ICLR 2024 Papers — Page 3
International Conference on Learning Representations · 2260 papers
BarLeRIa: An Efficient Tuning Framework for Referring Image Segmentation
Yaoming Wang (Shanghai Jiao Tong University), Qi Tian (Huawei Cloud)
SegmentationTransformerVision Language ModelImage
🎯 What it does: A parameter-efficient fine-tuning framework named BarLeRIa is proposed to accomplish reference image segmentation tasks while keeping the pre-trained model frozen.
Batch Calibration: Rethinking Calibration for In-Context Learning and Prompt Engineering
Han Zhou (Google Research), Subhrajit Roy (Google Research)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes the Batch Calibration (BC) method, which estimates contextual bias and performs calibration through batch unlabeled samples, significantly improving the performance of LLMs and VLMs in zero-shot and few-shot prompt learning.
Batch normalization is sufficient for universal function approximation in CNNs
Rebekka Burkholz (CISPA Helmholtz Center for Information Security)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper proves that in randomly initialized convolutional neural networks, training only the parameters of the batch normalization (BN) layers can achieve universal function approximation, and provides a feasible construction method; experiments were also conducted on CIFAR-10 and CIFAR-100 for validation.
Batched Low-Rank Adaptation of Foundation Models
Yeming Wen (University of Texas), Swarat Chaudhuri (University of Texas)
RecognitionComputational EfficiencyAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextAudio
🎯 What it does: This paper proposes FLORA (Fast LORA), a technique that allows for the allocation of independent low-rank adapters for each input example within the same batch, balancing the parameter efficiency of LORA with batch throughput.
BatchPrompt: Accomplish more with less
Jianzhe Lin (Microsoft), Robin Abraham (Microsoft)
Computational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Designed and evaluated a BatchPrompt method that improves the inference efficiency of large language models by inputting multiple data samples in a single prompt. Based on this, batch permutation and voting (BPE) and self-reflection early stopping (SEAS) techniques are proposed to further enhance performance.
BatteryML: An Open-source Platform for Machine Learning on Battery Degradation
Han Zhang (Tsinghua University), Jiang Bian (Microsoft Research)
TransformerTabularTime SeriesBenchmark
🎯 What it does: Built the BatteryML open-source platform, unifying data formats, feature engineering, and multi-model training, supporting battery degradation prediction.
Bayes Conditional Distribution Estimation for Knowledge Distillation Based on Conditional Mutual Information
Linfeng Ye (University of Waterloo), EN-HUI YANG
Knowledge DistillationImage
🎯 What it does: By simultaneously maximizing the log-likelihood and conditional mutual information in the training of the teacher model, the estimation of the Bayesian conditional distribution is improved, thereby enhancing the effect of knowledge distillation.
BayesDiff: Estimating Pixel-wise Uncertainty in Diffusion via Bayesian Inference
Siqi Kou (Shanghai Jiao Tong University), Zhijie Deng (Shanghai Jiao Tong University)
RestorationGenerationDiffusion modelImage
🎯 What it does: This paper presents BayesDiff, a method for pixel-level uncertainty estimation of images generated by diffusion models through Bayesian inference. It utilizes uncertainty metrics to achieve low-quality image filtering, enhance generation diversity, and repair defects.
Bayesian Bi-clustering of Neural Spiking Activity with Latent Structures
Ganchao Wei (Duke University)
Spiking Neural NetworkTime Series
🎯 What it does: A Bayesian nonparametric dual clustering model is proposed to simultaneously cluster the spatial (neuron groups) and temporal (states) aspects of neural spikes, capturing their dynamics through low-dimensional latent trajectories.
Bayesian Coreset Optimization for Personalized Federated Learning
Prateek Chanda (Indian Institute of Technology Bombay), Ganesh Ramakrishnan (Indian Institute of Technology Bombay)
OptimizationFederated LearningImageBiomedical Data
🎯 What it does: This paper proposes CORESET-PFEDBAYES, a framework that combines Bayesian coreset optimization with personalized federated learning, utilizing a representative subset from each client for model updates.
Bayesian Low-rank Adaptation for Large Language Models
Adam X. Yang (University of Bristol), Laurence Aitchison (University of Bristol)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes Laplace-LoRA, a Bayesian fine-tuning method that applies posterior Laplace approximation to LoRA parameters.
Bayesian Neural Controlled Differential Equations for Treatment Effect Estimation
Konstantin Hess (Munich Center for Machine Learning), Stefan Feuerriegel (Munich Center for Machine Learning)
Drug DiscoveryTime SeriesBiomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A Bayesian Neural Controlled Differential Equation (BNCDE) is proposed for estimating individualized treatment effects in continuous time, capable of outputting the posterior predictive distribution of potential outcomes.
Bayesian Optimization through Gaussian Cox Process Models for Spatio-temporal Data
Yongsheng Mei (George Washington University), Tian Lan (Northeastern University)
OptimizationTime Series
🎯 What it does: This paper proposes a Gaussian Cox process MAP inference method based on Laplace approximation and kernel transformation, which can simultaneously estimate the posterior mean and covariance of the latent intensity function, and embed it into a Bayesian optimization framework, supporting various acquisition functions (UCB, idle, cumulative arrival, change point detection).
