International Conference on Learning Representations · 1064 papers
Unbalancedness in Neural Monge Maps Improves Unpaired Domain Translation
Luca Eyring (Tübingen AI Center), Fabian J Theis
CodeImage TranslationDomain AdaptationDrug DiscoveryGenerative Adversarial NetworkImageBiomedical Data
🎯 What it does: The study investigates how to introduce imbalance into arbitrary neural Monge mappers within the framework of Unbalanced Optimal Transport (UOT) to enhance the performance of unpaired domain translation tasks.
🎯 What it does: This paper proposes an uncertainty-aware inverse constraint reinforcement learning framework (UAICRL) that can infer safety constraints from expert demonstrations and learn policies that adhere to these constraints.
Linlin Yu (University of Texas at Dallas), Feng Chen (University of Texas at Dallas)
CodeClassificationGraph Neural NetworkImage
🎯 What it does: The paper proposes a method for uncertainty quantification in hyperspectral image classification based on Graph Convolutional Networks (GCN). It models pixels using Evidential GCN and Graph Posterior Network, and incorporates Unmixing Regularization (UR) based on a physical mixing model and Total Variation Regularization (TV) based on edge preservation during training to enhance the detection capability for unknown classes (OOD) and misclassifications.
Unconstrained Stochastic CCA: Unifying Multiview and Self-Supervised Learning
James Chapman (University College London), Ana Lawry Aguila (University College London)
CodeOptimizationRepresentation LearningContrastive LearningImageMultimodalityBiomedical Data
🎯 What it does: This paper proposes a general GEP (Generalized Eigenvalue Problem) framework based on the Eckhart-Young unconstrained objective, unifying linear, deep, and multi-view CCA into a form suitable for optimization using stochastic gradient descent. Based on this, efficient algorithms such as CCA-EY, PLS-EY, DCCA-EY, and SSL-EY are designed.
🎯 What it does: Theoretical and experimental analysis of the provably robust network training mechanism based on Interval Bound Propagation (IBP) is conducted, and a measure of propagation tightness is proposed.
🎯 What it does: This paper explores whether domain generalization (DG) algorithms outperform traditional empirical risk minimization (ERM) in the presence of label noise, providing both theoretical and experimental analysis.
Understanding In-Context Learning from Repetitions
Jianhao Yan (Zhejiang University), Yue Zhang (Westlake Institute for Advanced Study)
CodeTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes a surface repetition-based word co-occurrence reinforcement mechanism to explain the behavior of large language models in context learning and the reasons for their successes and failures.
Understanding In-Context Learning in Transformers and LLMs by Learning to Learn Discrete Functions
Satwik Bhattamishra (University of Oxford), Varun Kanade (University of Oxford)
CodeTransformerLarge Language ModelSequential
🎯 What it does: This study investigates the contextual learning capabilities of Transformer and other architectures within an autoregressive framework for Boolean functions; it explores the enhancement of sample efficiency through teaching sequences and the potential of pre-trained LLMs to implement learning algorithms.
🎯 What it does: A more powerful dataset reconstruction attack is proposed, utilizing the Neural Tangent Kernel (NTK) and KKT conditions to recover all training samples from the trained network parameters, and this attack is associated with dataset distillation (KIP).
Understanding the Effects of RLHF on LLM Generalisation and Diversity
Robert Kirk (University College London), Roberta Raileanu (Meta)
CodeGenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper systematically evaluates three fine-tuning methods for large language models (SFT, RLHF, BoN) in terms of outlier distribution generalization and output diversity in summarization and instruction-following tasks, revealing that RLHF outperforms SFT in generalization but significantly reduces diversity, and proposes a multi-dimensional evaluation framework.
🎯 What it does: A lightweight defense method is proposed to inject random noise into intermediate layer features during the inference phase to reduce the success rate of query-based adversarial attacks.
Understanding when Dynamics-Invariant Data Augmentations Benefit Model-free Reinforcement Learning Updates
Nicholas Corrado, Josiah P. Hanna (University of Wisconsin)
CodeReinforcement Learning
🎯 What it does: This paper evaluates the impact of using dynamic invariant data augmentation on model-agnostic reinforcement learning in sparse reward tasks through experiments, exploring the roles of state-action coverage, reward density, and augmentation replay ratio.
