA
Yue Fan (University of California, Santa Cruz), Xin Wang
RecognitionData SynthesisTransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Proposed the MultipanelVQA benchmark to evaluate the ability of multimodal large language models (MLLMs) to understand multi-panel images (e.g., web screenshots, posters), along with systematic error analysis of the models.
MuggleMath: Assessing the Impact of Query and Response Augmentation on Math Reasoning
Chengpeng Li (University of Science and Technology of China), Chang Zhou (Alibaba Group)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper constructs two new datasets, AugGSM8K and AugMATH, by enhancing the original mathematical reasoning datasets GSM8K and MATH using GPT-3.5/GPT-4 for querying and reasoning path generation. Based on these datasets, the authors perform supervised fine-tuning on the LLaMA series models, resulting in the MuggleMath model, which achieves new state-of-the-art results for open-source models on both GSM8K and MATH.
MULFE: A Multi-Level Benchmark for Free Text Model Editing
Chenhao Wang (Chinese Academy of Sciences), Jun Zhao (Chinese Academy of Sciences)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed a multi-level free-text model editing benchmark (MULFE), systematically evaluated various editing methods on this benchmark, and introduced a simple and efficient method called SIDE based on context distillation.
Multi-Aspect Controllable Text Generation with Disentangled Counterfactual Augmentation
Yi Liu (Nanjing University), Wei Hu (Nanjing University)
GenerationTransformerPrompt EngineeringGenerative Adversarial NetworkText
🎯 What it does: Propose the MAGIC method to achieve multi-aspect controllable text generation;
Multi-Dimensional Optimization for Text Summarization via Reinforcement Learning
Sangwon Ryu (POSTECH), Jungseul Ok (POSTECH)
OptimizationReinforcement LearningText
🎯 What it does: Propose a text summarization framework based on multi-objective reinforcement learning, optimizing summaries through UniEval's four-dimensional evaluation (coherence, consistency, fluency, relevance)
Multi-Level Feedback Generation with Large Language Models for Empowering Novice Peer Counselors
Alicja Chaszczewicz (Stanford University), Diyi Yang (Stanford University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Developed a multi-layered feedback framework based on LLM to generate contextualized and structured feedback for novice peer counselors.
Multi-modal Preference Alignment Remedies Degradation of Visual Instruction Tuning on Language Models
Shengzhi Li (TIFIN Inc), Shichao Pei (University of Massachusetts Boston)
OptimizationData-Centric LearningReinforcement Learning from Human FeedbackSupervised Fine-TuningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: To address the degradation of language instruction capabilities in multi-modal large language models after visual instruction fine-tuning, this paper collects 5k samples of visual question-answer preference data and uses alignment methods such as Direct Preference Optimization (DPO) to fine-tune the model, restoring and enhancing both language and visual instruction performance.
Multi-Task Inference: Can Large Language Models Follow Multiple Instructions at Once?
Guijin Son (Yonsei University), Seungone Kim (KAIST AI)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Construct the MTI BENCH benchmark to evaluate the ability of large language models (LLMs) to simultaneously complete multiple tasks in a single inference
MultiLegalPile: A 689GB Multilingual Legal Corpus
Joel Niklaus (University of Bern), Daniel Ho (Stanford University)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Constructed and released a 689 GB multilingual legal corpus (MULTILEGALPILE), containing 24 languages and 17 jurisdictions, and further continued pretraining multilingual and monolingual legal pretraining models (Legal-XLM-R, Legal-XLM-LF, and 24 monolingual Legal-BERT).
Multimodal ArXiv: A Dataset for Improving Scientific Comprehension of Large Vision-Language Models
Lei Li (University of Hong Kong), Qi Liu (University of Hong Kong)
Large Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: Constructed the Multimodal ArXiv dataset, including ArXivCap and ArXivQA, to enhance the understanding capabilities of large vision-language models in scientific literature.