BayesPrompt: Prompting Large-Scale Pre-Trained Language Models on Few-shot Inference via Debiased Domain Abstraction
Jiangmeng Li (National Key Laboratory of Space Integrated Information System, Institute of Software Chinese Academy of Sciences, Beijing, China), Hui Xiong (Hong Kong University of Science and Technology)
ClassificationRetrievalDomain AdaptationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes BayesPrompt, which utilizes Bayesian distribution approximation and SVGD to debias the factual distribution in downstream domains, and then samples from it to generate prompts with domain discriminative information for few-shot inference in large-scale pre-trained language models.
Be Aware of the Neighborhood Effect: Modeling Selection Bias under Interference
Haoxuan Li (Peking University), Xiangnan He (University of Science and Technology of China)
Recommendation SystemTabular
🎯 What it does: This paper studies the issue of selection bias in recommendation systems when neighborhood effects are present, proposing a neighborhood intervention representation framework based on causal inference and providing the corresponding ideal loss function.
Be Careful What You Smooth For: Label Smoothing Can Be a Privacy Shield but Also a Catalyst for Model Inversion Attacks
Lukas Struppek (Technical University of Darmstadt), Kristian Kersting (Technical University of Darmstadt)
Safty and PrivacyAdversarial AttackImage
🎯 What it does: This study investigates the impact of label smoothing (positive and negative factors) on model inversion attacks (MIA) in deep learning models and proposes negative label smoothing as a defense strategy.
Beam Enumeration: Probabilistic Explainability For Sample Efficient Self-conditioned Molecular Design
Jeff Guo (Institut des Sciences et Ingénierie Chimiques National Centre of Competence in Research Catalysis), Philippe Schwaller (Institut des Sciences et Ingénierie Chimiques National Centre of Competence in Research Catalysis)
Explainability and InterpretabilityDrug DiscoveryLarge Language ModelReinforcement LearningTextBiomedical Data
🎯 What it does: This paper proposes the Beam Enumeration method, which utilizes high-probability token subsequence enumeration generated by a language model to extract molecular substructures and perform self-conditioning generation, thereby significantly improving sample efficiency and providing interpretability.
Beating Price of Anarchy and Gradient Descent without Regret in Potential Games
Iosif Sakos (Singapore University of Technology and Design), Georgios Piliouras (Singapore University of Technology and Design)
Optimization
🎯 What it does: This paper studies the convergence and equilibrium selection problem of q-replicator dynamics (QRD) in potential games, proving that in almost all potential games (PRPG), the QRD point converges to a Nash equilibrium. It also conducts an in-depth analysis of the relative performance of GD and RD in 2×2 symmetric potential games and the difference in their average prices.
BECLR: Batch Enhanced Contrastive Few-Shot Learning
Stylianos Poulakakis-Daktylidis (Delft University of Technology), Hadi Jamali-Rad (Shell Global Solutions International)
ClassificationRepresentation LearningContrastive LearningImage
🎯 What it does: This paper proposes an end-to-end unsupervised few-shot learning framework called BECLR, which combines a dynamic clustering memory module DyCE and a distribution alignment module OpTA, significantly improving the performance of unsupervised pre-training and few-shot inference.
Behaviour Distillation
Andrei Lupu (University of Oxford), Jakob Nicolaus Foerster
OptimizationKnowledge DistillationReinforcement LearningSequential
🎯 What it does: The paper proposes the task of behavior distillation and introduces the HaDES (Hallucinating Datasets with Evolution Strategies) method, which uses synthetic state-action pairs to train RL policies instead of the original dataset.
Belief-Enriched Pessimistic Q-Learning against Adversarial State Perturbations
Xiaolin Sun (Tulane University), Zizhan Zheng (Tulane University)
Adversarial AttackRecurrent Neural NetworkReinforcement LearningDiffusion modelSequential
🎯 What it does: This paper proposes a Bayesian-enhanced pessimistic Q-learning method that uniformly uses worst-case states during both training and testing phases, thereby improving resistance to robust state perturbation attacks.
Bellman Optimal Stepsize Straightening of Flow-Matching Models
Bao Nguyen (Vin University), Viet Anh Nguyen (Chinese University of Hong Kong)
GenerationOptimizationFlow-based ModelImageOrdinary Differential Equation
🎯 What it does: For the pre-trained flow matching generative model, we propose the BOSS (Bellman Optimal Stepsize Straightening) method, which first uses dynamic programming to obtain the optimal sampling step size, and then straightens the velocity network to achieve low NFE high-quality image generation.
Benchmarking Algorithms for Federated Domain Generalization
Ruqi Bai (Purdue University), David I. Inouye (Purdue University)
Domain AdaptationFederated LearningImageTextBenchmark
🎯 What it does: A complete benchmark framework for Federated Domain Generalization is proposed, along with a new controllable heterogeneous data partitioning method, systematically evaluating the performance of 14 methods on 7 multi-domain datasets.
Benchmarking and Improving Generator-Validator Consistency of Language Models
Xiang Lisa Li (Stanford University), Percy Liang (Stanford University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper designs a Generative-Validation Consistency (GV-consistency) metric to systematically evaluate the consistency of language models in both generative and validation queries, and proposes a label-free consistency fine-tuning method to enhance consistency as well as the quality of generation and validation.
BEND: Benchmarking DNA Language Models on Biologically Meaningful Tasks
Frederikke Isa Marin, Wouter Boomsma (University of Copenhagen)
Convolutional Neural NetworkTransformerSupervised Fine-TuningBiomedical DataBenchmark
🎯 What it does: This paper proposes the BEND benchmark, which systematically evaluates the performance of various DNA language models on seven biologically significant tasks.