Uni-RLHF: Universal Platform and Benchmark Suite for Reinforcement Learning with Diverse Human Feedback
Yifu Yuan (Tianjin University), YAN ZHENG
CodeAutonomous DrivingReinforcement Learning from Human FeedbackTransformerReinforcement LearningSequentialBenchmark
🎯 What it does: This paper proposes Uni-RLHF, a unified RLHF platform that includes annotation interfaces for multiple types of human feedback, a scalable crowdsourced annotation pipeline, and reproducible offline RLHF baselines.
🎯 What it does: We propose Uni3D, a scalable 3D foundation model with up to 1 billion parameters, which learns a unified 3D representation by transplanting the ViT structure as a point cloud encoder and aligning across modalities (point clouds, images, text) within a pre-training framework.
UniAdapter: Unified Parameter-Efficient Transfer Learning for Cross-modal Modeling
Haoyu Lu (Renmin University of China), Mingyu Ding (University of California)
CodeRetrievalTransformerVision Language ModelImageVideoTextMultimodality
🎯 What it does: A parameter-efficient cross-modal transfer learning framework named UniAdapter is designed, which can adapt to various cross-modal downstream tasks (such as retrieval, question answering, captioning, etc.) by fine-tuning only a small number of parameters on a frozen large-scale vision-language pre-trained model.
Unified Language-Vision Pretraining in LLM with Dynamic Discrete Visual Tokenization
Yang Jin (Peking University), Yadong MU
CodeGenerationData SynthesisRetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: A unified multimodal foundation model LaVIT is designed and trained, which converts images into discrete tokens that can be processed in parallel with text through a dynamic discrete visual tokenizer, predicting the next token in an autoregressive manner within the LLM, achieving understanding and generation of image-text pairs.
🎯 What it does: This paper proposes a 'Universal Backdoor Attack' that can misclassify any input into any target category during inference by contaminating only a very small proportion of samples (0.15%) in the training set.
Universal Jailbreak Backdoors from Poisoned Human Feedback
Javier Rando (ETH Zurich), Florian Tramèr (ETH Zurich)
CodeAdversarial AttackReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper studies the implantation of a universal 'jailbreak' backdoor by poisoning training data during the Reinforcement Learning from Human Feedback (RLHF) training process, allowing attackers to use hidden trigger words (such as SUDO) in any prompt to induce the language model to generate harmful content.
Unknown Domain Inconsistency Minimization for Domain Generalization
Seungjae Shin (Korea Advanced Institute of Science and Technology), Il-chul Moon
CodeDomain AdaptationOptimizationImage
🎯 What it does: This paper proposes the Unknown Domain Inconsistency Minimization (UDIM) method, which optimizes the consistency of the loss landscape between the source domain and the simulated unknown domain by perturbing both the parameter space and the data space, thereby enhancing domain generalization performance.
🎯 What it does: Proposes the FROND framework, which implements long-term memory characteristics in graph neural networks using Caputo fractional derivatives, improving node feature updates.
Unmasking and Improving Data Credibility: A Study with Datasets for Training Harmless Language Models
Zhaowei Zhu (Docta), Yang Liu (University of California)
CodeClassificationData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Evaluate and clean the label errors in publicly available datasets for training harmless language models, enhancing data credibility and downstream classification performance.
🎯 What it does: This paper proposes an unpaired image translation framework called UNSB based on the Schrödinger Bridge (SB), capable of performing multi-step unsupervised domain transfer at high resolution.
André Cruz, Moritz Hardt (Max Planck Institute for Intelligent Systems)
CodeOptimizationTabularBenchmark
🎯 What it does: This paper evaluates and compares all major algorithms regarding error rate fairness over the past seven years, concluding that post-processing only on the optimal predictor can achieve the accuracy-fairness Pareto frontier attainable by all methods.
