ICLR 2024 Papers — Page 21
International Conference on Learning Representations · 2260 papers
The Hidden Language of Diffusion Models
Hila Chefer (Google Research), Lior Wolf (Tel-Aviv University)
GenerationData SynthesisExplainability and InterpretabilityDiffusion modelImageText
🎯 What it does: The CONCEPTOR method is proposed to decompose the internal representations of text prompts in text-to-image diffusion models into a set of interpretable text elements.
The Human-AI Substitution game: active learning from a strategic labeler
Tom Yan (Carnegie Mellon University), Chicheng Zhang (University of Arizona)
🎯 What it does: The paper studies how to design active learning algorithms to quickly identify the target model when labelers have self-interested motives (hoping to obtain rewards by providing as many labels as possible); it proposes a 'human-machine substitution game' framework and provides a new state representation based on Effective Version Space (E-VS) and a near-optimal query complexity algorithm.
The importance of feature preprocessing for differentially private linear optimization
Ziteng Sun (Google Research), Aditya Krishna Menon (Google Research)
OptimizationSafty and PrivacyImage
🎯 What it does: The study proves that simply using DPSGD in differential privacy linear classification cannot achieve optimal solutions. Feature preprocessing is required first, followed by the proposal and implementation of the improved DPSGD-F algorithm.
The Joint Effect of Task Similarity and Overparameterization on Catastrophic Forgetting — An Analytical Model
Daniel Goldfarb (Northeastern University), PAul HAnd
TabularSequential
🎯 What it does: This paper studies the mechanism by which task similarity and over-parameterization jointly affect catastrophic forgetting, and provides an exact expected forgetting expression for two-task linear regression.
The Lipschitz-Variance-Margin Tradeoff for Enhanced Randomized Smoothing
Blaise Delattre (Foxstream), Alexandre Allauzen (Paris-Dauphine University)
ClassificationOptimizationAdversarial AttackTransformerGaussian SplattingImage
🎯 What it does: This study investigates the interaction between the Lipschitz constant of base classifiers, the prediction margin, and the Monte-Carlo sampling variance in Randomized Smoothing. A new method called LVM-RS is proposed, which introduces r-simplex mapping and the Empirical Bernstein inequality to balance variance and margin, thereby enhancing the certification radius and accuracy.
The LLM Surgeon
Tycho F. A. van der Ouderaa, Tijmen Blankevoort (University of Amsterdam)
CompressionTransformerLarge Language ModelText
🎯 What it does: The LLM Surgeon framework is proposed, which can perform unstructured, semi-structured, and structured pruning on large language models, supporting multi-step pruning and low-rank one-shot gradient correction.
The Marginal Value of Momentum for Small Learning Rate SGD
Runzhe Wang (Princeton University), Zhiyuan Li (Toyota Technological Institute at Chicago)
OptimizationConvolutional Neural NetworkTransformerImageTextStochastic Differential Equation
🎯 What it does: This paper studies the practical effects of Stochastic Gradient Descent with Momentum (SGDM) on optimization and generalization under a small learning rate, providing systematic theoretical and experimental validation.
The mechanistic basis of data dependence and abrupt learning in an in-context classification task
Gautam Reddy (Princeton University)
ClassificationOptimizationTransformerTabular
🎯 What it does: The study investigates the transition from weight learning (IWL) to context learning (ICL) in Transformer models and reveals its dependence on data distribution and mechanisms.
The Need for Speed: Pruning Transformers with One Recipe
Samir Khaki (University of Toronto), Konstantinos N Plataniotis
CompressionKnowledge DistillationTransformerImageTextMultimodality
🎯 What it does: One-shot pruning of pre-trained Transformers without the need for retraining, using trajectory importance measurement.
The optimality of kernel classifiers in Sobolev space
Jianfa Lai (Tsinghua University), Qian Lin (Tsinghua University)
ClassificationImage
🎯 What it does: This study investigates the generalization error of nuclear classifiers in Sobolev spaces and provides upper and lower bounds to prove that their convergence rate is optimal.
The Reasonableness Behind Unreasonable Translation Capability of Large Language Model
Tingchen Fu (Renmin University of China), Rui Yan (Renmin University of China)
TransformerLarge Language ModelText
🎯 What it does: This paper explores the reasons why multilingual large language models can acquire translation capabilities without parallel corpora, systematically analyzing the impact of different granularities of non-parallel bilingual data, such as sentence-level and word-level bilingual contamination and code-switching, on translation performance.
The Reversal Curse: LLMs trained on “A is B” fail to learn “B is A”
Lukas Berglund (Vanderbilt University), Owain Evans (University of Oxford)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Through fine-tuning and empirical evaluation of LLMs, it was found that they cannot automatically infer 'B is A' after learning 'A is B', leading to the phenomenon termed 'reverse curse'.
The Trickle-down Impact of Reward Inconsistency on RLHF
Lingfeng Shen (Johns Hopkins University), Dong Yu (Tencent AI Lab)
Reinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: The paper proposes a new evaluation method—Contrast Instructions—to measure the reward consistency of the Reward Model under different but similar instructions, and based on this, two techniques to enhance consistency are introduced: CONVEXDA during training and REWARDFUSION during inference.
