CodeAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes a backdoor attack method called ICLAttack that does not require fine-tuning, inducing large language models to generate predefined labels by implanting trigger words in the in-context learning demonstration context;
Unlabeled Debiasing in Downstream Tasks via Class-wise Low Variance Regularization
Shahed Masoudian, Markus Schedl (Thomson Reuters Labs)
CodeClassificationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose an unlabeled debiasing method based on category variance regularization, enabling the encoder LM to generate representations that are sensitive to category information but have low sensitivity to protected attribute information in downstream tasks.
CodeClassificationGraph Neural NetworkLarge Language ModelContrastive LearningTextGraph
🎯 What it does: Propose a self-supervised pre-training framework based on heterogeneous graphs that jointly learns representations for posts, words, and emojis, and integrates the learned emoji vectors into downstream tasks;
Unlocking Memorization in Large Language Models with Dynamic Soft Prompting
Zhepeng Wang (George Mason University), Yanfu Zhang (William And Mary)
CodeExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: A framework based on dynamic soft prompts was studied and implemented to extract and measure the training data memorized by large language models (LLMs), significantly improving the detection accuracy of the model's internal memory.
🎯 What it does: Developed a sign language tokenizer based on Residual Vector Quantization (RVQ), converting continuous joint coordinate sequences of American Sign Language (ASL) gestures into discrete tokens.
Unsupervised End-to-End Task-Oriented Dialogue with LLMs: The Power of the Noisy Channel
Brendan King (University of California, Santa Cruz), Jeffrey Flanigan (University of California, Santa Cruz)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Built an end-to-end task-oriented dialogue system that relies solely on unlabeled user-agent dialogue text and API schema, utilizing LLM to infer hidden dialogue states, API calls, and system behaviors through a noisy channel model and expectation maximization (Hard-EM), thereby completing dialogue generation.
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBiomedical DataMagnetic Resonance ImagingAudio
🎯 What it does: In an fMRI experiment, six participants watched 8.3 hours of TV shows and movies, and the videos were annotated with multi-level (speech, objects, background story, summary, spatiotemporal) semantic annotations; subsequently, these annotations were processed through LLMs (e.g., Llama 2) and multimodal models (e.g., LLaVA-v1.5) to extract latent representations, a linear encoding model was built to predict brain activity, and variance partitioning analysis was conducted to examine the unique contributions of different levels and modalities.
Varying Sentence Representations via Condition-Specified Routers
Ziyong Lin (National Key Laboratory of General Artificial Intelligence, Bigai), Zilong Zheng (National Key Laboratory of General Artificial Intelligence, Bigai)
🎯 What it does: Proposes a conditional sentence representation framework named CSR based on a conditionally specified router to enhance the performance of a three-encoder model in conditional semantic textual similarity and knowledge graph completion tasks.
Verifiable, Debuggable, and Repairable Commonsense Logical Reasoning via LLM-based Theory Resolution
Armin Toroghi (University of Toronto), Scott Sanner (University of Toronto)
CodeExplainability and InterpretabilityTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose the LLM-TRes framework, embedding LLM as a theoretical reasoner into a verifiable reasoning process to address reasoning errors and hallucinations;
Verification and Refinement of Natural Language Explanations through LLM-Symbolic Theorem Proving
Xin Quan (University Of Manchester), Andre Freitas (University Of Manchester)
CodeExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Studied the collaboration between large language models and theorem provers to automatically verify and improve natural language explanations in natural language inference (NLI).
Video-LLaVA: Learning United Visual Representation by Alignment Before Projection
Bin Lin (Peking University Shenzhen Graduate School), Li Yuan (Peking University Shenzhen Graduate School)
CodeRepresentation LearningTransformerLarge Language ModelVision Language ModelContrastive LearningImageVideoTextMultimodality
🎯 What it does: Pre-align the visual features of images and videos to the language feature space, then input unified visual representations into large language models (LLMs) via a shared projection layer, achieving a vision-language model that processes images and videos simultaneously in one go.
CodeGenerationTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Addressing the task of generating descriptive captions for news videos that include entities (entity-aware video captioning), and propose a modular framework: first use a visual model to detect named entities in the video (Entity Perceiver), then use a large language model to extract relevant background from an external knowledge base (Knowledge Extractor), and finally integrate visual information, entities, and background into the caption generation model.
Virtual Personas for Language Models via an Anthology of Backstories
Suhong Moon (University of California Berkeley), David M. Chan (University of California Berkeley)
CodeGenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringTextTabular
🎯 What it does: Proposed and validated the Anthology method, which utilizes open-ended life narratives (backstories) generated by LLMs as prefixes to modulate the outputs of large language models, creating representative, diverse, and consistent virtual human personalities for behavioral research simulations.
Vision-Language Model Fine-Tuning via Simple Parameter-Efficient Modification
Ming Li (University of Tokyo), Masashi Sugiyama (RIKEN Center for Advanced Intelligence Project)
CodeClassificationComputational EfficiencyKnowledge DistillationSupervised Fine-TuningVision Language ModelImage
🎯 What it does: Propose CLIPFit, which achieves efficient fine-tuning by only fine-tuning specific internal parameters of CLIP without adding external parameters;
VLEU: a Method for Automatic Evaluation for Generalizability of Text-to-Image Models
Jingtao Cao (Chinese University of Hong Kong), Kam-Fai Wong (Chinese University of Hong Kong)
CodeGenerationTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposed the VLEU evaluation metric to automatically measure the generalization capability of text-to-image models; samples visual-textual pairs using LLM and calculates KL divergence scores by assessing semantic consistency with CLIP.
VPTQ: Extreme Low-bit Vector Post-Training Quantization for Large Language Models
Yifei Liu (Microsoft), Mao Yang (Microsoft)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes an extremely low-bit weight quantization method based on Vector Post-Training Quantization (VPTQ), which can compress LLMs to 2-4 bits while maintaining high accuracy.
Watch Every Step! LLM Agent Learning via Iterative Step-level Process Refinement
Weimin Xiong (Peking University), Sujian Li (Peking University)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningAgentic AIContrastive LearningBenchmark
🎯 What it does: Proposed and implemented the IPR framework, which refines training of LLM agents through an iterative step-wise process, enabling agents to obtain more detailed process supervision in interactive tasks and significantly improving action decision quality.
What do Large Language Models Need for Machine Translation Evaluation?
Shenbin Qian (University of Surrey), Fred Blain
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper systematically evaluates the information needs and prompting strategies of large language models (LLMs) in machine translation quality assessment, covering zero-shot, chain-of-thought (CoT), and few-shot methods;
🎯 What it does: This paper systematically reviews the usage of 'typological diversity' in NLP literature and proposes two quantitative metrics (Average Pairwise Language Distance MPSD and Grambank Grammatical Feature Coverage) to objectively assess the linguistic sample diversity in research.
What the Harm? Quantifying the Tangible Impact of Gender Bias in Machine Translation with a Human-centered Study
Beatrice Savoldi (Fondazione Bruno Kessler), Luisa Bentivogli (Fondazione Bruno Kessler)
CodeGenerationTransformerText
🎯 What it does: This paper quantitatively evaluates the service quality gap and economic cost caused by gender bias for female users through post-editing experiments on machine translation (MT) outputs in real user scenarios.
When Context Leads but Parametric Memory Follows in Large Language Models
Yufei Tao (Portland State University), Ameeta Agrawal (Portland State University)
CodeExplainability and InterpretabilityTransformerPrompt EngineeringText
🎯 What it does: This paper conducts experiments on the responses of nine mainstream large language models under different context scales to investigate how models allocate local context knowledge and global parameter knowledge in knowledge-consistent scenarios, and evaluates their hallucination tendencies.
When Reasoning Meets Information Aggregation: A Case Study with Sports Narratives
Yebowen Hu (University of Central Florida), Fei Liu (Emory University)
CodeGenerationData SynthesisLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper analyzes step-by-step descriptions of NBA basketball games to study the performance of large language models in information aggregation and reasoning tasks. It proposes a game narrative generation method called SPORTSGEN and evaluates the model's score calculation ability under different segmentation strategies (batch splitting, player splitting, holistic processing).
CodeTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Systematically evaluate the alignment between large language models (LLMs) in question answering (QA) and evidence selection tasks under long context environments, constructing a diverse test set through semantic relevance and structural diversity, and exploring error sources between the two tasks.
Why Does New Knowledge Create Messy Ripple Effects in LLMs?
Jiaxin Qin (University of Illinois Urbana Champaign), Heng Ji (Stanford University)
CodeExplainability and InterpretabilityLarge Language ModelText
🎯 What it does: This paper proposes and validates the GradSim metric to evaluate the effectiveness of information propagation (ripple effect) in language models after knowledge editing.
Will LLMs Replace the Encoder-Only Models in Temporal Relation Classification?
Gabriel Roccabruna (University of Trento), Giuseppe Riccardi (University of Trento)
CodeClassificationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Evaluate and conduct interpretability analysis on the performance and decision-making processes of seven open-source and closed-source LLMs in the Temporal Relation Classification task, comparing their performance with the RoBERTa encoder model.
With Ears to See and Eyes to Hear: Sound Symbolism Experiments with Multimodal Large Language Models
Tyler Loakman (University of Sheffield), Chenghua Lin (University of Sheffield)
CodeRecognitionGenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: This study designs three types of experiments—shape-sound, size-sound, and word-image-sound—to evaluate the cognitive ability of multimodal large language models (VLM/LLM) in understanding sound symbolism.
Word Alignment as Preference for Machine Translation
Qiyu Wu (University of Tokyo), Yoshimasa Tsuruoka (University of Tokyo)
CodeGenerationOptimizationTransformerLarge Language ModelText
🎯 What it does: This paper proposes a framework based on Word Alignment Preference (WAP) to improve word alignment in large language models (LLMs) for machine translation through direct preference optimization (DPO), thereby reducing hallucination and omission phenomena.
CodeGenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: In non-drug-related Reddit communities, stigmatizing language toward people who use drugs (PWUS) was studied, and large language models (LLMs) were utilized to remove posts containing targeted stigma, generating more empathetic text.
🎯 What it does: Propose a multi-modal data construction pipeline named W2C, which utilizes existing Vision-Language Models (VLMs) to generate descriptions through self-instruction and organizes them in the form of Python code;
WPO: Enhancing RLHF with Weighted Preference Optimization
Wenxuan Zhou (Zoom Video Communications), Chenguang Zhu
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose Weighted Preference Optimization (WPO), which addresses distribution gaps and improves LLM alignment by weighted simulation of on-policy learning through offline preference pairs.
Zero-shot Cross-domain Dialogue State Tracking via Context-aware Auto-prompting and Instruction-following Contrastive Decoding
Xiaoyu Dong (Hong Kong Polytechnic University), Xiao-Ming Wu (Hong Kong Polytechnic University)
CodeDomain AdaptationTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This paper proposes the CAPID framework, which generates dynamic slot queries through context-aware automatic prompting and achieves zero-shot cross-domain dialogue state tracking using instruction-following contrastive decoding.
Zero-shot Cross-Lingual Transfer for Synthetic Data Generation in Grammatical Error Detection
Gaetan Lopez Latouche (Ubisoft La Forge), Benjamin Swanson (Ubisoft La Forge)
CodeClassificationData SynthesisDomain AdaptationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Leveraging the zero-shot cross-lingual transfer capability of multilingual pre-trained models, first train a multilingual artificial error generation model, then generate synthetic error data for the target language, followed by two-stage fine-tuning to achieve multi-language grammar error detection without human annotations.
Zero-Shot Detection of LLM-Generated Text using Token Cohesiveness
Shixuan Ma (Beijing University of Posts and Telecommunications), Quan Wang (Beijing University of Posts and Telecommunications)
CodeAnomaly DetectionLarge Language ModelText
🎯 What it does: Propose a zero-shot LLM text detection method called TOCSIN based on token cohesiveness, and use it as a general module to enhance the performance of existing detectors.