ACL 2024 Papers — Page 8
Annual Meeting of the Association for Computational Linguistics · 940 papers
Retrieval-Augmented Multilingual Knowledge Editing
Weixuan Wang (University of Edinburgh), Alexandra Birch (University of Edinburgh)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose ReMaKE, a retrieval-enhanced multilingual knowledge editor that enables parameter-free knowledge updates across different languages for large language models (LLMs).
Revealing the Parametric Knowledge of Language Models: A Unified Framework for Attribution Methods
Haeun Yu (University of Copenhagen), Isabelle Augenstein (University of Copenhagen)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a unified evaluation framework to compare instance attribution (IA) with neuron attribution (NA), and designed NA-Instances and IA-Neurons methods to align the two attribution results, assessing their ability to reveal model parameter knowledge.
Revisiting Code Similarity Evaluation with Abstract Syntax Tree Edit Distance
Yewei Song (University of Luxembourg), Jacques Klein (University of Luxembourg)
Large Language ModelTextBenchmark
🎯 What it does: Re-evaluate code similarity evaluation metrics, extend Tree Edit Distance (TSED) to multiple programming languages, and compare it with structural similarity generated by GPT-4, BLEU, Jaccard, and other traditional metrics.
Revisiting Demonstration Selection Strategies in In-Context Learning
Keqin Peng (Beihang University), Dacheng Tao (Nanyang Technological University)
ClassificationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Investigated model-related factors in demo selection during in-context learning (ICL) in large language models, and proposed a demo selection method called TopK+ConE that combines data and model information; experiments verified the impact of demos on the model's understanding of test inputs.
Revisiting Knowledge Distillation for Autoregressive Language Models
Qihuang Zhong (Wuhan University), Dacheng Tao (Nanyang Technological University)
Knowledge DistillationTransformerLarge Language ModelText
🎯 What it does: This paper proposes an Adaptive Teaching Knowledge Distillation (ATKD) method to improve knowledge distillation in autoregressive language models.
Revisiting Structured Sentiment Analysis as Latent Dependency Graph Parsing
Chengjie Zhou (Wuhan University), Donghong Ji (Wuhan University)
ClassificationLarge Language ModelText
🎯 What it does: View structured sentiment analysis (SSA) as a latent dependency graph parsing problem, utilizing tree CRF and constrained Inside algorithm to model flat sentiment spans as latent subtrees, followed by a two-stage parser to achieve joint extraction of expressions and their holders/targets.
Reward-based Input Construction for Cross-document Relation Extraction
Byeonghu Na (KAIST), Il-chul Moon
Recurrent Neural NetworkTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose a sentence selection module REIC based on reinforcement learning to select sentences most helpful for relation reasoning in cross-document relation extraction tasks, thereby effectively extracting information and constructing inputs from long documents;
Rewriting the Code: A Simple Method for Large Language Model Augmented Code Search
Haochen Li (Nanyang Technological University), Zhiqi Shen (Nanyang Technological University)
RetrievalAI Code AssistantTransformerLarge Language ModelContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose the ReCo framework, which combines generative augmentation retrieval (GAR) with code rewriting in the codebase to align the style of example code with database code, thereby improving code retrieval performance.
RLHFPoison: Reward Poisoning Attack for Reinforcement Learning with Human Feedback in Large Language Models
Jiongxiao Wang (University of Wisconsin Madison), Chaowei Xiao (University of Wisconsin Madison)
Adversarial AttackReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper studies human preference label poisoning attacks during the RLHF (Reinforcement Learning with Human Feedback) process and proposes a three-step poisoning sample selection method called RankPoison, which can induce LLMs to generate longer texts while maintaining model safety alignment and achieve trigger word-based backdoor attacks.
Robust Frame-Semantic Models with Lexical Unit Trees and Negative Samples
Jacob Devasier (University of Texas at Arlington), Chengkai Li (University of Texas at Arlington)
RecognitionTransformerLarge Language ModelText
🎯 What it does: This paper proposes a new method for object recognition and frame identification in frame semantic parsing.
Robust Singing Voice Transcription Serves Synthesis
Ruiqi Li (Zhejiang University), Zhou Zhao (Zhejiang University)
RecognitionData SynthesisConvolutional Neural NetworkTransformerAudio
🎯 What it does: Developed a robust singing voice note transcription model, ROSVOT, for automatically annotating SVS data and enhancing synthesis quality.
RomanSetu: Efficiently unlocking multilingual capabilities of Large Language Models via Romanization
Jaavid J, Anoop Kunchukuttan (Nilekani Centre at AI4Bharat)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: By continuously pre-training and instruction fine-tuning an English LLM with Romanized text as an interface, the model incorporates multi-lingual non-Latin scripts, enhancing its cross-lingual capabilities.
RORA: Robust Free-Text Rationale Evaluation
Zhengping Jiang (Johns Hopkins University), Anqi Liu (Johns Hopkins University)
Explainability and InterpretabilityText
🎯 What it does: A robust free-text reasoning explanation evaluation method called RORA is studied, aiming to assess whether explanatory text truly provides new information rather than leaking labels.