Multimodal Contextualized Semantic Parsing from Speech
Jordan Voas (University of Texas at Austin), Ray Mooney (University of Texas at Austin)
Explainability and InterpretabilityRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageMultimodalityBenchmarkAudio
🎯 What it does: Proposed the SPICE task for semantic parsing in multimodal context environments, constructed the VG-SPICE dataset based on vision and speech, and developed the AViD-SP model capable of dynamically updating knowledge graphs.
Multimodal Instruction Tuning with Conditional Mixture of LoRA
Ying Shen (Virginia Tech), Lifu Huang (Virginia Tech)
Computational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningMixture of ExpertsVision Language ModelMultimodality
🎯 What it does: Propose the Conditional Mixture of LoRA (MixLoRA) framework for multi-modal instruction tuning, dynamically constructing low-rank adaptation matrices to alleviate task interference.
Multimodal Prompt Learning with Missing Modalities for Sentiment Analysis and Emotion Recognition
Zirun Guo (Zhejiang University), Zhou Zhao (Zhejiang University)
ClassificationRecognitionTransformerPrompt EngineeringMultimodality
🎯 What it does: This paper proposes a multi-modal Transformer framework based on prompt learning, which can handle missing modalities in sentiment analysis and emotion recognition tasks;
Multimodal Reasoning with Multimodal Knowledge Graph
Junlin Lee (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
Computational EfficiencyRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningMultimodalityGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the MR-MKG method, which utilizes multi-modal knowledge graphs (MMKG) to map images and knowledge nodes into the LLM word embedding space via visual and knowledge adapters, and fuses them in the prompt to inject cross-modal knowledge into LLMs, thereby enhancing multi-modal reasoning capabilities.
Multimodal Table Understanding
Mingyu Zheng (Chinese Academy of Sciences), Weiping Wang (Beijing Normal University)
RecognitionData SynthesisLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose the multimodal table understanding task, construct the MMTab dataset, and develop the Table-LLaVA model based on LLaVA
Multipath parsing in the brain
Berta Franzluebbers (University of Georgia), John Hale
Explainability and InterpretabilityTransformerLarge Language ModelTextMagnetic Resonance Imaging
🎯 What it does: By combining the large language model BLOOM with an incremental generation dependency parser, the study investigates whether multi-path parsing exists in human sentence comprehension and associates it with brain fMRI signals;
MultiPICo: Multilingual Perspectivist Irony Corpus
Silvia Casola (University of Turin), Davide Bernardi (Amazon)
Data SynthesisTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Constructed and released MultiPICo, a multilingual, multi-perspective irony corpus containing 18,778 short dialogues from Twitter and Reddit.
Multistage Collaborative Knowledge Distillation from a Large Language Model for Semi-Supervised Sequence Generation
Jiachen Zhao (University Of Massachusetts Amherst), Andrew McCallum (Ibm Research Ai)
GenerationKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Propose a multi-stage collaborative knowledge distillation method (MCKD), which generates pseudo labels using a small amount of labeled data and a few examples prompted by a large language model (LLM). Subsequently, through cross-partition and multi-round distillation, the student model is gradually improved and ultimately achieves high performance on low-resource sequence generation tasks.
Must NLP be Extractive?
Steven Bird (Charles Darwin University)
TextAudio
🎯 What it does: This paper constructs a non-extractive NLP corpus tailored to the local context by conducting guided field tours in the Kabulwarnamyo community of Arnhem Land, Australia, manually recording and respeaking local spoken dialogues;
MYTE: Morphology-Driven Byte Encoding for Better and Fairer Multilingual Language Modeling
Tomasz Limisiewicz (Charles University in Prague), Luke Zettlemoyer (University of Washington)
Computational EfficiencyRepresentation LearningTransformerText
🎯 What it does: Proposed a morphology-based byte encoding scheme called MYTE to achieve fairer and more compact representations in multilingual text;
NACL: A General and Effective KV Cache Eviction Framework for LLM at Inference Time
Yilong Chen (Institute of Information Engineering, Chinese Academy of Sciences), Hua Wu (Baidu Inc)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposed a KV Cache eviction framework called NACL, which significantly reduces KV Cache memory usage during inference through one-time global eviction and multi-head random sampling, while maintaining the inference performance of large language models.