Benign Oscillation of Stochastic Gradient Descent with Large Learning Rate
Miao Lu (Stanford University), Difan Zou (University of Hong Kong)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: Analyzes the oscillatory behavior of SGD under large learning rates and proves its positive effect on the generalization of neural networks.
Benign Overfitting and Grokking in ReLU Networks for XOR Cluster Data
Zhiwei Xu (University of Michigan), Wei Hu (University of Michigan)
OptimizationRecurrent Neural NetworkTabular
🎯 What it does: This study investigates the training of a two-layer ReLU network on XOR clustering data with noisy labels using gradient descent. It demonstrates that the network overfits all training samples in the first iteration, but the testing performance is close to random. Subsequently, during further training, 'grokking' occurs, ultimately achieving 100% training accuracy and nearly zero testing error, thus exhibiting both harmful overfitting and benign overfitting.
BENO: Boundary-embedded Neural Operators for Elliptic PDEs
Haixin Wang (National Engineering Research Center for Software Engineering Peking University), Tailin Wu (Westlake University)
Graph Neural NetworkTransformerMeshPhysics Related
🎯 What it does: A new neural operator architecture called BENO is proposed to solve elliptic partial differential equations (Poisson/Laplace) with complex geometries and inhomogeneous boundary conditions.
BESA: Pruning Large Language Models with Blockwise Parameter-Efficient Sparsity Allocation
Peng Xu (OpenGVLab), Ping Luo (The University of Hong Kong)
CompressionOptimizationTransformerLarge Language ModelText
🎯 What it does: A block-level differentiable sparse allocation method called BESA is proposed for parameter pruning of large-scale language models, which minimizes reconstruction error layer by layer according to Transformer blocks and learns the sparsity rate of each layer.
Bespoke Solvers for Generative Flow Models
Neta Shaul (Weizmann Institute of Science), Yaron Lipman (Meta)
GenerationData SynthesisComputational EfficiencyDiffusion modelFlow-based ModelImageOrdinary Differential Equation
🎯 What it does: Proposes the 'Bespoke Solver' framework, which customizes a high-order consistency ODE solver for pre-trained generative flow models to significantly reduce the number of sampling steps.
Better Neural PDE Solvers Through Data-Free Mesh Movers
Peiyan Hu (Academy of Mathematics and Systems Science, Chinese Academy of Sciences), Zhi-Ming Ma (Academy of Mathematics and Systems Science, Chinese Academy of Sciences)
Graph Neural NetworkMeshPhysics RelatedOrdinary Differential Equation
🎯 What it does: An unsupervised 'Data-Driven Mesh Mover (DMM)' is proposed to learn adaptive meshes and build a Moving Mesh Neural PDE Solver (MM-PDE) based on it, achieving high-precision simulations of dynamic systems.
Beyond Accuracy: Evaluating Self-Consistency of Code Large Language Models with IdentityChain
Marcus J. Min (Columbia University), Baishakhi Ray (Columbia University)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a new framework called IdentityChain, which simultaneously evaluates the accuracy and self-consistency of code large language models (Code LLM) in natural language to program (NL-to-PL) and program to natural language (PL-to-NL) tasks, and quantifies self-consistency through a new metric TOM (Test Output Match);
Beyond IID weights: sparse and low-rank deep Neural Networks are also Gaussian Processes
Thiziri Nait Saada (University of Oxford), Jared Tanner (University of Oxford)
Tabular
🎯 What it does: This paper proves that deep networks with pseudo-independent and identically distributed (Pseudo-IID) weights (including low-rank, structured sparse, and orthogonal initialization) converge to a Gaussian process when the width is infinite.
Beyond Imitation: Leveraging Fine-grained Quality Signals for Alignment
Geyang Guo (Renmin University of China), Ji-Rong Wen (Renmin University of China)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A method called FIGA is proposed, which aligns large language models directly under conditions without reinforcement learning by utilizing fine-grained comparative signals between initial low-quality responses and their improved versions.
Beyond Memorization: Violating Privacy via Inference with Large Language Models
Robin Staab (ETH Zurich), Martin Vechev (ETH Zurich)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper explores and experiments with the ability of large language models (LLMs) to infer personal attributes (such as location, income, gender, etc.) from unseen text through reasoning, and assesses the privacy threats posed by this capability.
Beyond Reverse KL: Generalizing Direct Preference Optimization with Diverse Divergence Constraints
Chaoqi Wang (University of Chicago), Yuxin Chen (University of Chicago)
Recommendation SystemOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes a generalized direct preference optimization framework f-DPO, which allows for alignment fine-tuning of LLMs under various f-divergence constraints;
Beyond Spatio-Temporal Representations: Evolving Fourier Transform for Temporal Graphs
Anson Bastos (HERE Technologies), Toyotaro Suzumura (University of Tokyo)
Recommendation SystemComputational EfficiencyGraph Neural NetworkTransformerGraphTime Series
🎯 What it does: This paper proposes a reversible spectral transformation called the Evolving Graph Fourier Transform (EFT), which can simultaneously map dynamic graphs to the frequency domain in both the time domain and node domain, capturing the spectral features of the graph structure as it evolves over time.
Beyond Stationarity: Convergence Analysis of Stochastic Softmax Policy Gradient Methods
Sara Klein (University of Mannheim), Leif Döring
OptimizationReinforcement LearningTabular
🎯 What it does: This paper analyzes the global convergence properties of two policy gradient (PG) algorithms in finite-horizon Markov decision processes (MDP): traditional simultaneous training (Simultaneous PG) and backward training (Dynamic Policy Gradient) that incorporates dynamic programming ideas.