UNR-Explainer: Counterfactual Explanations for Unsupervised Node Representation Learning Models
Hyunju Kang (Sungkyunkwan University), Hogun Park (Sungkyunkwan University)
CodeExplainability and InterpretabilityRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: This paper proposes a counterfactual explanation method for unsupervised node representation learning models—UNR-Explainer, which can find subgraphs that significantly change the k-nearest neighbor set of the target node in the embedding space by minimizing perturbation edges.
🎯 What it does: This paper conducts experiments on the learning process of deep networks with different widths on a noisy dataset, studying how noisy data affects the double descent phenomenon, and uses k-NN to evaluate the 'isolation' degree of noisy samples in the learned feature space.
🎯 What it does: An unsupervised sequence learning method (UOL) is proposed, capable of achieving ordered clustering and ranking estimation on unlabeled data.
Unsupervised Pretraining for Fact Verification by Language Model Distillation
Adrián Bazaga (University of Cambridge), Gos Micklem (University of Cambridge)
CodeKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelContrastive LearningTextGraph
🎯 What it does: An unsupervised pre-training framework SFAVEL is proposed, which achieves high-quality alignment features between text claims and knowledge graph facts by distilling the semantic representations of pre-trained language models into a knowledge graph attention network.
Unveiling and Manipulating Prompt Influence in Large Language Models
Zijian Feng (Nanyang Technological University), Kezhi Mao (Nanyang Technological University)
CodeGenerationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A method based on Token Distribution Dynamics (TDD) is proposed to explain and manipulate the influence of prompts on the generation results of large language models.
Unveiling Options with Neural Network Decomposition
Mahdi Alikhasi (University of Alberta), Levi Lelis
CodeReinforcement Learning
🎯 What it does: An algorithm is proposed to decompose a trained neural network policy into sub-policies and package them as options to achieve knowledge transfer across tasks.
Unveiling the Pitfalls of Knowledge Editing for Large Language Models
Zhoubo Li (Zhejiang University), Huajun Chen (Zhejiang University)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper systematically studies the potential side effects that may occur during knowledge editing in large language models, proposing and quantifying two main issues: knowledge conflict (contradictory edits) and knowledge distortion (edits leading to irreversible distortion of knowledge structure). It experimentally verifies the shortcomings of existing editing methods in these two aspects and proposes a multi-label editing (MLE) strategy to mitigate knowledge distortion.
🎯 What it does: Utilizing neural radiance fields (NeRF) to jointly optimize the continuous time motion trajectory of rolling shutter cameras, thereby learning a distortion-free 3D scene representation from rolling shutter images and recovering the camera trajectory;
🎯 What it does: Proposes the VALUES framework to systematically validate the effectiveness and practicality of uncertainty estimation in semantic segmentation.
🎯 What it does: The study investigates the gradient vanishing problem during the reinforcement learning fine-tuning (RFT) process of language models and proposes alleviation through a small amount of supervised fine-tuning (SFT).
Variance Reduced Halpern Iteration for Finite-Sum Monotone Inclusions
Xufeng Cai (University of Wisconsin Madison), Jelena Diakonikolas (University of Wisconsin Madison)
CodeOptimization
🎯 What it does: This paper proposes two variance-reduced algorithms based on Halpern iteration to solve finite-sum monotone inclusion problems, covering applications such as variational inequalities (VI) and convex-concave minimization.
Variance-aware Regret Bounds for Stochastic Contextual Dueling Bandits
Qiwei Di (University of California), Quanquan Gu (University of California)
CodeReinforcement LearningTabular
🎯 What it does: This paper addresses the contextual two-armed bandit problem and proposes a variance-aware algorithm VACDB based on generalized linear models, providing its variance-aware cumulative regret upper bound.
🎯 What it does: This paper proposes a Variational Poisson Learning (VPL) framework for graph-structured semi-supervised learning, particularly suitable for extremely sparse label scenarios, and presents efficient implementations of the V-Laplace and V-Poisson algorithms, as well as the V-GPN model applicable to graph neural networks.
