ICLR 2024 Papers — Page 18
International Conference on Learning Representations · 2260 papers
Rethinking Label Poisoning for GNNs: Pitfalls and Attacks
Vijay Lingam (CISPA Helmholtz Center for Information Security), Aleksandar Bojchevski (University of Cologne)
Adversarial AttackMeta LearningGraph Neural NetworkNeural Radiance FieldGraph
🎯 What it does: This paper studies the robustness of Graph Neural Networks (GNN) under label poisoning and points out six major evaluation flaws in previous research; by correcting the evaluation framework, it systematically assesses various attack methods and proposes two new efficient attacks (linear proxy attack and meta-learning attack), revealing that binary label poisoning is often more effective.
Rethinking Model Ensemble in Transfer-based Adversarial Attacks
Huanran Chen (Beijing Institute of Technology), Jun Zhu (Tsinghua University)
Object DetectionAdversarial AttackConvolutional Neural NetworkTransformerVision Language ModelImage
🎯 What it does: This paper researches and implements a new model ensemble-based adversarial attack method aimed at improving the attack success rate in black-box environments.
Rethinking the Benefits of Steerable Features in 3D Equivariant Graph Neural Networks
Shih-Hsin Wang (University of Utah), Bao Wang (University of Utah)
Graph Neural NetworkGraph
🎯 What it does: The paper studies the impact of using different types (L) of steerable features in three-dimensional equivariant graph neural networks (EGNN) on the network's expressive power and performance, providing corresponding theoretical analysis and numerical validation.
Rethinking the Power of Graph Canonization in Graph Representation Learning with Stability
Zehao Dong (Washington University in St. Louis), Yixin Chen (Washington University in St. Louis)
Representation LearningGraph Neural NetworkGraphBiomedical Data
🎯 What it does: This paper constructs the GC-GNN framework by using the discrete node colors generated by graph canonization as positional encoding, significantly enhancing the expressive power of GNNs, and further proposes the Universal Graph Canonization (UGC-GNN) to address the trade-off between expressiveness and stability.
Rethinking the symmetry-preserving circuits for constrained variational quantum algorithms
Ge Yan (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationDrug DiscoveryReinforcement LearningTabularPhysics Related
🎯 What it does: This paper studies the feasibility of using Hamming Weight (HW) to maintain ansatz in constrained variational quantum algorithms (VQA). It provides a theoretical mapping from symmetry constraints to HW maintaining ansatz and constructs a new gate (BS gate) that is globally controllable in the HW subspace using quantum optimal control and overparameterization theory. Subsequently, experimental validation was conducted on ground state energy estimation and feature selection tasks based on the subspace.
Rethinking the Uniformity Metric in Self-Supervised Learning
Xianghong Fang (Chinese University of Hong Kong), Benyou Wang (Chinese University of Hong Kong)
Representation LearningContrastive LearningImage
🎯 What it does: A uniformity metric based on the squared Wasserstein distance is proposed and introduced as an auxiliary loss in various self-supervised learning methods to enhance the uniformity of representations and suppress dimensional collapse.
Retrieval is Accurate Generation
Bowen Cao (Peking University), Shuming Shi (Tencent AI Lab)
GenerationRetrievalLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: A language model (CoG-2) is proposed that directly retrieves contextually relevant phrases instead of generating them word by word.
Retrieval meets Long Context Large Language Models
Peng Xu (NVIDIA), Bryan Catanzaro (NVIDIA)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper systematically compares the performance of two approaches, retrieval-augmentation and long-context window, in large-scale language models on zero-shot and few-shot tasks, and conducts large-scale experiments on GPT-43B and Llama2-70B.
Retrieval-based Disentangled Representation Learning with Natural Language Supervision
Jiawei Zhou (Hong Kong University of Science and Technology), Lei Chen (Hong Kong University of Science and Technology)
RetrievalRepresentation LearningTransformerContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a retrieval-based separable representation learning framework (VDR) that utilizes natural language as a supervisory signal to map data and text into the same lexical space, achieving dimension-level separation through sparse word vectors.
Retrieval-Enhanced Contrastive Vision-Text Models
Ahmet Iscen (Google Research), Cordelia Schmid (Google Research)
RetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A lightweight retrieval-fusion module is added to the pre-trained CLIP model, which retrieves cross-modal information through external memory during inference and fuses it into the original embeddings.
Retrieval-Guided Reinforcement Learning for Boolean Circuit Minimization
Animesh Basak Chowdhury (New York University), Siddharth Garg (New York University)
OptimizationGraph Neural NetworkTransformerReinforcement LearningGraph
🎯 What it does: This study proposes a retrieval-based reinforcement learning method called ABC-RL, which is used to automatically find high-quality Boolean circuit synthesis paths, significantly improving synthesis quality and reducing search time.