Rule or Story, Which is a Better Commonsense Expression for Talking with Large Language Models?
Ning Bian (University of Chinese Academy of Sciences), Le Sun (Institute of Software, Chinese Academy of Sciences)
RetrievalRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Systematically compared and verified the effectiveness of using stories and rules in large language models (LLMs) for retrieving and utilizing common sense, and proposed a self-supervised iterative fine-tuning approach to improve story generation quality.
S^2GSL: Incorporating Segment to Syntactic Enhanced Graph Structure Learning for Aspect-based Sentiment Analysis
Bingfeng Chen (Guangdong University of Technology), Zhifeng Hao (Shantou University)
ClassificationGraph Neural NetworkTransformerText
🎯 What it does: Propose the S2GSL framework, which combines dual branches of segmented semantic graphs and syntactic latent graphs through adaptive fusion to enhance ABSA performance.
SAC-KG: Exploiting Large Language Models as Skilled Automatic Constructors for Domain Knowledge Graph
Hanzhu Chen (University of Science and Technology of China), Jieping Ye (University of Science and Technology of China)
GenerationTransformerLarge Language ModelTextGraphBenchmarkAgriculture RelatedRetrieval-Augmented Generation
🎯 What it does: Proposed a general framework for automatically constructing domain knowledge graphs (KG) based on large language models (LLM), named SAC-KG, which iteratively builds multi-layer domain KGs through three stages: Generator, Verifier, and Pruner.
SafeDecoding: Defending against Jailbreak Attacks via Safety-Aware Decoding
Zhangchen Xu (University of Washington), Radha Poovendran (University of Washington)
Adversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a security-aware decoding strategy, SafeDecoding, which utilizes probability information from the original model and a specially trained safety expert model to suppress jailbreak attacks, thereby enhancing the safety of LLMs.
Safety Alignment in NLP Tasks: Weakly Aligned Summarization as an In-Context Attack
Yu Fu (University of California, Riverside), Yue Dong (University of California, Riverside)
Adversarial AttackReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Investigate the safety alignment of large language models (LLMs) across different natural language processing (NLP) tasks, and propose context attacks leveraging tasks with weaker safety alignment (e.g., summarization), leading to decreased safety in other tasks.
SafetyBench: Evaluating the Safety of Large Language Models
Zhexin Zhang (Tsinghua University), Minlie Huang (Tsinghua University)
Safty and PrivacyLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Constructed SafetyBench, a bilingual multiple-choice safety evaluation benchmark containing 11,435 questions across 7 categories of safety issues.
Same Task, More Tokens: the Impact of Input Length on the Reasoning Performance of Large Language Models
Mosh Levy (Bar-Ilan University), Yoav Goldberg (Bar-Ilan University)
TransformerTextChain-of-Thought
🎯 What it does: Investigate the impact of input length on the reasoning performance of large language models (LLMs), and construct the scalable FLenQA QA reasoning dataset for systematic evaluation.
SAPT: A Shared Attention Framework for Parameter-Efficient Continual Learning of Large Language Models
Weixiang Zhao (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
Computational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose the SAPT framework, which aligns the learning and selection of parameter-efficient fine-tuning (PET) blocks through shared attention, enabling LLMs to achieve continuous learning without task IDs, significantly mitigating catastrophic forgetting and promoting knowledge transfer.
SBAAM! Eliminating Transcript Dependency in Automatic Subtitling
Marco Gaido (Fondazione Bruno Kessler), Luisa Bentivogli (Fondazione Bruno Kessler)
GenerationTransformerVideoTextMultimodality
🎯 What it does: Proposed the first fully end-to-end, transcription-free automatic caption generation model that directly outputs translation, segmentation, and timestamps.
SC2: Towards Enhancing Content Preservation and Style Consistency in Long Text Style Transfer
Jie Zhao (Xidian University), Yue Jiang (Xidian University)
GenerationTransformerContrastive LearningText
🎯 What it does: Propose an SC2 model for long-text style transfer, focusing on enhancing content preservation and style consistency.
SceMQA: A Scientific College Entrance Level Multimodal Question Answering Benchmark
Zhenwen Liang (University of Notre Dame), Xiangliang Zhang (University of Notre Dame)
TransformerLarge Language ModelVision Language ModelImageTextMultimodalityTabularBenchmark
🎯 What it does: Proposes SceMQA, a multimodal science question-answering benchmark targeting the college entrance exam level, covering four subjects: mathematics, physics, chemistry, and biology;
SciMON: Scientific Inspiration Machines Optimized for Novelty
Qingyun Wang (University of Illinois at Urbana-Champaign), Tom Hope (Allen Institute for Artificial Intelligence)
GenerationTransformerLarge Language ModelContrastive LearningTextGraphRetrieval-Augmented Generation
🎯 What it does: Propose the SCIMON framework to achieve literature-based scientific direction generation, taking contextual input and outputting novel natural language suggestions
Search-Adaptor: Embedding Customization for Information Retrieval
Jinsung Yoon (Google Cloud AI), Tomas Pfister (Google Cloud AI)
RetrievalRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityBenchmark
🎯 What it does: Propose the Search-Adaptor method, which performs low-cost, controllable fine-tuning on text or multimodal embeddings of pre-trained LLMs to improve information retrieval performance.