Beyond task performance: evaluating and reducing the flaws of large multimodal models with in-context-learning
Mustafa Shukor (Sorbonne University), Matthieu Cord (Valeo)
Explainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Evaluate and mitigate the shortcomings of large-scale multimodal models (LMMs), proposing various untrained in-context learning (ICL) variants to enhance the model's capabilities in interpretability, refusal, compositional reasoning, and instruction following.
Beyond Vanilla Variational Autoencoders: Detecting Posterior Collapse in Conditional and Hierarchical Variational Autoencoders
Hien Dang (FPT Software), Nhat Ho (University of Texas at Austin)
GenerationData SynthesisAuto EncoderImage
🎯 What it does: This paper systematically analyzes the causes and triggering conditions of posterior collapse in Linear Conditional Variational Autoencoders (CVAE) and Markov Hierarchical Variational Autoencoders (MHVAE) through a combination of theoretical and experimental approaches, and proposes corresponding mitigation strategies.
Beyond Weisfeiler-Lehman: A Quantitative Framework for GNN Expressiveness
Bohang Zhang (Peking University), Liwei Wang (Peking University)
Graph Neural NetworkGraph
🎯 What it does: A quantitative framework for the expressive power of graph neural networks based on isomorphic mapping counting is proposed, defining isomorphic expressiveness and providing a complete description of various mainstream GNN models.
Beyond Worst-case Attacks: Robust RL with Adaptive Defense via Non-dominated Policies
Xiangyu Liu (University of Maryland), Furong Huang (J.P. Morgan AI Research)
OptimizationAdversarial AttackRobotic IntelligenceReinforcement LearningAgentic AISequential
🎯 What it does: The study investigates the robustness of reinforcement learning in adversarial attack environments, proposing the construction of a finite non-dominated policy set during the training phase and adaptive defense through online no-regret learning during the testing phase.
Bias Runs Deep: Implicit Reasoning Biases in Persona-Assigned LLMs
Shashank Gupta (Allen Institute for AI), Tushar Khot (Allen Institute for AI)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper systematically evaluates and quantifies the impact of 'character roles' on the reasoning performance of four mainstream LLMs (ChatGPT-3.5, ChatGPT-3.5 (latest version), GPT-4-Turbo, Llama-2-70b-chat) by assigning 19 types of sociodemographic character roles (race, gender, religion, politics, disability) across 24 reasoning datasets covering various fields such as mathematics, programming, medicine, law, and ethics.
Biased Temporal Convolution Graph Network for Time Series Forecasting with Missing Values
Xiaodan Chen (Harbin Institute of Technology), Zhijun Li (Harbin Institute of Technology)
Graph Neural NetworkTime Series
🎯 What it does: The BiTGraph model is proposed for handling multivariate time series forecasting with missing values.
Bidirectional Temporal Diffusion Model for Temporally Consistent Human Animation
Tserendorj Adiya (AI Center, CJ Corporation), Hwasup Lim (Korea Institute of Science and Technology)
GenerationPose EstimationDiffusion modelImageVideo
🎯 What it does: This paper proposes a Bidirectional Temporal Diffusion Model (BTDM) that can generate temporally consistent human animations from a single image, video, or noise.
Bilevel Optimization under Unbounded Smoothness: A New Algorithm and Convergence Analysis
Jie Hao (George Mason University), Mingrui Liu (George Mason University)
OptimizationRepresentation LearningHyperparameter SearchMeta LearningText
🎯 What it does: This paper proposes a new second-order layer optimization algorithm BO-REP, designed to handle non-convex second-order problems with potentially unbounded smoothness in the upper layer function, and theoretically provides an operator complexity of ˜O(1/ε⁴).
BioBridge: Bridging Biomedical Foundation Models via Knowledge Graphs
Zifeng Wang (University of Illinois Urbana-Champaign), RISHITA ANUBHAI
GenerationRetrievalDrug DiscoveryTransformerContrastive LearningMultimodalityBiomedical Data
🎯 What it does: This paper proposes the BioBRIDGE framework, which learns cross-modal conversion modules through knowledge graphs, achieving multi-modal retrieval and generation without fine-tuning the base foundational models (FMs).
Blending Imitation and Reinforcement Learning for Robust Policy Improvement
Xuefeng Liu (University of Chicago), Yuxin Chen (University of Chicago)
Robotic IntelligenceReinforcement LearningAgentic AISequential
🎯 What it does: A Robust Policy Improvement (RPI) algorithm is proposed, which dynamically switches between imitation learning (IL) and reinforcement learning (RL) to achieve robust policy improvement in scenarios with multiple sub-optimal black-box experts.
Bongard-OpenWorld: Few-Shot Reasoning for Free-form Visual Concepts in the Real World
Rujie Wu (Peking University), Yizhou Wang (Peking University)
ClassificationRecognitionMeta LearningTransformerLarge Language ModelVision Language ModelImageBenchmark
🎯 What it does: A Bongard-OpenWorld benchmark is proposed to evaluate few-shot visual reasoning capabilities under real images.
BooookScore: A systematic exploration of book-length summarization in the era of LLMs
Yapei Chang (University of Massachusetts Amherst), Mohit Iyyer (University of Massachusetts Amherst)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper systematically studies the process and evaluation of summarizing long documents (>100K tokens) using large language models.