The Truth is in There: Improving Reasoning in Language Models with Layer-Selective Rank Reduction
Pratyusha Sharma (Massachusetts Institute of Technology), Dipendra Misra (Microsoft Research)
TransformerLarge Language ModelText
🎯 What it does: Perform layer-selective rank reduction (LASER) on Transformer LLMs by retaining low-order SVD components in specific layers and removing high-order components to improve inference accuracy;
The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning
Bill Yuchen Lin (Allen Institute for Artificial Intelligence), Yejin Choi (University of Washington)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: This study analyzes the impact of alignment fine-tuning on foundational large language models (LLMs), revealing that alignment primarily alters the stylized position of token distribution. It proposes a no-fine-tuning alignment method called URIAL, which achieves alignment effects comparable to or even better than SFT/RLHF through in-context learning using a small number of fixed examples and system prompts.
The Unreasonable Effectiveness of Linear Prediction as a Perceptual Metric
Daniel Severo (University of Toronto), Johannes Ballé (Google)
RetrievalCompressionImage
🎯 What it does: A linear autoregressive similarity index (LASI) is proposed, which constructs perceptual embeddings and calculates image similarity by solving a weighted least squares problem during inference without the need for training or pre-trained networks.
The Update-Equivalence Framework for Decision-Time Planning
Samuel Sokota (Carnegie Mellon University), Noam Brown (OpenAI)
OptimizationReinforcement Learning
🎯 What it does: This paper studies a decision time planning framework that does not rely on public information states and proposes a DTP algorithm based on update equivalence (Mirror Descent Search, MDS, and Magnetic Mirror Descent Search, MMDS), achieving more efficient search in Hanabi and zero-sum games.
The Wasserstein Believer: Learning Belief Updates for Partially Observable Environments through Reliable Latent Space Models
Raphaël Avalos (Vrije Universitait Brussel), Diederik M Roijers
OptimizationRecurrent Neural NetworkReinforcement LearningFlow-based ModelAuto EncoderSequential
🎯 What it does: This paper proposes an algorithm called Wasserstein Belief Updater (WBU), which can construct reliable Bayesian state estimates in partially observable environments by learning latent models and performing Bayesian updates, thus achieving effective compression of history without using RNNs.
Theoretical Analysis of Robust Overfitting for Wide DNNs: An NTK Approach
Shaopeng Fu (King Abdullah University of Science and Technology), Di Wang (King Abdullah University of Science and Technology)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper presents a theoretical analysis of robust overfitting in wide DNNs during adversarial training using the NTK method. It proves that wide networks can be approximated linearly, provides a closed-form dynamic for adversarial training, and reveals the phenomenon of 'AT degradation,' where long-term adversarial training leads the network to degrade to a state without adversarial training. Based on this, the first infinite-width adversarial training algorithm, Adv-NTK, is designed.
Theoretical Understanding of Learning from Adversarial Perturbations
Soichiro Kumano (University of Tokyo), Toshihiko Yamasaki (University of Tokyo)
Adversarial AttackImage
🎯 What it does: The theoretical foundation of learning from adversarial perturbations is studied, proving that the perturbations contain class features and can achieve generalization.
Thin-Shell Object Manipulations With Differentiable Physics Simulations
Yian Wang (University of Massachusetts Amherst), Chuang Gan (Massachusetts Institute of Technology)
OptimizationRobotic IntelligenceReinforcement LearningAgentic AIBenchmarkPhysics Related
🎯 What it does: ThinShellLab is proposed, a benchmark for thin shell material manipulation based on a differentiable physics simulation platform, which includes various thin shell objects and tasks, supporting robot learning and inverse design.
Think before you speak: Training Language Models With Pause Tokens
Sachin Goyal (Carnegie Mellon University), Vaishnavh Nagarajan (Google Research)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Introduce learnable <pause> tokens in the Transformer language model to delay computation before generating the next token.
Think-on-Graph: Deep and Responsible Reasoning of Large Language Model on Knowledge Graph
Jiashuo Sun (IDEA Research), Jian Guo
Graph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraph
🎯 What it does: A new LLM⊗KG coupling paradigm is proposed—Think-On-Graph (ToG), which allows large language models to perform adaptive beam search on knowledge graphs and make reasoning decisions;
THOUGHT PROPAGATION: AN ANALOGICAL APPROACH TO COMPLEX REASONING WITH LARGE LANGUAGE MODELS
Junchi Yu (Institute of Automation Chinese Academy of Sciences), Zhitao Ying
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: A Thought Propagation framework is proposed, utilizing solutions to similar problems to improve multi-step reasoning in large language models.
Threaten Spiking Neural Networks through Combining Rate and Temporal Information
Zecheng Hao (Peking University), Tiejun Huang (Peking University)
Adversarial AttackSpiking Neural NetworkImage
🎯 What it does: This study investigates the rate information and temporal information of SNNs, proposing a hybrid attack method called HART that efficiently attacks SNNs in various attack scenarios.