SeeClick: Harnessing GUI Grounding for Advanced Visual GUI Agents
Kanzhi Cheng (Nanjing University), Zhiyong Wu (Shanghai AI Laboratory)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVision-Language-Action ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposes SeeClick, a visual GUI agent that can perform click and typing operations relying solely on screen screenshots.
SeeGULL Multilingual: a Dataset of Geo-Culturally Situated Stereotypes
Mukul Bhutani (Google Research), Sunipa Dev (Google Research)
Data-Centric LearningTransformerLarge Language ModelText
🎯 What it does: The paper created a cross-lingual, cross-cultural stereotype dataset called SeeGULL Multilingual, covering 23 regions, 20 languages, and containing over 25,861 stereotypes based on nationality and region, annotated as stereotypes and their level of offensiveness.
SEER: Facilitating Structured Reasoning and Explanation via Reinforcement Learning
Guoxin Chen (Institute of Computing Technology, Chinese Academy of Sciences), Yiming Qian (Shanghai Artificial Intelligence Laboratory)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextGraph
🎯 What it does: Propose the SEER method, which combines reinforcement learning with structured rewards to achieve high-quality tree/graph structure reasoning and explanation generation.
SEGO: Sequential Subgoal Optimization for Mathematical Problem-Solving
Xueliang Zhao (The University of Hong Kong), Lingpeng Kong (The University of Hong Kong)
OptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes the SEGO (Sequential Subgoal Optimization) framework, which leverages LLMs to generate, optimize, and select subgoals for problem decomposition, significantly enhancing mathematical problem-solving capabilities.
Selene: Pioneering Automated Proof in Software Verification
Lichen Zhang (Peking University), Nan Duan (Microsoft Research Asia)
TransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the Selene benchmark to evaluate the automatic proof generation capability of large language models (LLMs) in industrial-scale software verification (based on the seL4 microkernel), and build a lightweight verification environment along with a complete end-to-end proof generation pipeline.
Self-Alignment for Factuality: Mitigating Hallucinations in LLMs via Self-Evaluation
Xiaoying Zhang (Chinese University of Hong Kong), Helen Meng (Chinese University of Hong Kong)
OptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Leverage the LLM's self-evaluation capability to conduct self-factuality assessment of generated text, and enhance factual accuracy through self-alignment
Self-Augmented In-Context Learning for Unsupervised Word Translation
Yaoyiran Li (University of Cambridge), Ivan Vulić (University of Cambridge)
Representation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a self-incremental context learning-based unsupervised bilingual lexicon induction method called SAIL, which leverages LLMs to generate high-confidence word pairs and iteratively improves translation quality through repeated iterations.
Self-chats from Large Language Models Make Small Emotional Support Chatbot Better
Zhonghua Zheng (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
Knowledge DistillationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Utilize large language models (LLMs) as 'tutors' to iteratively expand the emotional support dialogue dataset (ExTES), and employ Diverse Response Inpainting (DRI) to generate rich response examples for fine-tuning the small ChatPal model, thereby enhancing the performance of small models in emotional support tasks.
Self-Contrast: Better Reflection Through Inconsistent Solving Perspectives
Wenqi Zhang (Zhejiang University), Weiming Lu (Zhejiang University)
Explainability and InterpretabilityComputational EfficiencyTransformerPrompt EngineeringContrastive LearningTextChain-of-Thought
🎯 What it does: Propose the Self-Contrast method, which first allows the LLM to generate multiple solution perspectives autonomously, then compares answers from different perspectives, extracts differences, generates a checklist, and uses this checklist to guide the LLM's self-correction.
Self-Distillation Bridges Distribution Gap in Language Model Fine-Tuning
Zhaorui Yang (Zhejiang University), Qian Liu (Sea AI Lab)
Knowledge DistillationTransformerSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose the Self-Distillation Fine-Tuning (SDFT) method, which bridges the gap between task data and the original model distribution during large language model fine-tuning by leveraging 'distilled' data generated by the model itself, thereby mitigating catastrophic forgetting.
Self-Evolving GPT: A Lifelong Autonomous Experiential Learner
Jinglong Gao (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
Meta LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed and implemented a lifelong autonomous experience learning framework called SE-GPT based on LLM, which can automatically identify task types, perform experience transfer, automatically practice, and generalize experience, thereby continuously accumulating and updating task experience without requiring manually written experience, and using stored experience to answer user questions.
Self-Modifying State Modeling for Simultaneous Machine Translation
Donglei Yu (State Key Laboratory of Multimodal Artificial Intelligence Systems Institute of Automation Chinese Academy of Sciences), Chengqing Zong (State Key Laboratory of Multimodal Artificial Intelligence Systems Institute of Automation Chinese Academy of Sciences)
GenerationTransformerText
🎯 What it does: Propose a self-modification state modeling (SM²) training framework that evaluates read/write decisions at each state through a self-modification process and thoroughly explores all potential states via prefix sampling, without constructing complete decision paths;
Self-Training with Direct Preference Optimization Improves Chain-of-Thought Reasoning
Tianduo Wang (Singapore University of Technology and Design), Wei Lu (Singapore University of Technology and Design)
OptimizationTransformerSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Propose a method that integrates Direct Preference Optimization (DPO) into a self-training framework to enhance the chain-of-thought reasoning ability of small language models in mathematical reasoning tasks.