Retro-fallback: retrosynthetic planning in an uncertain world
Austin Tripp (University of Cambridge), José Miguel Hernández-Lobato (University of Cambridge)
OptimizationDrug DiscoveryGraphBenchmarkStochastic Differential Equation
🎯 What it does: This paper proposes a stochastic process framework that considers the uncertainty of reaction feasibility and availability in chemical retro-synthesis planning, and based on this, defines a success synthesis probability (SSP) evaluation metric. Subsequently, a greedy search algorithm called Retro-Fallback is designed to maximize this metric.
RetroBridge: Modeling Retrosynthesis with Markov Bridges
Ilia Igashov (Ecole Polytechnique Federale de Lausanne), Bruno Correia (Ecole Polytechnique Federale de Lausanne)
GenerationDrug DiscoveryGraph Neural NetworkTransformerGraph
🎯 What it does: This paper proposes a generative framework called RetroBridge based on Markov bridges, which directly generates possible precursor molecules from target molecules to achieve single-step retro-synthesis prediction.
Retroformer: Retrospective Large Language Agents with Policy Gradient Optimization
Weiran Yao (Salesforce AI Research), Silvio Savarese (Salesforce AI Research)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Proposes the Retroformer framework, which utilizes a trainable retrospective language model to self-reflect and correct the prompts of LLM agents, learning policy gradients from environmental rewards;
RETSim: Resilient and Efficient Text Similarity
Marina Zhang (Google), Elie Bursztein (Google)
RetrievalData-Centric LearningTransformerTextBenchmark
🎯 What it does: A lightweight, multilingual text similarity model called RETSim has been developed, which can efficiently retrieve and cluster approximately duplicate texts in the presence of spelling errors, character-level attacks, and adversarial transformations, and is used for dataset deduplication and spam clustering.
REValueD: Regularised Ensemble Value-Decomposition for Factorisable Markov Decision Processes
David Ireland (University of Warwick), Giovanni Montana (Alan Turing Institute)
Reinforcement LearningTabular
🎯 What it does: An algorithm named REValueD is proposed to solve the Factorized Markov Decision Process (FMDP) problem with high-dimensional discrete sub-action spaces. It combines value decomposition, Critic ensemble, and regularization loss to reduce target variance and alleviate the credit assignment problem caused by exploring sub-actions.
Reverse Diffusion Monte Carlo
Xunpeng Huang (Hong Kong University of Science and Technology), Tong Zhang (University of Illinois Urbana-Champaign)
Diffusion modelStochastic Differential Equation
🎯 What it does: A Monte Carlo sampling algorithm based on the reverse diffusion process, rdMC, is proposed, which utilizes the analytical solution of the normal OU process to transform score matching into posterior mean estimation without the need to train a parametric model.
Reverse Forward Curriculum Learning for Extreme Sample and Demo Efficiency
Stone Tao (University of California San Diego), Hao Su (University of California San Diego)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: A reverse + forward curriculum learning framework is proposed, which can train complex robot control strategies with only a small number of demonstrations under sparse rewards.
Revisit and Outstrip Entity Alignment: A Perspective of Generative Models
Lingbing Guo (Zhejiang University), Huajun Chen (Zhejiang University)
GenerationData SynthesisRetrievalAuto EncoderGenerative Adversarial NetworkMultimodalityGraphBenchmark
🎯 What it does: This paper proposes a generative model-based entity alignment framework called GEEA, which can simultaneously perform entity alignment and entity synthesis.
Revisiting Data Augmentation in Deep Reinforcement Learning
Jianshu Hu (Shanghai Jiao Tong University), Paul Weng (Duke Kunshan University)
Reinforcement LearningImageVideo
🎯 What it does: Theoretical and experimental analysis of data augmentation techniques in deep reinforcement learning on image benchmarks is conducted, proposing a novel actor-critic algorithm that includes explicit/implicit regularization and tangent propagation, along with a unified implementation framework.
Revisiting Deep Audio-Text Retrieval Through the Lens of Transportation
Manh Luong (Monash University), Lizhen Qu (Monash University)
RetrievalContrastive LearningTextAudio
🎯 What it does: A mini-batch Learning-to-Match (m-LTM) framework for audio-text retrieval is proposed, combining Mahalanobis distance and partial optimal transport to enhance the quality of the embedding space and noise tolerance.
Revisiting Link Prediction: a data perspective
Haitao Mao (Michigan State University), Jiliang Tang (Michigan State University)
Graph Neural NetworkGraphBenchmark
🎯 What it does: Revisiting link prediction from a data perspective, this paper systematically explores three key data factors (local structural proximity, global structural proximity, and feature proximity) and their interrelationships. It demonstrates the incompatibility between feature proximity and structural proximity, revealing potential flaws in the GNN4LP model on feature-dominant edges, and subsequently proposes a decoupled SEAL architecture along with new dataset categories and selection guidelines.