Threshold-Consistent Margin Loss for Open-World Deep Metric Learning
Qin ZHANG, Yifan Xing (Alexa AI)
RetrievalConvolutional Neural NetworkTransformerImage
🎯 What it does: This study investigates the issue of threshold inconsistency in deep metric learning, proposing a variance-based OPIS metric and designing a Threshold-Consistent Margin (TCM) loss to enhance threshold consistency and retrieval accuracy.
TiC-CLIP: Continual Training of CLIP Models
Saurabh Garg (Carnegie Mellon University), Fartash Faghri (Apple)
ClassificationRetrievalTransformerContrastive LearningMultimodalityTime SeriesBenchmark
🎯 What it does: A time-continuous CLIP training benchmark (TIC-DataComp, TIC-YFCC, TIC-RedCaps) is proposed, and dynamic retrieval and classification evaluation tasks are designed on it.
Tight Rates in Supervised Outlier Transfer Learning
Mohammadreza Mousavi Kalan, Samory Kpotufe (Columbia University)
Domain AdaptationAnomaly Detection
🎯 What it does: This paper studies how to transfer known rare category (anomaly) distribution information to the target task under limited samples, using the Neyman-Pearson framework for theoretical analysis, providing information-theoretic lower bounds and optimal convergence rates.
Time Fairness in Online Knapsack Problems
Adam Lechowicz (University of Massachusetts Amherst), Mohammad Hajiesmaili (University of Massachusetts Amherst)
OptimizationTabular
🎯 What it does: This paper studies the time fairness of the online knapsack problem, introduces α-conditional time-independent fairness, designs threshold-based deterministic and learning-enhanced algorithms, and conducts experimental evaluations.
Time Travel in LLMs: Tracing Data Contamination in Large Language Models
Shahriar Golchin (University of Arizona), Mihai Surdeanu (University of Arizona)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Using guided instruction to have large language models complete partial instances and assess data contamination in the training or test set through matching evaluation.
Time-Efficient Reinforcement Learning with Stochastic Stateful Policies
Firas Al-Hafez (TU Darmstadt), Davide Tateo (TU Darmstadt)
Reinforcement LearningSequentialOrdinary Differential Equation
🎯 What it does: A stateful policy gradient estimation method called S2PG is proposed, which does not use BPTT. It utilizes a stochastic internal state transition kernel to decompose the stateful policy into a stochastic kernel and a non-stateful policy, optimizing them jointly to address the issues of gradient vanishing, explosion, and sequence computation bottlenecks associated with traditional BPTT.
Time-LLM: Time Series Forecasting by Reprogramming Large Language Models
Ming Jin (Monash University), Qingsong Wen (Alibaba Group)
TransformerLarge Language ModelPrompt EngineeringTime Series
🎯 What it does: The TIME-LLM framework is proposed, which reprograms time series data into text prototypes and adds Prompt-as-Prefix in front of the LLM, enabling the frozen LLM to perform general time series prediction tasks.
Time-Varying Propensity Score to Bridge the Gap between the Past and Present
Rasool Fakoor (Amazon Web Services), Alex Smola
Domain AdaptationReinforcement LearningTabularTime Series
🎯 What it does: A time-varying propensity score is proposed to capture the gradual distribution drift of data over time, assigning different weights to historical samples during training;
TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting
Shiyu Wang (Ant Group), JUN ZHOU
Time Series
🎯 What it does: This paper proposes a multi-scale hybrid framework called TimeMixer based on a fully MLP architecture to address the complex non-stationarity issues in time series forecasting.
To Grok or not to Grok: Disentangling Generalization and Memorization on Corrupted Algorithmic Datasets
Darshil Doshi (University of Maryland), Andrey Gromov (Meta AI)
Representation LearningTransformerSupervised Fine-TuningTabular
🎯 What it does: This study investigates the relationship between generalization and memory in deep learning, particularly on algorithm datasets with corrupted labels, using interpretable models to analyze the differences between generalization representations and memory representations.
To the Cutoff... and Beyond? A Longitudinal Perspective on LLM Data Contamination
Manley Roberts (Abacus AI), Samuel Dooley (Abacus AI)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: By utilizing a natural experiment on the training cutoff time of LLMs, longitudinal data contamination detection was conducted on the two major coding/mathematics problem repositories, Codeforces and Project Euler;
TokenFlow: Consistent Diffusion Features for Consistent Video Editing
Michal Geyer, Tali Dekel
GenerationData SynthesisDiffusion modelVideoText
🎯 What it does: Using a pre-trained text-to-image diffusion model, we perform text-driven editing on input videos and achieve high-quality, temporally coherent results by maintaining inter-frame consistency in the diffusion feature space.
Tool-Augmented Reward Modeling
Lei Li (Zhejiang University), Hua Wu (Baidu Inc.)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a tool-augmented reward model framework called Themis, which enables the reward model to dynamically call external tools (such as calculators, search engines, etc.) during the reasoning process to obtain information and evaluate the quality of outputs.