Self-Training with Pseudo-Label Scorer for Aspect Sentiment Quad Prediction
Yice Zhang (Harbin Institute of Technology), Ruifeng Xu (Harbin Institute of Technology)
ClassificationData-Centric LearningTransformerLarge Language ModelScore-based ModelText
🎯 What it does: Propose a self-training framework based on a pseudo-label scorer for the Aspect Sentiment Quad Prediction (ASQP) task, enhancing model performance through the filtering of pseudo-labels.
Semi-Supervised Spoken Language Glossification
Huijie Yao (University of Science and Technology of China), Houqiang Li (Baidu Inc)
GenerationTransformerSupervised Fine-TuningTextAudio
🎯 What it does: Propose a semi-supervised spoken language to gloss translation framework called S3LG, which utilizes a large amount of monolingual data to generate pseudo labels, iteratively expanding the training data and improving model performance.
Semiparametric Token-Sequence Co-Supervision
Hyunji Lee (KAIST AI), Minjoon Seo (KAIST AI)
RetrievalRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Proposed a semi-parametric word sequence co-supervised training method, enabling a single autoregressive language model to receive supervision simultaneously in both parametric word embedding space and non-parametric sequence embedding space.
Semisupervised Neural Proto-Language Reconstruction
Liang Lu (Carnegie Mellon University), David Mortensen
GenerationData SynthesisRecurrent Neural NetworkTransformerContrastive LearningText
🎯 What it does: Propose a semi-supervised protolanguage reconstruction method under limited historical records, and design a bidirectional reconstructor (DPD-BiReconstructor) to simultaneously predict proto-words and their descendant words;
Shifting Attention to Relevance: Towards the Predictive Uncertainty Quantification of Free-Form Large Language Models
Jinhao Duan (Drexel University), Kaidi Xu (Drexel University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Explored how to improve uncertainty quantification in free-text LLMs by focusing on semantic relevance, and proposed the Shift Attention to Relevance (SAR) method.
Shoulders of Giants: A Look at the Degree and Utility of Openness in NLP Research
Surangika Ranathunga (Massey University), Aloka Fernando (University of Moratuwa)
Data-Centric LearningText
🎯 What it does: This paper quantifies the openness of the NLP community and the benefits brought by openness by semi-automatically extracting and manually annotating NLP papers from the ACL Anthology between 2015-2023. It systematically evaluates the reuse and publication status of research outcomes (data, code, pre-trained models), and analyzes the impact of different publication scenarios (Main, LREC, Other) and language resource levels on open practices.
Sign Language Translation with Sentence Embedding Supervision
Yasser Hamidullah (German Research Center for Artificial Intelligence), Cristina España-Bonet (German Research Center for Artificial Intelligence)
GenerationRepresentation LearningConvolutional Neural NetworkTransformerVision Language ModelVideoTextMultimodality
🎯 What it does: Propose an end-to-end sign language translation model that uses an unlabeled gloss as an intermediate representation, leveraging sentence embeddings (SEM) as a supervisory signal to achieve video-to-text translation;
Silent Signals, Loud Impact: LLMs for Word-Sense Disambiguation of Coded Dog Whistles
Julia Kruk, Diyi Yang
ClassificationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Leveraged large language models (LLMs) for word sense disambiguation, identifying and distinguishing dog whistles (coded political language) from common usage, ultimately constructing the Silent Signals dataset, which contains 16,550 high-confidence dog whistle instances;
Simple but Effective Compound Geometric Operations for Temporal Knowledge Graph Completion
Rui Ying (Nankai University), Renhong Cheng (Nankai University)
Representation LearningGraphTime Series
🎯 What it does: Propose and implement a temporal knowledge graph embedding model named TCompoundE, which uses combined translation and scaling geometric operations to capture entity, relation, and temporal information to complete knowledge graph completion tasks.
Simpson’s Paradox and the Accuracy-Fluency Tradeoff in Translation
Zheng Wei Lim (University of Melbourne), Charles Kemp (University of Melbourne)
GenerationTransformerText
🎯 What it does: Investigates the trade-off between translation accuracy and fluency at the source sentence level, and links it to a positive correlation at the corpus level, revealing their relationship as Simpson's paradox.
Simul-LLM: A Framework for Exploring High-Quality Simultaneous Translation with Large Language Models
Victor Agostinelli (Oregon State University), Lizhong Chen (Oregon State University)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Developed the Simul-LLM framework to achieve high-quality synchronous translation fine-tuning and evaluation for LLMs;
SIP: Injecting a Structural Inductive Bias into a Seq2Seq Model by Simulation
Matthias Lindemann (University of Edinburgh), Ivan Titov (University of Amsterdam)
Representation LearningTransformerSupervised Fine-TuningTextSequential
🎯 What it does: Inject structured inductive bias into seq2seq models by first pretraining Transformers to simulate input/output relationships of finite state transducers (FST), then fine-tuning with tunable prefixes on downstream tasks.