Revisiting Plasticity in Visual Reinforcement Learning: Data, Modules and Training Stages
Guozheng Ma (Tsinghua University), Dacheng Tao (Nanyang Technological University)
Reinforcement LearningImage
🎯 What it does: This paper systematically studies the plasticity loss of neural networks in visual reinforcement learning, exploring the impact of data augmentation, modules, and training phases on plasticity, and proposes an adaptive replay ratio method.
Revisiting the Last-Iterate Convergence of Stochastic Gradient Methods
Zijian Liu (New York University), Zhengyuan Zhou (New York University)
Optimization
🎯 What it does: This paper conducts a unified analysis of the Composite Stochastic Mirror Descent (CSMD) algorithm and studies the convergence of the last iteration of stochastic gradient methods.
Reward Design for Justifiable Sequential Decision-Making
Aleksa Sukovic (Max Planck Institute for Software Systems), Goran Radanovic (Max Planck Institute for Software Systems)
Explainability and InterpretabilityDrug DiscoveryReinforcement LearningBiomedical DataElectronic Health Records
🎯 What it does: A debate-based reward model is proposed to train reinforcement learning agents to provide decisions and corresponding evidence that can be recognized by human reviewers at each state, and its effectiveness is validated in the sepsis treatment scenario.
Reward Model Ensembles Help Mitigate Overoptimization
Thomas Coste (University of Cambridge), David Krueger (University College London)
OptimizationReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This study investigates the use of reward model ensemble and conservative optimization (WCO, UWO) in RLHF to eliminate over-optimization, validating its effectiveness through two strategies: BoN and PPO.
Reward-Consistent Dynamics Models are Strongly Generalizable for Offline Reinforcement Learning
Fan-Ming Luo (Nanjing University), Yang Yu (Nanjing University)
Reinforcement Learning
🎯 What it does: The MOREC method is proposed, which learns a 'dynamic reward' in offline reinforcement learning and uses it to filter transitions generated by the model, thereby enhancing the authenticity of model rollouts and the performance of the policy.
Reward-Free Curricula for Training Robust World Models
Marc Rigter (University of Oxford), Ingmar Posner (University of Oxford)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningWorld ModelSequential
🎯 What it does: A reward-agnostic adaptive curriculum learning method (WAKER) is proposed for training world models that maintain robustness across different environments.
Rigid Protein-Protein Docking via Equivariant Elliptic-Paraboloid Interface Prediction
Ziyang Yu (Tsinghua University), Yang Liu (Tsinghua University)
Protein Structure PredictionGraph Neural NetworkTransformerGraphBenchmark
🎯 What it does: A rigid protein-protein docking method called ElliDock was developed, which predicts the interface based on elliptical paraboloid and aligns the docking transformation directly from two unbound proteins.
Ring-A-Bell! How Reliable are Concept Removal Methods For Diffusion Models?
Yu-Lin Tsai (National Yang Ming Chiao Tung University), Chun-Ying Huang (National Yang Ming Chiao Tung University)
Adversarial AttackPrompt EngineeringDiffusion modelImageText
🎯 What it does: Ring-A-Bell, a model-agnostic red team tool, has been developed to offline generate adversarial prompts to evaluate the security mechanisms of text-image diffusion models (such as filters and concept elimination).
RingAttention with Blockwise Transformers for Near-Infinite Context
Hao Liu (University of California Berkeley), Pieter Abbeel (University of California Berkeley)
TransformerLarge Language ModelReinforcement LearningTextSequential
🎯 What it does: Proposes an architecture that combines RingAttention with Blockwise Transformers, utilizing ring communication and block-level parallelism to address the memory bottleneck of Transformers when processing long sequences, achieving near-infinite context for training and inference.
Risk Bounds of Accelerated SGD for Overparameterized Linear Regression
Xuheng Li (University of California Los Angeles), Quanquan Gu (University of California Los Angeles)
OptimizationTabular
🎯 What it does: This study investigates the generalization risk of Accelerated Stochastic Gradient Descent (ASGD) in over-parameterized linear regression and provides an instance-dependent risk upper bound.
RLCD: Reinforcement Learning from Contrastive Distillation for LM Alignment
Kevin Yang (Meta AI), Yuandong Tian (Meta AI)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes a new language model alignment method—RLCD (Reinforcement Learning from Contrastive Distillation), which automatically labels preference pairs by generating positive and negative contrastive prompts and is used in the RLHF process.