ToolChain*: Efficient Action Space Navigation in Large Language Models with A* Search
Yuchen Zhuang (Georgia Institute of Technology), Chao Zhang (Georgia Institute of Technology)
TransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: A tool usage planning algorithm called ToolChain is proposed, which is based on A* tree search and utilizes the action space generated by LLM for efficient navigation, addressing planning and reasoning problems under a large-scale tool set.
ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs
Yujia Qin (Tsinghua University), Maosong Sun (Tsinghua University)
TransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Constructed the ToolBench dataset and trained ToolLLaMA on it to enhance the tool usage capabilities of open-source LLMs.
Topic Modeling as Multi-Objective Contrastive Optimization
Thong Thanh Nguyen (Institute of Data Science National University of Singapore), Anh Tuan Luu (Nanyang Technological University)
OptimizationAuto EncoderContrastive LearningText
🎯 What it does: A neural topic model framework that combines set-based contrastive learning with multi-objective optimization is proposed to enhance topic coherence, topic diversity, and downstream task performance.
Topological data analysis on noisy quantum computers
Ismail Yunus Akhalwaya (IBM Research), Lior Horesh
Point CloudMeshPhysics Related
🎯 What it does: A topology data analysis algorithm NISQ-TDA is proposed, which can be executed on NISQ devices to estimate Betti numbers.
TopoMLP: A Simple yet Strong Pipeline for Driving Topology Reasoning
Dongming Wu (Beijing Institute of Technology), Jianbing Shen
Object DetectionAutonomous DrivingPoint Cloud
🎯 What it does: A complete process is proposed that combines query-based 3D lane detection with 2D traffic element detection, using a simple MLP to achieve topological reasoning between lanes and traffic elements.
ToRA: A Tool-Integrated Reasoning Agent for Mathematical Problem Solving
Zhibin Gou (Tsinghua University), Weizhu Chen (Microsoft Research)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Designed and trained a series of Tool Integrated Reasoning Agents (TORA) to solve complex mathematical problems through interaction with external computational/symbolic tools using natural language reasoning.
TorchRL: A data-driven decision-making library for PyTorch
Albert Bou (Acellera), Vincent Moens (Meta)
Robotic IntelligenceReinforcement Learning
🎯 What it does: This paper presents TorchRL, a general decision-making and control library based on PyTorch, featuring a composable and modular component design.
TOSS: High-quality Text-guided Novel View Synthesis from a Single Image
Yukai Shi, Heung-Yeung Shum
GenerationData SynthesisPose EstimationTransformerDiffusion modelImageText
🎯 What it does: A text-guided single-image novel view synthesis model TOSS is proposed, significantly improving the rationality and controllability of the generated views.
Toward effective protection against diffusion-based mimicry through score distillation
Haotian Xue (Georgia Institute of Technology), Yongxin Chen (Georgia Institute of Technology)
Computational EfficiencyAdversarial AttackDiffusion modelScore-based ModelImage
🎯 What it does: This study investigates how to protect images from imitation attacks based on diffusion models and proposes a method for generating adversarial perturbations against potential threats.
Toward Optimal Policy Population Growth in Two-Player Zero-Sum Games
Stephen Marcus McAleer (Carnegie Mellon University), Roy Fox (University of California)
OptimizationReinforcement LearningTabular
🎯 What it does: An algorithm is proposed that can expand the strategy set in two-player zero-sum games without increasing exploitability: Anytime Double Oracle (ADO), its approximate extension Anytime PSRO (APSRO), and the improved Self-Play PSRO (SP-PSRO), which are validated in various games.
Toward Student-oriented Teacher Network Training for Knowledge Distillation
Chengyu Dong (University of California), Jingbo Shang (University of California)
Knowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: A student performance-oriented teacher network training method called SoTeacher is proposed, which utilizes Lipschitz regularization and consistency regularization to enhance the teacher's approximation of the true label distribution, thereby significantly improving the student's distillation effect.
Towards 3D Molecule-Text Interpretation in Language Models
Sihang Li (University of Science and Technology of China), Qi Tian (Huawei Cloud)
GenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextMultimodality
🎯 What it does: This paper proposes 3D-MoLM, which integrates a 3D molecular encoder with a language model to achieve the parsing and generation of 3D molecules into text.
Towards a statistical theory of data selection under weak supervision
Germain Kolossov (Granica Computing Inc), Pulkit Tandon (Granica Computing Inc)
Autonomous DrivingData-Centric LearningImage
🎯 What it does: The study focuses on subsample selection using unlabeled samples and proxy models under weak supervision conditions, thereby reducing labeling costs and improving learning outcomes.
Towards Aligned Layout Generation via Diffusion Model with Aesthetic Constraints
Jian Chen (University at Buffalo), Changyou Chen (University at Buffalo)
GenerationDiffusion modelImage
🎯 What it does: A unified continuous space diffusion model LACE is proposed for controllable layout generation, capable of handling various tasks such as unconditional generation, category/size conditional generation, partial layout completion, and refinement.