Boosting Graph Anomaly Detection with Adaptive Message Passing
Jingyan Chen (Nanjing University), Yihua Huang (Nanjing University)
Anomaly DetectionGraph Neural NetworkAuto EncoderContrastive LearningGraph
🎯 What it does: This paper proposes a two-stage unsupervised graph anomaly detection framework called GADAM. It first uses MLP for contrastive learning to obtain local anomaly scores based on local inconsistencies, and then generates global anomaly scores through adaptive message passing with mixed attention and global consistency discrimination, ultimately fusing the scores from both stages.
Boosting of Thoughts: Trial-and-Error Problem Solving with Large Language Models
Sijia Chen (University of Toronto), Di Niu (University of Alberta)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: The Boosting of Thoughts (BoT) framework is proposed, utilizing LLM to generate, aggregate, and evaluate tree-like thinking during the iterative process, and continuously incorporating error analysis and suggestions as experience into the prompts, thereby achieving complex mathematical problem solving without human annotations.
Boosting the Adversarial Robustness of Graph Neural Networks: An OOD Perspective
Kuan Li (Hong Kong University of Science and Technology), Xiang Ao (Institute of Computing Technology, Chinese Academy of Sciences)
Anomaly DetectionAdversarial AttackGraph Neural NetworkGraph
🎯 What it does: This paper proposes a method for adversarial training based on OOD detection called GOOD-AT, aimed at enhancing the robustness of Graph Neural Networks (GNNs) against graph structure attacks.
Boosting Vanilla Lightweight Vision Transformers via Re-parameterization
Zhentao Tan (Alibaba Group), Jieping Ye (Alibaba Group)
ClassificationObject DetectionSegmentationTransformerImage
🎯 What it does: This paper proposes a two-dimensional reparameterized linear module (TDRL), which enhances the model's capacity during the training phase by introducing a multi-branch pyramid structure and batch normalization in the linear layers of a lightweight visual Transformer, while keeping the inference cost unchanged.
Bootstrapping Variational Information Pursuit with Large Language and Vision Models for Interpretable Image Classification
Aditya Chattopadhyay (Johns Hopkins University), Rene Vidal
ClassificationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVision Language ModelImage
🎯 What it does: This paper proposes an interpretable image classification framework called Concept-QA+V-IP, which automatically generates semantic queries using the large language model GPT and trains a lightweight Concept-QA network to utilize pseudo-labels from CLIP and GPT to answer these queries, thereby achieving information tracking and interpretation in V-IP without the need for manual annotation.
Boundary Denoising for Video Activity Localization
Mengmeng Xu (King Abdullah University of Science and Technology), Bernard Ghanem (King Abdullah University of Science and Technology)
Object DetectionSegmentationTransformerVideoMultimodality
🎯 What it does: This paper proposes a boundary denoising-based encoding-decoding framework called DenoiseLoc to address the boundary blur problem in video activity localization.
Bounding Box Stability against Feature Dropout Reflects Detector Generalization across Environments
Yang Yang (Australian National University), Liang Zheng (Australian National University)
Object DetectionAutonomous DrivingImage
🎯 What it does: A detection box stability score (BoS) based on feature map masking is proposed for label-free evaluation of the accuracy of object detection models.
Bounding the Expected Robustness of Graph Neural Networks Subject to Node Feature Attacks
Yassine ABBAHADDOU, Henrik Boström (KTH Stockholm)
Adversarial AttackGraph Neural NetworkGraph
🎯 What it does: This paper studies the robustness of Graph Neural Networks (GNN) against node feature attacks, proposes the concept of expected robustness and provides theoretical upper bounds, designs the GCORN model based on orthogonalized weights to enhance robustness, and presents an attack-independent probability assessment method.
Bounds on Representation-Induced Confounding Bias for Treatment Effect Estimation
Valentyn Melnychuk (LMU Munich), Stefan Feuerriegel (LMU Munich)
Representation LearningFlow-based ModelTabular
🎯 What it does: This paper studies the representation-induced confounding bias (RICB) that may arise when using low-dimensional representation learning to estimate the conditional average treatment effect (CATE) in causal inference. It proposes a neural rebuttal framework that is independent of specific representation learning methods to estimate the upper and lower bounds of RICB, thereby enhancing the reliability of decision-making.
Brain decoding: toward real-time reconstruction of visual perception
Yohann Benchetrit (Meta), Jean-Remi King (Meta)
GenerationRetrievalConvolutional Neural NetworkDiffusion modelContrastive LearningImageTime SeriesMagnetic Resonance Imaging
🎯 What it does: Using MEG signals to train deep models, real-time decoding and generating images of viewers' visual perception;
BrainLM: A foundation model for brain activity recordings
Josue Ortega Caro (Yale University), David van Dijk (Yale University)
TransformerSupervised Fine-TuningBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Developed BrainLM, a self-supervised pre-training foundation model based on Transformer, which learns the whole-brain spatiotemporal dynamics from 6,700 hours of fMRI recordings and can be used for various downstream tasks such as clinical variable prediction, future brain state prediction, and functional network discovery.
BrainSCUBA: Fine-Grained Natural Language Captions of Visual Cortex Selectivity
Andrew Luo, Leila Wehbe (Carnegie Mellon University)
GenerationData SynthesisRetrievalTransformerLarge Language ModelDiffusion modelImageTextMultimodalityMagnetic Resonance Imaging
🎯 What it does: The BrainSCUBA method is proposed, which trains a CLIP-based fMRI encoder to project the weights of each voxel into the CLIP embedding space. It then uses the pre-trained CLIPCap+GPT-2 to generate natural language descriptions that best activate each voxel, and the generated text is used in a text-to-image diffusion model to synthesize images related to that voxel. At the same time, these texts are utilized for a fine-grained exploration of the encoding of 'people' in the brain.