SirLLM: Streaming Infinite Retentive LLM
Yao Yao (Shanghai Jiao Tong University), Hai Zhao (Shanghai Jiao Tong University)
Large Language ModelText
🎯 What it does: Propose SirLLM, which utilizes token entropy to filter key tokens in the KV cache, maintaining long-term memory in infinite-length dialogues without requiring fine-tuning;
Sketch-Guided Constrained Decoding for Boosting Blackbox Large Language Models without Logit Access
Saibo Geng (EPFL), Robert West (EPFL)
TransformerLarge Language ModelText
🎯 What it does: Proposed a two-stage method named Sketch-Guided Constrained Decoding (SketchGCD) to achieve constrained decoding on black-box large language models without logit access. First, a powerful black-box LLM generates a sketch (unconstrained output), followed by a local lightweight auxiliary model performing constraint correction based on the sketch.
Skin-in-the-Game: Decision Making via Multi-Stakeholder Alignment in LLMs
Bilgehan Sel (Virginia Tech), Ming Jin (Virginia Tech)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposed the Skin‑in‑the‑Game (SKIG) framework, which enhances the ethical reasoning capabilities of large language models by leveraging multi-round multi-agent reasoning and simulating responsibility.
Small But Funny: A Feedback-Driven Approach to Humor Distillation
Sahithya Ravi (University of British Columbia), Arash Einolghozati (Meta AI)
GenerationKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningText
🎯 What it does: Knowledge distillation for humor generation in large language models (LLMs), where the LLM simultaneously assumes the roles of teacher and critic (reviewer) to train and iteratively improve small language models (SLMs).
Small Models, Big Insights: Leveraging Slim Proxy Models To Decide When and What to Retrieve for LLMs
Jiejun Tan (Renmin University of China), Ji-Rong Wen (Renmin University of China)
RetrievalComputational EfficiencyKnowledge DistillationTransformerTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: A smaller proxy language model is used to generate 'heuristic answers,' and based on these answers, it is determined whether the LLM needs to perform retrieval. If retrieval is required, query rewriting and filtering are conducted, ultimately enhancing the quality of retrieval-augmented generation in question-answering.
Soft Knowledge Prompt: Help External Knowledge Become a Better Teacher to Instruct LLM in Knowledge-based VQA
Qunbo Wang (Institute of Automation, Chinese Academy of Sciences), Jing Liu (Institute of Automation, Chinese Academy of Sciences)
Prompt EngineeringVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose a soft knowledge prompting (SKP) method that converts external knowledge into soft prompt vectors, which are fused with visual embeddings to guide LLMs in knowledge-driven visual question answering.
Soft Self-Consistency Improves Language Models Agents
Han Wang (University of North Carolina Chapel Hill), Mohit Bansal (University of North Carolina Chapel Hill)
Computational EfficiencyLarge Language ModelAgentic AITextSequential
🎯 What it does: Propose a new post-generation diversification selection method called Soft Self-Consistency (SOFT-SC), which replaces traditional majority voting with continuous scoring based on the token probabilities of generated answers;
SoftDedup: an Efficient Data Reweighting Method for Speeding Up Language Model Pre-training
Nan He (Tencent AI Lab), Yang Wei
Computational EfficiencyData-Centric LearningText
🎯 What it does: Propose a soft deduplication method called SoftDedup, which reduces the negative impact of duplicate data on the pretraining of large language models by calculating the commonness of samples and dynamically adjusting sampling weights;
SOTOPIA-π: Interactive Learning of Socially Intelligent Language Agents
Ruiyi Wang (Carnegie Mellon University), Hao Zhu (Carnegie Mellon University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Through the interactive learning method SOTOPIA-π, GPT-4 is utilized to automatically generate diverse social tasks and score the model's performance on these tasks. Subsequently, the model's social intelligence and safety are enhanced through two-stage training involving behavioral cloning and self-reinforcement.
Soul-Mix: Enhancing Multimodal Machine Translation with Manifold Mixup
Xuxin Cheng (Peking University), Yuexian Zou (Peking University)
Convolutional Neural NetworkTransformerVision Language ModelTextMultimodality
🎯 What it does: Propose the Soul-Mix framework, which utilizes manifold mixup to blend the translation results of complete and missing texts, thereby enhancing the performance of multimodal machine translation.
SpaRC and SpaRP: Spatial Reasoning Characterization and Path Generation for Understanding Spatial Reasoning Capability of Large Language Models
Md Imbesat Rizvi (Technical University of Darmstadt), Iryna Gurevych (Technical University of Darmstadt)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: This paper studies the spatial reasoning capabilities of large language models (LLMs) and proposes the SpaRC spatial reasoning characterization framework and the SpaRP reasoning path dataset based on symbolic reasoning;
SparseFit: Few-shot Prompting with Sparse Fine-tuning for Jointly Generating Predictions and Natural Language Explanations
Jesus Solano (ETH Zürich), Pasquale Minervini (University of Edinburgh)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose the SPARSEFIT method, which under limited natural language explanation (NLE) data, jointly generates task predictions and corresponding natural language explanations by combining sparse fine-tuning (updating only a small portion of model parameters) with prompt-based learning.