NaijaHate: Evaluating Hate Speech Detection on Nigerian Twitter Using Representative Data
Manuel Tonneau (World Bank), Samuel Fraiberger
ClassificationDomain AdaptationData-Centric LearningTransformerSupervised Fine-TuningTextBenchmark
🎯 What it does: Created the NaiJAHATE hate speech detection dataset and proposed the NAIJAXLM-T pre-trained model, evaluating its real-world effectiveness on Nigerian tweets.
Naming, Describing, and Quantifying Visual Objects in Humans and LLMs
Alberto Testoni (University of Amsterdam), Sandro Pezzelle (University of Amsterdam)
Large Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Evaluated the generation diversity of Vision-Language Large Models (VLLMs) in object naming, attribute description, and quantifier selection, attempting to measure model performance using human-generated label distributions.
Narrowing the Knowledge Evaluation Gap: Open-Domain Question Answering with Multi-Granularity Answers
Gal Yona (Google Research), Mor Geva (Tel Aviv University)
TransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposed a multi-granularity answer evaluation framework called GRANOLA QA, constructed a multi-granularity version of ENTITYQUESTIONS (GRANOLA-EQ), and designed a decoding method named DRAG based on answer aggregation to align with model uncertainty and provide answers at appropriate granularity.
Natural Language Satisfiability: Exploring the Problem Distribution and Evaluating Transformer-based Language Models
Tharindu Madusanka (University of Manchester), Riza Batista-Navarro (University of Manchester)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Conduct experiments on natural language satisfiability problems with varying computational complexity to evaluate the reasoning capabilities of Transformer-based language models and investigate the impact of the phase transition region of training data on model performance.
Navigate through Enigmatic Labyrinth A Survey of Chain of Thought Reasoning: Advances, Frontiers and Future
Zheng Chu (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
Large Language ModelPrompt EngineeringTextMultimodalityReview/Survey PaperBenchmarkChain-of-Thought
🎯 What it does: A systematic review of Chain-of-Thought (CoT) and its extensions (XoT), summarizing related methods, benchmarks, frontiers, and future research directions.
Navigating the Dual Facets: A Comprehensive Evaluation of Sequential Memory Editing in Large Language Models
Zihao Lin (Virginia Tech), Lifu Huang (Virginia Tech)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Conducted memory editing on large language models and systematically evaluated the impact of continuous editing on multiple core model capabilities.
Navigating the Metrics Maze: Reconciling Score Magnitudes and Accuracies
Tom Kocmi (Microsoft), Matt Post (Microsoft)
Data-Centric LearningTextBenchmark
🎯 What it does: This study systematically evaluates the relationship between the score differences (metric delta) of various MT automatic evaluation metrics and human judgments using the large-scale human assessment dataset ToShip23. It proposes a 'threshold-accuracy' mapping table under different metrics, helping researchers align metric differences with the human-perceived discriminability. The robustness of this mapping is validated across the WMT public dataset and various scenarios involving different languages, domains, and system relatedness.
Navigating the OverKill in Large Language Models
Chenyu Shi (Fudan University), Dahua Lin (Fudan University)
Safty and PrivacyTransformerContrastive LearningTextBenchmark
🎯 What it does: Systematically study the 'Overkill' phenomenon, characterized by excessive safety and rejection, in large language models, and propose a no-training, model-agnostic Self-Contrastive Decoding (Self-CD) method to reduce rejection rates.