VCR-Graphormer: A Mini-batch Graph Transformer via Virtual Connections
Dongqi Fu (University of Illinois Urbana-Champaign), Bo Long (Meta AI)
CodeGraph Neural NetworkTransformerGraph
🎯 What it does: This paper proposes a graph Transformer that can be used under small-batch training conditions—VCR-Graphormer. It generates a personalized PageRank (PPR) sampling token list for each node and learns node representations through self-attention in the Transformer.
VDC: Versatile Data Cleanser based on Visual-Linguistic Inconsistency by Multimodal Large Language Models
Zihao Zhu (Chinese University of Hong Kong), Baoyuan Wu (Chinese University of Hong Kong)
CodeAnomaly DetectionTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: This paper proposes a general dirty sample detection framework called VDC, which utilizes a multimodal large language model (MLLM) to assess the semantic inconsistency between images and labels, thereby identifying backdoor attack samples, noisy labels, and mixed dirty samples.
VFLAIR: A Research Library and Benchmark for Vertical Federated Learning
Tianyuan Zou (Institute for AI Industry Research Tsinghua University), Ya-Qin Zhang (Institute for AI Industry Research Tsinghua University)
CodeFederated LearningSafty and PrivacyComputational EfficiencyAdversarial AttackImageTextTabularBenchmark
🎯 What it does: This paper proposes and implements a lightweight, scalable vertical federated learning framework called VFLAIR, and conducts a unified evaluation of 11 types of attacks and 8 types of defense methods under this framework, providing benchmarks for communication/computation efficiency, model performance, and defense capabilities.
Vision-Language Models are Zero-Shot Reward Models for Reinforcement Learning
Juan Rocamonde (FAR AI), David Lindner (ETH Zurich)
CodeTransformerReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: A framework utilizing pre-trained visual-language models (such as CLIP) as zero-shot reward models (VLM-RM) is proposed and implemented, which is used to drive reinforcement learning agents to complete various tasks.
🎯 What it does: This paper presents Vocos, a speech synthesis model that directly generates complex STFT coefficients using GANs and achieves upsampling through inverse STFT, completely avoiding traditional transposed convolutions.
🎯 What it does: This paper proposes an unsupervised video object learning framework called VONet, which can simultaneously generate attention masks for multiple objects while maintaining temporal consistency.
🎯 What it does: Learn discrete encoding of local subgraphs of nodes through an improved VQ-VAE structure, and use this codebook as a benchmark to distill the structural knowledge of the teacher GNN into an MLP without adjacency queries, achieving efficient inference;
Weakly Supervised Virus Capsid Detection with Image-Level Annotations in Electron Microscopy Images
Hannah Kniesel (Ulm University), Pedro Hermosilla (Vienna University of Technology)
CodeObject DetectionTransformerImage
🎯 What it does: A weakly supervised virus capsid detection method based solely on image-level labels is proposed, utilizing a pre-trained classifier and gradient optimization of a Gaussian mask to achieve particle localization.
🎯 What it does: A weakly supervised language-conditioned audio separation framework is proposed, which utilizes a pre-trained multimodal audio-text embedding model (CLAP) to provide weak supervision signals for single-source extraction, enabling target source separation without the need for single-source audio data during the training phase.
🎯 What it does: Use text prompt-driven generative models to expand the training set, and improve the retrieval step of visual localization by combining geometric consistency filtering.
What does the Knowledge Neuron Thesis Have to do with Knowledge?
Jingcheng Niu (University of Toronto), Gerald Penn (University of Toronto)
CodeLarge Language ModelText
🎯 What it does: Evaluate the Knowledge Neuron (KN) theory and its model editing methods based on MLP weights (KN Edit, ROME), and expand the evaluation scope to synthetic grammatical phenomena, introducing two new editing evaluation metrics: symmetry and synonymy.
What Makes Good Data for Alignment? A Comprehensive Study of Automatic Data Selection in Instruction Tuning
Wei Liu (ShanghaiTech University), Junxian He (Hong Kong University of Science and Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: A comprehensive study on data selection in instruction tuning is conducted, proposing the EVOL COMPLEXITY/QUALITY and REPR FILTER strategies. Based on this, the DEITA model is trained, achieving alignment performance comparable to traditional datasets of 100K–300K scale with only 6K samples.