Towards Assessing and Benchmarking Risk-Return Tradeoff of Off-Policy Evaluation
Haruka Kiyohara (Cornell University), Yuta Saito (Cornell University)
Recommendation SystemReinforcement LearningTabularBenchmarkFinance Related
🎯 What it does: This paper proposes a new metric, SharpeRatio@k, to evaluate the risk-return trade-off of OPE estimators in online A/B testing.
Towards Best Practices of Activation Patching in Language Models: Metrics and Methods
Fred Zhang (University of California Berkeley), Neel Nanda
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: The system evaluates three major hyperparameters of activation patching in language models: corruption methods (Gaussian noise vs. symmetric word replacement), evaluation metrics (probability, logit difference, KL divergence), and sliding window patches, and analyzes their impact on interpretability results.
Towards Category Unification of 3D Single Object Tracking on Point Clouds
Jiahao Nie (Hangzhou Dianzi University), Fei Xie (Shanghai Jiao Tong University)
Object TrackingAutonomous DrivingTransformerPoint Cloud
🎯 What it does: Designed and implemented two unified 3D single-object tracking models, SiamCUT and MoCUT, capable of simultaneously tracking different categories of objects on the same network.
Towards Characterizing Domain Counterfactuals for Invertible Latent Causal Models
Zeyu Zhou (Purdue University), David I. Inouye (Purdue University)
Data SynthesisDomain AdaptationFlow-based ModelAuto EncoderImage
🎯 What it does: The researchers propose to directly estimate domain adversarial counterfactuals using domain labels within the framework of Invertible Latent Causal Models (ILD), avoiding the need for complete causal structure recovery.
Towards Cheaper Inference in Deep Networks with Lower Bit-Width Accumulators
Yaniv Blumenfeld (Technion), Daniel Soudry (Technion)
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformerSupervised Fine-TuningImageText
🎯 What it does: This study investigates how to use low-bit-width accumulators (12 bits and below) during the inference phase of deep networks to achieve acceptable accuracy through fine-grained gradient estimation.
Towards Codable Watermarking for Injecting Multi-Bits Information to LLMs
Lean Wang (Peking University), Xu Sun (Peking University)
GenerationOptimizationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper studies the technique of Coded Text Watermarking (CTWL) and proposes the Balance-Marking method, which balances vocabulary partitioning using a proxy language model to embed multi-bit information in LLM-generated text while maintaining text quality.
Towards Cross Domain Generalization of Hamiltonian Representation via Meta Learning
Yeongwoo Song (Korea Advanced Institute of Science and Technology), Hawoong Jeong (Korea Advanced Institute of Science and Technology)
Domain AdaptationRepresentation LearningMeta LearningGraph Neural NetworkTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: Using meta-learning (MAML) and graph neural networks (GCN) to construct Hamiltonian neural networks, learning a unified Hamiltonian representation across physical domains, and achieving rapid adaptation to unknown physical systems.
Towards Diverse Behaviors: A Benchmark for Imitation Learning with Human Demonstrations
Xiaogang Jia (Karlsruhe Institute of Technology), Gerhard Neumann (Karlsruhe Institute of Technology)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerDiffusion modelMultimodalityBenchmark
🎯 What it does: A D3IL benchmark and dataset are proposed, specifically designed to evaluate the ability to learn multimodal behaviors from human demonstrations, along with a quantifiable measure of behavioral diversity.
Towards domain-invariant Self-Supervised Learning with Batch Styles Standardization
Marin Scalbert (CentraleSupélec, Université Paris-Saclay), Florent Couzinie-Devy (VitaDX)
ClassificationDomain AdaptationContrastive LearningImageBiomedical Data
🎯 What it does: This paper proposes a method called Batch Styles Standardization (BSS), which utilizes amplitude information in the Fourier domain to unify the styles of images within the same batch, thereby reducing style/domain-related spurious correlations in self-supervised learning (SSL) and enhancing generalization ability in unknown domains.
Towards Eliminating Hard Label Constraints in Gradient Inversion Attacks
Yanbo Wang (University of Chinese Academy of Sciences), Ran He (Chinese Academy of Sciences)
RestorationOptimizationAdversarial AttackConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: An algorithm is proposed to analytically recover soft labels (label smoothing, mixup) and corresponding fully connected layer features from a single sample gradient inversion.
Towards Energy Efficient Spiking Neural Networks: An Unstructured Pruning Framework
Xinyu Shi (Peking University), Zhaofei Yu (Peking University)
Spiking Neural NetworkImage
🎯 What it does: A new framework that combines unstructured weight pruning and unstructured neuron pruning is proposed, aiming to maximize the sparsity of Spiking Neural Networks (SNN) on neuromorphic hardware to enhance energy efficiency.
Towards Enhancing Time Series Contrastive Learning: A Dynamic Bad Pair Mining Approach
Xiang Lan (National University of Singapore), Mengling Feng (National University of Singapore)
Anomaly DetectionRepresentation LearningContrastive LearningTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: This paper proposes a Dynamic Bad Positive Sample Mining (DBPM) algorithm, which utilizes a memory module to track the loss of positive samples and dynamically identify noise and erroneous positive samples. Subsequently, it adjusts their weights through a transformation module to enhance the representation quality of temporal contrastive learning.