Branch-GAN: Improving Text Generation with (not so) Large Language Models
Fredrik Carlsson (RISE Research Institutes of Sweden), Joakim Nivre (AI Sweden)
GenerationData SynthesisTransformerReinforcement LearningGenerative Adversarial NetworkText
🎯 What it does: Introduces the Branch-GAN framework, which generates branch sequences in parallel on the Transformer and enhances text generation quality through adversarial training with GAN.
Breaking Physical and Linguistic Borders: Multilingual Federated Prompt Tuning for Low-Resource Languages
Wanru Zhao (University of Cambridge), Nicholas Donald Lane
Federated LearningSafty and PrivacyTransformerPrompt EngineeringText
🎯 What it does: A multi-language federated prompt tuning framework is proposed, which efficiently fine-tunes parameters for low-resource languages without sharing data.
Bridging Neural and Symbolic Representations with Transitional Dictionary Learning
Junyan Cheng (Dartmouth), Peter Chin (Dartmouth)
SegmentationRepresentation LearningReinforcement LearningDiffusion modelImage
🎯 What it does: Proposes an unsupervised Transitional Dictionary Learning (TDL) framework that utilizes diffusion decomposition of visual input and online prototype clustering to automatically extract interpretable visual components and their relationships from images, represented in a dictionary form.
Bridging State and History Representations: Understanding Self-Predictive RL
Tianwei Ni (Mila), Pierre-Luc Bacon (Mila)
Representation LearningReinforcement LearningSequential
🎯 What it does: Through a unified analysis, the state and history are abstracted as self-predictive abstractions, proposing a minimized RL algorithm that learns self-predictive representations end-to-end in model-free learning.
Bridging Vision and Language Spaces with Assignment Prediction
Jungin Park (NAVER AI Lab), Kwanghoon Sohn (Korea Institute of Science and Technology)
GenerationRetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Using a single-layer linear mapping, the embedding space of the pre-trained visual model is mapped to the word embedding space of the LLM, and the alignment of visual and language representations is achieved through optimal transport allocation prediction, enabling the frozen LLM to understand visual inputs and generate text.
BroGNet: Momentum-Conserving Graph Neural Stochastic Differential Equation for Learning Brownian Dynamics
Suresh Bishnoi (Indian Institute of Technology Delhi), N M Anoop Krishnan
Graph Neural NetworkGraphPhysics RelatedStochastic Differential Equation
🎯 What it does: The BROGNET framework is proposed, utilizing graph neural networks and stochastic differential equations to learn Brownian dynamics, and achieving momentum conservation through hard constraints.
BRUSLEATTACK: A QUERY-EFFICIENT SCORE- BASED BLACK-BOX SPARSE ADVERSARIAL ATTACK
Quoc Viet Vo, Damith Ranasinghe
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A query-efficient, sparse black-box attack method named BRUSLEATTACK based on a Bayesian framework has been developed to generate sparse adversarial samples under the condition of only obtaining model scores.
BTR: Binary Token Representations for Efficient Retrieval Augmented Language Models
Qingqing Cao (University of Washington), Hannaneh Hajishirzi (University of Washington)
RetrievalCompressionComputational EfficiencyKnowledge DistillationTransformerTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the BTR (Binary Token Representations) technology, which uses a 1-bit vector to precompute token representations for retrieving paragraphs, significantly improving the inference speed and storage efficiency of Retrieval-Augmented Language Models.
Building Cooperative Embodied Agents Modularly with Large Language Models
Hongxin Zhang (Shanghai Jiao Tong University), Chuang Gan (Massachusetts Institute of Technology)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningMultimodality
🎯 What it does: A modular collaborative embodied agent named CoELA has been developed, capable of planning, communicating, and collaborating to complete long-term tasks in decentralized, multi-task, limited observation, and costly communication multi-agent environments.
Butterfly Effects of SGD Noise: Error Amplification in Behavior Cloning and Autoregression
Adam Block (Massachusetts Institute of Technology), Cyril Zhang (Microsoft Research)
Autonomous DrivingReinforcement LearningSequential
🎯 What it does: This study explores the instability of deep neural networks in behavior cloning training, finding that small-batch SGD updates lead to severe fluctuations in long-term rewards, even though the behavior cloning loss is almost unaffected.
Byzantine Robust Cooperative Multi-Agent Reinforcement Learning as a Bayesian Game
Simin Li (Beihang University), Xianglong Liu (Beihang University)
Reinforcement Learning
🎯 What it does: This study investigates Byzantine robust cooperative multi-agent reinforcement learning (c-MARL), constructs BARDec-POMDP through a Bayesian game framework, and proposes the ex interim Markov Perfect Bayesian equilibrium (RMPBE) to achieve robust cooperative strategies against Byzantine faults.
C-TPT: Calibrated Test-Time Prompt Tuning for Vision-Language Models via Text Feature Dispersion
Hee Suk Yoon (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
ClassificationDomain AdaptationTransformerPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: A prompt tuning method called C-TPT is proposed for calibrating CLIP during unlabeled testing;
CABINET: Content Relevance-based Noise Reduction for Table Question Answering
Sohan Patnaik (Adobe), Balaji Krishnamurthy (Adobe)
TransformerLarge Language ModelTabular
🎯 What it does: Proposes the CABINET framework, which uses an unsupervised relevance scorer and a weakly supervised parsing statement generator to weight table content, helping large language models focus more on question-relevant table information to complete table question-answering tasks.