SparseFlow: Accelerating Transformers by Sparsifying Information Flows
Yeachan Kim (Korea University), SangKeun Lee (Korea University)
Computational EfficiencyTransformerMixture of ExpertsTextMultimodality
🎯 What it does: Propose SPARSEFLOW, which learns the information flow of sparse Transformers, reducing computational costs in self-attention and feed-forward layers of each layer.
Spatially-Aware Speaker for Vision-and-Language Navigation Instruction Generation
Muraleekrishna Gopinathan (Edith Cowan University), David Suter (Edith Cowan University)
GenerationRecurrent Neural NetworkVision-Language-Action ModelGenerative Adversarial NetworkMultimodality
🎯 What it does: Propose Spatially‑Aware Speaker (SAS), an Encoder‑Decoder model that generates navigation instructions by leveraging spatial structure and semantic information.
Speaker Verification in Agent-generated Conversations
Yizhe Yang (Beijing Institute of Technology), Ee-Peng Lim (Singapore Management University)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Constructed a speaker verification task and collected multi-source dialogue data to evaluate the personalization performance of role-play models.
Spectral Filters, Dark Signals, and Attention Sinks
Nicola Cancedda (Meta)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Proposed a spectral filters method, extending the logit lens to logit spectroscopy to investigate dark signals in LLM residual flows and their role in attention sinking
Speculative Contrastive Decoding
Hongyi Yuan (Tsinghua University), Chang Zhou (Alibaba Inc)
GenerationComputational EfficiencyLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: Propose the Speculative Contrastive Decoding (SCD) method, which accelerates inference using predictions from a small model while combining contrastive decoding to enhance generation quality.
Speech language models lack important brain-relevant semantics
Subba Reddy Oota (Inria), Mariya Toneva (Max Planck Institute for Software Systems)
Representation LearningTransformerLarge Language ModelBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Evaluate the impact of low-level textual, auditory, and visual features in language model representations on brain alignment in text and speech models during reading and listening tasks using fMRI recordings.
Speech Sense Disambiguation: Tackling Homophone Ambiguity in End-to-End Speech Translation
Tengfei Yu (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
TransformerContrastive LearningTextAudio
🎯 What it does: Constructed comprehensive homonym dictionaries for English, French, German, and Spanish, and annotated the MuST-C and CoVoST-2 end-to-end speech translation datasets based on these dictionaries; subsequently proposed the AmbigST model, which utilizes homonym masks and three-level contrastive learning to resolve speech homonym ambiguity.
Speech Translation with Speech Foundation Models and Large Language Models: What is There and What is Missing?
Marco Gaido (Fondazione Bruno Kessler), Luisa Bentivogli (Fondazione Bruno Kessler)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringReview/Survey PaperAudio
🎯 What it does: This paper systematically reviews the current research on combining Speech Foundation Models (SFM) with Large Language Models (LLM) for Speech-to-Text Translation (ST), proposes a unified framework, analyzes the architectures, training strategies, and evaluation methods of nine related papers, and identifies the shortcomings in current research as well as future improvement directions.
Speech vs. Transcript: Does It Matter for Human Annotators in Speech Summarization?
Roshan Sharma (Carnegie Mellon University), Bhiksha Raj (Carnegie Mellon University)
Large Language ModelTextBenchmarkRetrieval-Augmented GenerationAudio
🎯 What it does: This paper systematically evaluates the impact of directly listening to recordings versus reading transcribed text on generating summaries, and further explores the differences in summary quality caused by transcription errors, expert versus non-expert annotations.
SpikeVoice: High-Quality Text-to-Speech Via Efficient Spiking Neural Network
Kexin Wang (Chinese Academy of Sciences), Guoqi Li (Chinese Academy of Sciences)
GenerationSpiking Neural NetworkTransformerAudio
🎯 What it does: This paper proposes a high-quality text-to-speech (TTS) model called SpikeVoice based on Spiking Neural Networks (SNN), achieving for the first time the 'speaking' functionality of SNN.
Spiral of Silence: How is Large Language Model Killing Information Retrieval?—A Case Study on Open Domain Question Answering
Xiaoyang Chen (University of Chinese Academy of Sciences), Yingfei Sun (University of Chinese Academy of Sciences)
RetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Build and run an iterative simulation pipeline to systematically evaluate the short-term and long-term impacts of LLM-generated text in Retrieval-Augmented Generation (RAG) systems, particularly on open-domain question answering (ODQA) tasks, and observe whether the 'spiral of silence' effect occurs.