Navigating the Shadows: Unveiling Effective Disturbances for Modern AI Content Detectors
Ying Zhou (University of Chinese Academy of Sciences), Le Sun (University of Chinese Academy of Sciences)
ClassificationAnomaly DetectionAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: Systematically evaluate the detection of ChatGPT-generated text and design 12 black-box perturbation methods to explore the robustness of detectors
NEO-BENCH: Evaluating Robustness of Large Language Models with Neologisms
Jonathan Zheng (Georgia Institute of Technology), Wei Xu (Georgia Institute of Technology)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes and implements NEO-BENCH, a benchmark to evaluate the robustness and understanding capabilities of large language models (LLMs) when encountering neologisms that have emerged in recent years.
Never Lost in the Middle: Mastering Long-Context Question Answering with Position-Agnostic Decompositional Training
Junqing He (International Digital Economy Academy), Jiaxing Zhang (International Digital Economy Academy)
GenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: To address the 'middle failure' issue of large language models in long-text multi-document question answering, this paper proposes a position-agnostic multi-step QA (PAM QA) training method. By decomposing tasks into question repetition, index prediction, and answer summarization, the method enhances the model's attention focusing and information retrieval capabilities in long contexts.
NewsBench: A Systematic Evaluation Framework for Assessing Editorial Capabilities of Large Language Models in Chinese Journalism
Miao Li (University of Melbourne), Yi Luo (State Key Laboratory of Media Convergence Production Technology and Systems)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the NewsBench framework for systematic evaluation of Chinese news editing domain LLMs' writing and safety performance;
NextLevelBERT: Masked Language Modeling with Higher-Level Representations for Long Documents
Tamara Czinczoll (Hasso Plattner Institute / University of Potsdam), Gerard De Melo (Hasso Plattner Institute / University of Potsdam)
ClassificationRetrievalRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: NextLevelBERT for long documents was constructed by performing masked language modeling on text block vectors.
NICE: To Optimize In-Context Examples or Not?
Pragya Srivastava (Microsoft Research), Amit Sharma (Microsoft Research)
ClassificationGenerationOptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The study investigates the interaction between task instruction refinement and in-context example (ICE) optimization in large language models, proposing a metric called NICE to measure the importance of ICE, and systematically evaluates the impact of different tasks and instructions on model performance.
Noise Correction on Subjective Datasets
Uthman Jinadu (Georgia State University), Yi Ding (Georgia State University)
Data-Centric LearningTransformerText
🎯 What it does: Combines multi-task learning with loss-based label correction to address annotator fatigue, opinion discrepancies, and noise issues in subjective datasets.
Not All Countries Celebrate Thanksgiving: On the Cultural Dominance in Large Language Models
Wenxuan Wang (Chinese University of Hong Kong), Michael Lyu
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Investigate the issue of cultural hegemony in large language models (LLMs) under multilingual scenarios and construct a benchmark to measure their cultural bias.
Not All Experts are Equal: Efficient Expert Pruning and Skipping for Mixture-of-Experts Large Language Models
Xudong Lu (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
Computational EfficiencyTransformerLarge Language ModelMixture of ExpertsTextBenchmark
🎯 What it does: This paper proposes a post-training based expert-level sparsification method, including expert pruning and dynamic skipping, to significantly reduce memory consumption and improve inference speed in Mixture-of-Experts LLMs.
NounAtlas: Filling the Gap in Nominal Semantic Role Labeling
Roberto Navigli (Sapienza University of Rome), Alessandro Scirè (Sapienza University of Rome)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningText
🎯 What it does: This paper constructs NounAtlas, a large-scale noun predicate lexicon, and generates the first silver-annotated noun semantic role labeling (SRL) dataset based on it. Training and validation on this dataset demonstrate a unified SRL method for both nouns and verbs.
NPHardEval: Dynamic Benchmark on Reasoning Ability of Large Language Models via Complexity Classes
Lizhou Fan (University of Michigan), Yongfeng Zhang (Rutgers University)
Data SynthesisTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Proposed NPHardEval, a dynamic algorithm problem benchmark based on computational complexity categories, for strictly evaluating the reasoning capabilities of large language models at P, NP-complete, and NP-hard levels;