🎯 What it does: This paper proposes a Learning-based Proximal Operator Network (LPN), which achieves an exact proximal operator for any non-convex regularization function by parameterizing the network as the gradient of a convex function. It learns the logarithmic prior of the data using proximal matching loss in an unsupervised manner, and subsequently embeds LPN into the Plug-and-Play (PnP) framework to solve inverse problems, providing convergence guarantees.
When should we prefer Decision Transformers for Offline Reinforcement Learning?
Prajjwal Bhargava (Meta AI), Amy Zhang (University of Texas)
CodeTransformerReinforcement LearningTabular
🎯 What it does: A systematic evaluation of three mainstream algorithms in offline reinforcement learning (CQL, BC, DT) was conducted under different conditions of data quality, task complexity, reward sparsity, and randomness, and a guideline for algorithm selection in various scenarios was provided.
🎯 What it does: It is proven that under specific conditions, trained deep neural networks only encode a small number of sparse interaction primitives, thereby explaining the emergence of symbolic concepts.
Wenting Zhao (Cornell University), Yuntian Deng (Allen Institute for Artificial Intelligence)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Collected and publicly released over 1 million real user multi-turn dialogue logs with ChatGPT, forming a multilingual, multi-turn, anonymous dataset with geographical information called WILDCHAT.
🎯 What it does: Identified and fixed a bug that occurred during high-resolution fine-tuning when using window attention and absolute position embeddings, and resolved the issue by introducing the 'absolute win' position embedding scheme.
CodeTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the YaRN method, which improves the interpolation of RoPE positional encoding and temperature scaling to extend the context window of LLMs from thousands to hundreds of thousands without the need for large amounts of data or training steps.
Yet Another ICU Benchmark: A Flexible Multi-Center Framework for Clinical ML
Robin van de Water (Hasso Plattner Institute), Patrick Rockenschaub (Amsterdam UMC)
CodeHyperparameter SearchRecurrent Neural NetworkTransformerTabularBiomedical DataElectronic Health RecordsBenchmark
🎯 What it does: A scalable ICU prediction benchmark framework YAIB has been developed, unifying data extraction, task definition, preprocessing, and model training.
🎯 What it does: A novel membership inference attack called YOQO is proposed, which queries the label only once to determine whether the target sample belongs to the training set by generating query samples in an improved region.
Zero and Few-shot Semantic Parsing with Ambiguous Inputs
Elias Stengel-Eskin (University of North Carolina at Chapel Hill), Benjamin Van Durme (Johns Hopkins University)
CodeTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the AMP framework and dataset, specifically designed to address the ambiguity issues in semantic parsing, and designs zero-shot and few-shot evaluation tasks to assess the performance of large language models when faced with ambiguous sentences.
Zero-Shot Continuous Prompt Transfer: Generalizing Task Semantics Across Language Models
Zijun Wu (University of Alberta), Lili Mou (University of Alberta)
CodeTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A zero-shot continuous prompt transfer method for cross-lingual models is proposed, utilizing encode-then-search relative space encoding and target model search.
Dyah Adila (University of Wisconsin Madison), Frederic Sala (University of Wisconsin Madison)
CodeClassificationLarge Language ModelPrompt EngineeringTextMultimodality
🎯 What it does: A zero-shot robustification method called ROBOSHOT is proposed, which utilizes harmful and beneficial concept embedding vectors generated by a language model to project and amplify the embedding space of a pre-trained model in an unsupervised manner, thereby enhancing the robustness of zero-shot classification.
🎯 What it does: A new ASR encoder called Zipformer is proposed, which improves the U-Net structure, block structure, normalization, activation functions, and optimizers, enhancing speed and performance.
Zoology: Measuring and Improving Recall in Efficient Language Models
Simran Arora (Stanford University), Christopher Re
CodeConvolutional Neural NetworkTransformerLarge Language ModelText
🎯 What it does: This paper focuses on the non-attentive gated convolutional language model, systematically evaluating and improving its associative recall performance on real data, and further explaining the gap between it and Transformers.