Split and Rephrase with Large Language Models
David Ponce (Fundación Vicomtech, Basque Research and Technology Alliance), Harritxu Gete (Fundación Vicomtech, Basque Research and Technology Alliance)
GenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper evaluates and improves the performance of large language models (LLMs) on the Split and Rephrase (SPRP) task, proposing multiple methods through LoRA fine-tuning and instruction fine-tuning;
SPOR: A Comprehensive and Practical Evaluation Method for Compositional Generalization in Data-to-Text Generation
Ziyao Xu (Peking University), Houfeng Wang (Peking University)
GenerationData-Centric LearningLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposes the SPOR (Systematicity, Productivity, Order invariance, Rule learnability) evaluation framework to comprehensively assess compositional generalization in data-to-text generation.
SportsMetrics: Blending Text and Numerical Data to Understand Information Fusion in LLMs
Yebowen Hu (University of Central Florida), Fei Liu (Emory University)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed the SportsMetrics benchmark to evaluate large language models' ability to integrate information and perform numerical reasoning on mixed text and numerical data, particularly in sports events (NBA, NFL), designing four adversarial tasks (new rules, lengthy descriptions, scrambled narratives, and statistical fill-in) and conducting experiments using LLMs.
SPZ: A Semantic Perturbation-based Data Augmentation Method with Zonal-Mixing for Alzheimer’s Disease Detection
FangFang Li, Jie Yin (University of Sydney)
ClassificationTransformerTextAlzheimer's Disease
🎯 What it does: Propose a Semantically Perturbed Regional Mixture (SPZ) data augmentation framework to enhance the robustness and accuracy of Alzheimer's Disease (AD) text detection.
Stealthy Attack on Large Language Model based Recommendation
Jinghao Zhang (Institute of Automation, Chinese Academy of Sciences), Liang Wang (Northeastern University)
Recommendation SystemAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: Proposed a text attack method that subtly rewrites product titles during the test phase in large language model recommendation systems, significantly increasing the exposure rate of target products;
Steering Llama 2 via Contrastive Activation Addition
Nina Rimsky (Anthropic), Alexander Turner
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: Propose the Contrastive Activation Addition (CAA) method, which calculates a steering vector by comparing the activation differences between positive and negative examples during the forward pass, and then adds this vector to all token positions to precisely control the behavior of LLMs.
StepCoder: Improving Code Generation with Reinforcement Learning from Compiler Feedback
Shihan Dou (Fudan University), Xuanjing Huang (Fudan University)
AI Code AssistantReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Improve code generation in large language models through reinforcement learning and compiler feedback, proposing the StepCoder framework.
STICKERCONV: Generating Multimodal Empathetic Responses from Scratch
Yiqun Zhang (Northeastern University), Kaisong Song (Alibaba Group)
GenerationData SynthesisTransformerLarge Language ModelAgentic AIVision Language ModelDiffusion modelTextMultimodality
🎯 What it does: Proposed a sticker-based multimodal empathetic dialogue system and constructed the first sticker empathetic dataset
STREAM: Simplified Topic Retrieval, Exploration, and Analysis Module
Anton Frederik Thielmann (Clausthal University of Technology), Gillian Kant (University of Göttingen)
RetrievalExplainability and InterpretabilityLarge Language ModelTextMultimodality
🎯 What it does: This paper proposes the STREAM framework, which integrates multiple topic models, evaluation metrics, interactive visualization, and interpretable downstream models to help non-technical users easily complete topic modeling and analysis.
StreamAtt: Direct Streaming Speech-to-Text Translation with Attention-based Audio History Selection
Sara Papi (Fondazione Bruno Kessler), Luisa Bentivogli (Fondazione Bruno Kessler)
Knowledge DistillationTransformerTextAudio
🎯 What it does: Proposes StreamAtt, a direct speech-to-text translation (StreamST) strategy for continuous unbounded audio streams, and introduces the StreamLAAL latency evaluation metric, which is comparable to SimulST.
StreamSpeech: Simultaneous Speech-to-Speech Translation with Multi-task Learning
Shaolei Zhang (Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences), Yang Feng (Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences)
RecognitionGenerationData SynthesisTransformerAudio
🎯 What it does: Proposed StreamSpeech, an 'integrated' end-to-end synchronous speech-to-speech translation model, capable of generating target speech in real-time while receiving continuous speech input, and providing intermediate ASR and text translation results;
StreamVoice: Streamable Context-Aware Language Modeling for Real-time Zero-Shot Voice Conversion
Zhichao Wang (Northwestern Polytechnical University), Yuping Wang (Northwestern Polytechnical University)
GenerationCompressionTransformerLarge Language ModelAudio
🎯 What it does: Propose StreamVoice, a streaming zero-shot acoustic conversion system based on language models, which can achieve real-time speech conversion without using future information.
Strengthened Symbol Binding Makes Large Language Models Reliable Multiple-Choice Selectors
Mengge Xue (Tencent), Chengguo Yin (Tencent)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper investigates the selection bias problem in multiple-choice questions (MCQ) during the supervised fine-tuning (SFT) stage of large language models and proposes the Point-wise Intelligent Feedback (PIF) method to significantly reduce bias and improve accuracy.
Striking Gold in Advertising: Standardization and Exploration of Ad Text Generation
Masato Mita (CyberAgent), Peinan Zhang (CyberAgent)
GenerationTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Standardize the task definition for Ad Text Generation (ATG), construct the first publicly available benchmark dataset CAMERA /camera_retro, and conduct multi-model, multi-modal baseline experiments and system-level evaluations on it.