Towards Establishing Guaranteed Error for Learned Database Operations
Sepanta Zeighami (University of California Berkeley), Cyrus Shahabi (University of Southern California)
Tabular
🎯 What it does: This study investigates the theoretical lower bounds of model size for learning-based database operations (indexing, cardinality estimation, interval summation), providing guarantees in the worst-case and average error scenarios.
Towards Faithful Explanations: Boosting Rationalization with Shortcuts Discovery
Linan Yue (University of Science and Technology of China), Yanqing An (University of Science and Technology of China)
ClassificationExplainability and InterpretabilityTransformerSupervised Fine-TuningText
🎯 What it does: A semi-supervised interpretability method SSR based on the discovery and utilization of 'shortcut keys' is proposed, significantly improving prediction accuracy and interpretability in text classification tasks.
Towards Faithful XAI Evaluation via Generalization-Limited Backdoor Watermark
Mengxi Ya (Tsinghua University), Shu-Tao Xia (Tsinghua University)
OptimizationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a 'Generalized Limited Backdoor Watermark' (GLBW) to improve the automated evaluation of saliency-based representation visualization (SRV).
Towards Few-Shot Adaptation of Foundation Models via Multitask Finetuning
Zhuoyan Xu (University of Wisconsin Madison), Yingyu Liang (University of Wisconsin Madison)
Domain AdaptationMeta LearningTransformerSupervised Fine-TuningContrastive LearningImageText
🎯 What it does: Multi-task fine-tuning of the base model is conducted, theoretically analyzing its advantages in downstream tasks with a small number of labels, and proposing a task selection algorithm based on diversity and consistency.
Towards Foundation Models for Knowledge Graph Reasoning
Mikhail Galkin (Intel AI Lab), Zhaocheng Zhu (Mila University of Montreal)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes ULTRA, a foundational model that enables zero-shot reasoning and fine-tuning across different knowledge graphs by constructing a relation graph and learning conditional relation representations to achieve cross-graph knowledge inference.
Towards Foundational Models for Molecular Learning on Large-Scale Multi-Task Datasets
Dominique Beaini (Mila - Quebec AI Institute), Dominic Masters (Graphcore)
Drug DiscoveryGraph Neural NetworkGraphBiomedical Data
🎯 What it does: A large-scale multi-task molecular dataset covering nearly 100 billion molecules and over 3000 tasks has been developed, along with the introduction of the Graphium library for efficient training.
Towards Generative Abstract Reasoning: Completing Raven’s Progressive Matrix via Rule Abstraction and Selection
Fan Shi (Fudan University), Xiangyang Xue (Fudan University)
GenerationData SynthesisAuto EncoderImage
🎯 What it does: A generative Raven’s Progressive Matrix solver called RAISE is proposed, which is based on a conditional deep latent variable model. It can abstract attribute concepts from contextual images and select corresponding atomic rules to generate missing images at any position.
Towards Green AI in Fine-tuning Large Language Models via Adaptive Backpropagation
Kai Huang (University of Pittsburgh), Wei Gao (University of Pittsburgh)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: For fine-tuning large language models, GreenTrainer is proposed to dynamically select training tensors through adaptive backpropagation, significantly reducing FLOPs while maintaining or even improving performance.
Towards Identifiable Unsupervised Domain Translation: A Diversified Distribution Matching Approach
Sagar Shrestha (Oregon State University), Xiao Fu (Oregon State University)
Image TranslationDomain AdaptationGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes an unsupervised domain translation method based on diversified distribution matching, addressing the identifiability issues present in traditional methods.
Towards image compression with perfect realism at ultra-low bitrates
Marlene Careil (Meta AI), Stéphane Lathuilière (Telecom Paris)
CompressionDiffusion modelImage
🎯 What it does: A diffusion model-based image compression framework called PerCo is proposed, achieving realistic reconstruction at low bit rates.
Towards Imitation Learning to Branch for MIP: A Hybrid Reinforcement Learning based Sample Augmentation Approach
Changwen Zhang (Lenovo Research), Junchi Yan (Shanghai Jiao Tong University)
OptimizationGraph Neural NetworkReinforcement LearningTabular
🎯 What it does: A hybrid online and offline reinforcement learning sample augmentation framework (HRL-Aug) is proposed, which generates training samples by dynamically selecting experts or learned policies and performs quality screening on the samples to enhance the learning strategy for variable selection in branch-and-bound.
Towards LLM4QPE: Unsupervised Pretraining of Quantum Property Estimation and A Benchmark
Yehui Tang (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Recurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTime SeriesSequentialBenchmarkPhysics Related
🎯 What it does: This paper proposes an unsupervised pre-training framework called LLM4QPE, which first pre-trains a Transformer+LSTM model using a large amount of unlabeled quantum measurement records, and then fine-tunes it on a limited amount of labeled data to achieve task-independent estimation of quantum system properties.