CADS: Unleashing the Diversity of Diffusion Models through Condition-Annealed Sampling
Seyedmorteza Sadat (ETH Zurich), Romann M. Weber (Disney Research Studios)
GenerationData SynthesisPose EstimationDiffusion modelImage
🎯 What it does: The Condition-Annealed Diffusion Sampler (CADS) introduces noise to the conditional vector during the inference phase and gradually removes it, significantly enhancing the output diversity of conditional diffusion models while maintaining image quality.
CALICO: Self-Supervised Camera-LiDAR Contrastive Pre-training for BEV Perception
Jiachen Sun (University of Michigan), Chaowei Xiao (University of Wisconsin)
Object DetectionSegmentationAutonomous DrivingKnowledge DistillationContrastive LearningMultimodalityPoint Cloud
🎯 What it does: The CALICO framework is proposed for self-supervised dual-modal BEV pre-training of LiDAR and cameras, which includes two stages: point-region contrast (PRC) and region-aware distillation (RAD);
CAMBranch: Contrastive Learning with Augmented MILPs for Branching
Jiacheng Lin (University of Illinois Urbana-Champaign), Huangang Wang (Tsinghua University)
OptimizationGraph Neural NetworkContrastive LearningGraph
🎯 What it does: The CAMBranch framework is proposed, which generates Augmented MILP (AMILP) by translating the original MILP variables, using these samples for imitation learning and combining contrastive learning to enhance the branching strategy.
Cameras as Rays: Pose Estimation via Ray Diffusion
Jason Y. Zhang (Carnegie Mellon University), Shubham Tulsiani (Carnegie Mellon University)
Pose EstimationTransformerDiffusion modelImage
🎯 What it does: A distributed representation of camera parameters is proposed, parameterized as a bundle of light rays, and based on this, an end-to-end regression network and diffusion model are designed to achieve sparse view camera pose estimation.
CAMIL: Context-Aware Multiple Instance Learning for Cancer Detection and Subtyping in Whole Slide Images
Olga Fourkioti (Institute of Cancer Research), Chris Bakal (Institute of Cancer Research)
ClassificationSegmentationTransformerContrastive LearningImage
🎯 What it does: This paper proposes a context-aware multi-instance learning framework CAMIL that combines neighborhood-constrained attention with Nystromformer for cancer detection and subtype classification in whole slide images (WSI).
Can Large Language Models Infer Causation from Correlation?
Zhijing Jin (ETH Zürich), Bernhard Schölkopf (Max Planck Institute for Intelligent Systems)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposes the CORR2CAUSE task and dataset, exploring whether large language models can infer causal relationships from correlations;
Can LLM-Generated Misinformation Be Detected?
Canyu Chen (Illinois Institute of Technology), Kai Shu (Illinois Institute of Technology)
ClassificationGenerationTransformerLarge Language ModelText
🎯 What it does: This study investigates the difficulty of detecting misinformation generated by LLMs (such as ChatGPT) for both humans and machine detectors, constructs a five-dimensional classification framework for LLM misinformation, and validates three types of generation methods (Hallucination, Arbitrary, Controllable), generating the LLMFake dataset for empirical evaluation.
Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs
Miao Xiong (National University of Singapore), Bryan Hooi (National University of Singapore)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This study investigates the self-confidence expression methods of black-box LLMs, proposing a three-component framework (prompt, sampling, aggregation) and conducting systematic evaluations across multiple models and tasks.
Can LLMs Keep a Secret? Testing Privacy Implications of Language Models via Contextual Integrity Theory
Niloofar Mireshghallah (University of Washington), Yejin Choi (Allen Institute for Artificial Intelligence)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: A four-level privacy reasoning benchmark called CONFAIDE, based on the theory of contextual integrity, is proposed to assess the privacy leakage risks of LLMs in multi-source interactive environments.
Can Sensitive Information Be Deleted From LLMs? Objectives for Defending Against Extraction Attacks
Vaidehi Patil (University of North Carolina Chapel Hill), Mohit Bansal (University of North Carolina Chapel Hill)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: In pre-trained language models, an attack and defense framework is proposed for the direct deletion and extraction defense of sensitive information, exploring 'complete deletion' through model editing and verifying the deletion effect.
Can Transformers Capture Spatial Relations between Objects?
Chuan Wen (Institute for Interdisciplinary Information Sciences Tsinghua University), Yang Gao (University of Pennsylvania)
RecognitionObject DetectionTransformerVision Language ModelImage
🎯 What it does: For the task of precise, physics-based spatial relationship prediction (SRP), we redefine the semantics of spatial relationships, relabel the SpatialSense dataset as SpatialSense+, and propose an end-to-end RelatiViT Transformer architecture.
Can We Evaluate Domain Adaptation Models Without Target-Domain Labels?
Jianfei Yang (Nanyang Technological University), Lihua Xie (Nanyang Technological University)
Domain AdaptationImage
🎯 What it does: A target domain label-free UDA model evaluation metric called Transfer Score is proposed for model selection, hyperparameter tuning, and checkpoint selection.
Can we get the best of both Binary Neural Networks and Spiking Neural Networks for Efficient Computer Vision?
Gourav Datta (University of Southern California), Peter Anthony Beerel
ClassificationObject DetectionComputational EfficiencyConvolutional Neural NetworkSpiking Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A sparse binary activation neural network (BANN) is designed using Hoyer regularization and threshold layers, balancing the high accuracy of binary neural networks with the sparsity of spiking neural networks to achieve low-power visual tasks.