STRUCTSUM Generation for Faster Text Comprehension
Parag Jain (Google DeepMind), Francesco Piccinno (University of Edinburgh)
GenerationLarge Language ModelPrompt EngineeringText
🎯 What it does: Leverage large language models to generate structured summaries (tables and mind maps) from text, and enhance output quality through segmented generation and iterative expansion.
Structured Tree Alignment for Evaluation of (Speech) Constituency Parsing
Freda Shi (Toyota Technological Institute at Chicago), Karen Livescu (Toyota Technological Institute at Chicago)
TextAudio
🎯 What it does: Proposed a new Structured Average IOU (STRUCT-IOU) metric for evaluating constituent syntactic parse trees of text and speech, particularly suitable for textless speech parsing.
Stumbling Blocks: Stress Testing the Robustness of Machine-Generated Text Detectors Under Attacks
Yichen Wang (Xi'an Jiaotong University), Tianxing He (University of Washington)
Anomaly DetectionAdversarial AttackText
🎯 What it does: This paper systematically evaluates the robustness of eight machine-generated text detectors under twelve real-world attacks (editing, rewriting, collaborative generation, prompting), constructing a complete attack budget and evaluation metric system;
Subtle Biases Need Subtler Measures: Dual Metrics for Evaluating Representative and Affinity Bias in Large Language Models
Abhishek Kumar (Brock University), Ali Emami (Brock University)
GenerationTransformerLarge Language ModelMixture of ExpertsTextBenchmark
🎯 What it does: This paper constructs a Creativity-Oriented Generation Toolkit (CoGS) and introduces Representativeness Bias Score (RBS) and Affinity Bias Score (ABS) to evaluate subtle biases in large language models during the generation and evaluation processes.
Superfiltering: Weak-to-Strong Data Filtering for Fast Instruction-Tuning
Ming Li (University of Maryland), Tianyi Zhou (University of Maryland)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the Superfiltering method, which uses weak models (e.g., GPT-2) to filter instruction tuning data, significantly accelerating and enhancing the training effectiveness of large LLMs.
Surgical Feature-Space Decomposition of LLMs: Why, When and How?
Arnav Chavan (Nyun AI), Deepak Gupta (Transmute AI Hub)
Computational EfficiencyKnowledge DistillationLarge Language ModelText
🎯 What it does: Applies hierarchical feature space low-rank decomposition (SFSD) to large language models, achieving training-free compression while maintaining or even enhancing reasoning and common sense reasoning performance.
SwapMoE: Serving Off-the-shelf MoE-based Large Language Models with Tunable Memory Budget
Rui Kong (Shanghai Jiao Tong University), Yunxin Liu (Institute for AI Industry Research Tsinghua University)
Computational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Designed and implemented the SwapMoE framework, enabling efficient inference of large-scale MoE (Mixture of Experts) language models on memory-constrained edge devices.
SyllabusQA: A Course Logistics Question Answering Dataset
Nigel Fernandez (University of Massachusetts Amherst), Andrew Lan (University of Massachusetts Amherst)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper constructs a public course schedule question-answering dataset called SYLLABUSQA and evaluates the factual accuracy and quality of LLM-generated answers based on this dataset.
Symbol-LLM: Towards Foundational Symbol-centric Interface For Large Language Models
Fangzhi Xu (Xi'an Jiaotong University), Jun Liu (National University Of Singapore)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: A series of Symbol-LLM models were constructed, utilizing a two-stage fine-tuning process (injection stage and infusion stage) to inject symbolic knowledge into LLMs, thereby enhancing symbolic task performance while maintaining general NLP capabilities.
SymKGQA: Few-Shot Knowledge Graph Question Answering via Symbolic Program Generation and Execution
Prerna Agarwal (Indian Institute of Technology Delhi), Srikanta Bedathur (Indian Institute of Technology Delhi)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Under cold start conditions, few-shot in-context learning (ICL) with a small number of examples is used to prompt large language models (LLMs) to generate progressive symbolic logical forms (KoPL), which are then matched and corrected with knowledge graph (KG) context via the RAG-enhanced QUACK parser, ultimately leading to the execution of answers.
Synchronized Video Storytelling: Generating Video Narrations with Structured Storyline
Dingyi Yang (Renmin University of China), Qin Jin (Renmin University of China)
GenerationTransformerLarge Language ModelSupervised Fine-TuningVideoTextMultimodality
🎯 What it does: This paper designs and implements a synchronized video storytelling task and the corresponding framework VideoNarrator, and constructs a new dataset named E-SyncVidStory based on this.
Synergetic Event Understanding: A Collaborative Approach to Cross-Document Event Coreference Resolution with Large Language Models
Qingkai Min (Zhejiang University), Yue Zhang (Shanghai AI Laboratory)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes a collaborative method that first uses a large language model to perform multi-step summarization of event abstracts, then inputs the summary information along with the original text into a small language model for fine-tuning, addressing cross-document event coreference resolution.