Towards Lossless Dataset Distillation via Difficulty-Aligned Trajectory Matching
Ziyao Guo (National University of Singapore), Yang You (National University of Singapore)
Knowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: A Difficulty-Aligned Trajectory Matching (DATM) method is proposed to achieve lossless dataset distillation under both low IPC and high IPC conditions, enhancing distillation stability through soft label learning and sequential generation.
Towards Meta-Pruning via Optimal Transport
Alexander Theus (ETH Zurich), Sidak Pal Singh (ETH Zurich)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: A new structured pruning method called Intra-Fusion is proposed, which utilizes Optimal Transport to fuse the pruned neurons instead of simply discarding them, thereby maintaining model performance without data or fine-tuning.
Towards Non-Asymptotic Convergence for Diffusion-Based Generative Models
Gen Li (Chinese University of Hong Kong), Yuejie Chi (Carnegie Mellon University)
GenerationData SynthesisDiffusion modelScore-based ModelOrdinary Differential Equation
🎯 What it does: This paper develops a non-asymptotic theory to understand the data generation process of diffusion models in discrete time, assuming that accurate (Stein) score function estimates can be obtained.
Towards Offline Opponent Modeling with In-context Learning
Yuheng Jing (Institute of Automation Chinese Academy of Sciences), Jian Cheng (Institute of Automation Chinese Academy of Sciences)
Meta LearningTransformerReinforcement LearningContrastive LearningSequential
🎯 What it does: This paper proposes the Offline Opponent Modeling (OOM) task and introduces a three-stage method based on Transformer called TAO. It first learns opponent strategy embeddings, then trains response strategies using in-scene contextual learning, and finally achieves rapid adaptation in new environments through an opponent context window.
Towards Optimal Feature-Shaping Methods for Out-of-Distribution Detection
Qinyu Zhao (Australian National University), Stephen Gould (Australian National University)
Anomaly DetectionOptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: A feature reshaping method is proposed that can be trained without OOD samples to enhance the generalization ability of OOD detection.
Towards Optimal Regret in Adversarial Linear MDPs with Bandit Feedback
Haolin Liu (University of Virginia), Julian Zimmert (Google Research)
OptimizationReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: This paper proposes two algorithms to handle adversarial loss in linear MDPs with unknown transitions and bandit feedback, aiming to minimize the cumulative loss difference (regret) with respect to the optimal fixed policy.
Towards Poisoning Fair Representations
Tianci Liu (Purdue University), Jing Gao (Purdue University)
Representation LearningAdversarial AttackTabular
🎯 What it does: A white-box clean label data poisoning attack is proposed for Fair Representation Learning (FRL) models, aiming to make the learned representations highly correlated with sensitive attributes, thereby undermining fairness.
Towards Principled Representation Learning from Videos for Reinforcement Learning
Dipendra Misra (Microsoft Research), John Langford (Microsoft Research)
Representation LearningReinforcement LearningAuto EncoderContrastive LearningVideo
🎯 What it does: This paper constructs offline pre-trained representations based on video data and theoretically analyzes their feasibility in reinforcement learning (RL). It proves that under a Block MDP without external noise, forward modeling and temporal contrastive learning can learn latent state representations that are useful for efficient downstream RL. It also provides a lower bound showing that when external noise is present, the sample complexity of video pre-training deteriorates exponentially compared to trajectory pre-training. Experiments are then conducted in visual environments such as GridWorld and ViZDoom to validate these findings.
Towards Reliable and Efficient Backdoor Trigger Inversion via Decoupling Benign Features
Xiong Xu (Zhejiang University), Kui Ren (Zhejiang University)
OptimizationAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A new backdoor trigger inversion (BTI) method called BTI-DBF is proposed to address backdoor attacks in third-party models. This method recovers triggers by decoupling normal features rather than directly decoupling backdoor features, and based on this, two types of backdoor defenses (removal and preprocessing) are designed.
Towards Robust and Efficient Cloud-Edge Elastic Model Adaptation via Selective Entropy Distillation
Yaofo Chen (South China University of Technology), Mingkui Tan (South China University of Technology)
Domain AdaptationComputational EfficiencyKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: A Cloud-Edge Elastic Model Adaptation (CEMA) framework is proposed, where edge devices only perform forward inference, while the cloud is responsible for self-supervised entropy minimization, replay-based knowledge distillation, and parameter updates, achieving dynamic and low-cost online model adaptation.
Towards Robust Fidelity for Evaluating Explainability of Graph Neural Networks
Xu Zheng (Florida International University), Dongsheng Luo (Florida International University)
Explainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: A robust Fidelity metric for evaluating the interpretability of graph neural networks is proposed, addressing the bias and instability of traditional Fidelity metrics under subgraph distribution shifts.
Towards Robust Multi-Modal Reasoning via Model Selection
Xiangyan Liu (National University of Singapore), Tao Lin (Westlake University)
OptimizationRepresentation LearningGraph Neural NetworkVision Language ModelMultimodalityBenchmark
🎯 What it does: A framework named M³ is proposed for dynamic model selection during multi-modal multi-step reasoning processes, thereby enhancing overall reasoning robustness.