Candidate Label Set Pruning: A Data-centric Perspective for Deep Partial-label Learning
Shuo He (University of Electronic Science and Technology of China), Lei Feng (Nanyang Technological University)
ClassificationData-Centric LearningContrastive LearningImage
🎯 What it does: A candidate label set pruning (CLSP) method is proposed, which removes potential pseudo-labels through 'down voting' based on the inconsistency between the k-NN representation space of instances and the candidate label space, thereby reducing bias without training the model.
CARD: Channel Aligned Robust Blend Transformer for Time Series Forecasting
Xue Wang (Alibaba Group), Rong Jin (Meta)
Anomaly DetectionOptimizationTransformerTime Series
🎯 What it does: This work proposes the Channel Aligned Robust Blend Transformer (CARD), a Transformer model specifically designed for multivariate time series forecasting, which enhances prediction performance through channel and hidden dimension aligned attention, a Token Blend module, and a robust signal attenuation loss.
CAS: A Probability-Based Approach for Universal Condition Alignment Score
Chunsan Hong (KAIST), Tae-Hyun Oh (POSTECH)
GenerationData SynthesisDiffusion modelImageMultimodalityOrdinary Differential EquationAudio
🎯 What it does: A Conditional Alignment Score (CAS) is proposed, which utilizes the conditional probabilities of the diffusion model itself to evaluate the consistency of generated samples with given conditions, achieving a self-rejection mechanism without the need for additional training.
Cascading Reinforcement Learning
Yihan Du (University of Illinois Urbana-Champaign), Wei Chen (Microsoft)
Recommendation SystemReinforcement LearningTabular
🎯 What it does: A new framework called 'Cascading Reinforcement Learning (Cascading RL)' is proposed to consider user states and their transitions in scenarios such as recommendation systems and online advertising, with the goal of maximizing long-term cumulative rewards.
Catastrophic Jailbreak of Open-source LLMs via Exploiting Generation
Yangsibo Huang (Princeton University), Danqi Chen (Princeton University)
GenerationAdversarial AttackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes a method to bypass the alignment of open-source LLMs through changes in decoding configurations (removing system prompts, adjusting temperature/TopK/TopP, etc.) and, based on this, introduces a multi-decoding strategy for a generation-aware alignment method for defense.
Cauchy-Schwarz Divergence Information Bottleneck for Regression
Shujian Yu (Vrije Universiteit Amsterdam), Jose C Principe
CompressionOptimizationRecurrent Neural NetworkImageTabular
🎯 What it does: This study investigates the deep information bottleneck for regression tasks and proposes using Cauchy-Schwarz divergence as a substitute for KL/MSE for prediction and compression.
Causal Fairness under Unobserved Confounding: A Neural Sensitivity Framework
Maresa Schröder, Stefan Feuerriegel (Munich Center for Machine Learning)
Supervised Fine-TuningReinforcement LearningTabular
🎯 What it does: This study investigates the sensitivity of causal fairness in the presence of unobserved confounding variables and proposes a neural network framework that provides upper and lower bounds for worst-case causal fairness under a given sensitivity model, thereby training a predictive model that is robust to confounding uncertainty.
Causal Inference with Conditional Front-Door Adjustment and Identifiable Variational Autoencoder
Ziqi Xu (University of South Australia), Kui Yu (Hefei University of Technology)
Auto EncoderTabular
🎯 What it does: Proposes a Conditional Front-Door (CFD) framework and learns potential CFD adjustment variables based on an identifiable Variational Autoencoder (CFDiVAE) to estimate the average causal effect in the presence of unobserved confounding variables from observational data.
Causal Modelling Agents: Causal Graph Discovery through Synergising Metadata- and Data-driven Reasoning
Ahmed Abdulaal (University College London), Daniel C. Alexander (University College London)
TransformerLarge Language ModelMultimodalityBiomedical DataMagnetic Resonance ImagingAlzheimer's DiseaseBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the Causal Modelling Agent (CMA), a framework that combines metadata inference generated by large language models (LLMs) with deep structural causal models (DSCM) for causal graph discovery on multimodal data, and conducts a case study on Alzheimer's disease (AD) data.
Causal Structure Recovery with Latent Variables under Milder Distributional and Graphical Assumptions
Xiu-Chuan Li (Sydney AI Centre, University of Sydney), Tongliang Liu (Sydney AI Centre, University of Sydney)
Graph
🎯 What it does: A method for fully recovering the causal structure of linear latent variable models is proposed without the assumptions of non-Gaussianity, purity, and two purity.
Causal-StoNet: Causal Inference for High-Dimensional Complex Data
Yaxin Fang (Purdue University), Faming Liang (Purdue University)
Tabular
🎯 What it does: A deep learning model named Causal-StoNet is proposed, which jointly estimates the average causal effect, propensity score, and outcome function, and can handle high-dimensional nonlinear data and missing values.
Causality-Inspired Spatial-Temporal Explanations for Dynamic Graph Neural Networks
Kesen Zhao (City University of Hong Kong), Liang Zhang (Shenzhen Research Institute of Big Data)
Explainability and InterpretabilityGraph Neural NetworkAuto EncoderContrastive LearningGraph
🎯 What it does: This paper proposes a causal heuristic spatiotemporal interpreter for dynamic graph neural networks, DyGNNExplainer, which can generate subgraphs that conform to causal relationships without compromising the model.