Towards Robust Offline Reinforcement Learning under Diverse Data Corruption
Rui Yang (Hong Kong University of Science and Technology), Tong Zhang (University of Texas at Austin)
Reinforcement LearningTabular
🎯 What it does: This paper evaluates the robustness of offline reinforcement learning algorithms under various types of data contamination (state, action, reward, dynamics) and proposes Robust IQL (RIQL) to improve performance in these contaminated scenarios.
Towards Robust Out-of-Distribution Generalization Bounds via Sharpness
Yingtian Zou (National University of Singapore), Wynne Hsu (National University of Singapore)
Domain AdaptationOptimizationConvolutional Neural NetworkImageTabular
🎯 What it does: This paper studies the OOD generalization theory, proposing an OOD generalization upper bound based on algorithm robustness and loss sharpness, and provides theoretical proofs and experimental validation.
Towards Seamless Adaptation of Pre-trained Models for Visual Place Recognition
Feng Lu (Tsinghua University), Chun Yuan (Tsinghua University)
RecognitionRetrievalTransformerContrastive LearningImage
🎯 What it does: A method that integrates global and local features, called SelaVPR, is proposed to achieve seamless transfer of the pre-trained visual foundation model (DINOv2) for the visual place recognition (VPR) task.
Towards the Fundamental Limits of Knowledge Transfer over Finite Domains
Qingyue Zhao (Tsinghua University), Banghua Zhu (University of California)
Knowledge DistillationTabular
🎯 What it does: The study investigates the statistical efficiency of knowledge transfer in finite fields, providing the optimal sampling complexity for three levels of teacher information acquisition (hard labels, partial soft labels, and complete soft labels);
Towards Training Without Depth Limits: Batch Normalization Without Gradient Explosion
Alexandru Meterez (ETH Zurich), Hadi Daneshmand (MIT)
Image
🎯 What it does: This paper proposes a multilayer perceptron (MLP) using orthogonal random weights under batch normalization (BN) and demonstrates that it can avoid gradient explosion while maintaining good signal propagation.
Towards Transparent Time Series Forecasting
Krzysztof Kacprzyk (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
Explainability and InterpretabilityRecurrent Neural NetworkTime SeriesOrdinary Differential Equation
🎯 What it does: A time series transparent prediction framework called TIMEVIEW based on static features is proposed, utilizing two-layer (trend and attribute) transparency for interpretation;
Towards Understanding Factual Knowledge of Large Language Models
Xuming Hu (Hong Kong University of Science and Technology), Zhijiang Guo (University of Cambridge)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposed the Pinocchio benchmark, constructing 20,713 multilingual, multi-domain, and multi-temporal factual knowledge and reasoning questions to systematically evaluate the capabilities of LLMs in factual memory and reasoning;
Towards Understanding Sycophancy in Language Models
Mrinank Sharma (Anthropic), Ethan Perez (Anthropic)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper systematically evaluates the phenomenon of sycophancy (catering to user preferences) in five mainstream AI assistants during open-text generation tasks. It explores the incentive mechanisms of human feedback and the Preference Model (PM) for sycophancy by transforming human preference data into interpretable features and using Bayesian logistic regression modeling. Furthermore, it validates the impact of PM on model behavior through Best-of-N sampling and RL optimization, and finally assesses the preference differences between humans and PM in correcting misunderstandings.
Towards Unified Multi-Modal Personalization: Large Vision-Language Models for Generative Recommendation and Beyond
Tianxin Wei (University of Illinois Urbana Champaign), Xianfeng Tang (Amazon)
Recommendation SystemTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: A unified multimodal personalized system called UniMP is proposed, which can simultaneously process various user historical information such as text, images, and IDs, and generate multimodal results.
Tractable MCMC for Private Learning with Pure and Gaussian Differential Privacy
Yingyu Lin (University of California San Diego), Zhiqi Bu (Amazon)
OptimizationSafty and Privacy
🎯 What it does: A new algorithm called ASAP (Approximate Sample Perturbation) is proposed, which achieves pure differential privacy or pure Gaussian differential privacy by adding noise (the noise scale is related to the Wasserstein-∞ distance) to the approximate posterior samples obtained from MCMC sampling.
Tractable Probabilistic Graph Representation Learning with Graph-Induced Sum-Product Networks
Federico Errica (NEC Laboratories Europe), Mathias Niepert (University of Stuttgart)
Representation LearningDrug DiscoveryGraph Neural NetworkGraph
🎯 What it does: This paper proposes Graph-Induced Sum-Product Networks (GSPN), which achieve inferable probabilistic graphical representation learning by mapping the computational tree of a graph to a hierarchical SPN structure.
Training Bayesian Neural Networks with Sparse Subspace Variational Inference
Junbo Li (Purdue University), Ruqi Zhang (Purdue University)
ClassificationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A fully sparse Bayesian neural network training framework SSVI is proposed, which maintains high sparsity during both training and inference.