RLIF: Interactive Imitation Learning as Reinforcement Learning
Jianlan Luo (University of California Berkeley), Sergey Levine (University of California Berkeley)
Robotic IntelligenceReinforcement Learning
🎯 What it does: This paper proposes a reinforcement learning method based on user intervention feedback, RLIF, which uses the timing of interventions in interactive demonstrations as rewards to train a control policy for a reward function that does not require assistance.
Robot Fleet Learning via Policy Merging
Lirui Wang (Massachusetts Institute of Technology), Russ Tedrake (Massachusetts Institute of Technology)
Federated LearningRobotic IntelligenceRecurrent Neural NetworkReinforcement LearningTabular
🎯 What it does: This paper studies a distributed strategy merging method for robot teams and proposes the FLEET-MERGE algorithm to enable multi-robot learning without the need to share raw data.
Robust Adversarial Reinforcement Learning via Bounded Rationality Curricula
Aryaman Reddi (Technical University of Darmstadt), Carlo D'Eramo (University of Würzburg)
Robotic IntelligenceReinforcement Learning
🎯 What it does: A quantized adversarial reinforcement learning algorithm QARL based on entropy regularization is proposed to achieve adversarial robustness in high-dimensional control tasks.
Robust agents learn causal world models
Jonathan Richens (Google DeepMind), Tom Everitt (Google DeepMind)
Reinforcement LearningAgentic AIGraph
🎯 What it does: This paper proves that any agent capable of maintaining low regret under significant distribution shifts must have learned the causal model of the data generation process.
Robust Angular Synchronization via Directed Graph Neural Networks
Yixuan He (University of Oxford), Mihai Cucuringu (University of Oxford)
Graph Neural NetworkPoint CloudGraph
🎯 What it does: This paper proposes the GNNSync framework based on directed graph neural networks to address the angle synchronization and its k-synchronization extension problem, capable of accurately estimating unknown angles in high noise environments.
Robust Classification via Regression for Learning with Noisy Labels
Erik Englesson (KTH Royal Institute of Technology), Hossein Azizpour (KTH Royal Institute of Technology)
ClassificationImage
🎯 What it does: A unified method is proposed to transform classification tasks into regression tasks, combining loss reweighting and label correction to enhance robustness against noisy labels.
Robust Model Based Reinforcement Learning Using $\mathcal{L}_1$ Adaptive Control
Minjun Sung (University of Illinois), Naira Hovakimyan (University of Illinois)
Reinforcement LearningSequential
🎯 What it does: A robust enhancement module based on $L_{1}$ adaptive control is proposed, which can non-invasively improve the performance and sample efficiency of model-driven reinforcement learning algorithms in the presence of system uncertainties.
Robust Model-Based Optimization for Challenging Fitness Landscapes
Saba Ghaffari (University of Illinois Urbana-Champaign), Saurabh Sinha (Georgia Institute of Technology)
OptimizationDrug DiscoveryAuto EncoderSequentialBiomedical Data
🎯 What it does: A property-prioritized variational autoencoder (PPGVAE) is proposed and implemented for model-driven optimization, capable of overcoming the local optimum challenges caused by high imbalance and separation in protein design and continuous optimization problems.
Robust NAS under adversarial training: benchmark, theory, and beyond
Yongtao Wu (École Polytechnique Fédérale de Lausanne), Volkan Cevher (École Polytechnique Fédérale de Lausanne)
Adversarial AttackNeural Architecture SearchConvolutional Neural NetworkGenerative Adversarial NetworkImageBenchmark
🎯 What it does: We released NAS-RobBench-201, a benchmark containing the performance of 6466 networks trained adversarially on CIFAR-10/100 and ImageNet-16-120; we also provided a theoretical analysis of NTK related to this benchmark and its robustness.
Robust Similarity Learning with Difference Alignment Regularization
Shuo Chen (RIKEN Center for Advanced Intelligence Project), Masashi Sugiyama (RIKEN Center for Advanced Intelligence Project)
ClassificationRetrievalContrastive LearningImageText
🎯 What it does: This paper proposes a Differential Alignment Regularization (DAR) that enhances the robustness and generalization ability of similarity learning by encouraging feature differences among samples of different classes to converge.
Robust Training of Federated Models with Extremely Label Deficiency
Yonggang Zhang (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
Federated LearningContrastive LearningImage
🎯 What it does: Proposes the Twin-sight dual model framework, which addresses the gradient conflict problem caused by label scarcity in federated semi-supervised learning by separately training a supervised model and an unsupervised model, and aligning their features using neighborhood-preserving constraints.
Robustifying and Boosting Training-Free Neural Architecture Search
Zhenfeng He (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
Neural Architecture Search
🎯 What it does: This paper proposes a RoBoT algorithm that utilizes Bayesian optimization to learn a weighted linear combination of various training-free metrics to construct a more robust evaluation measure, and further enhances NAS performance through greedy search.
Robustifying State-space Models for Long Sequences via Approximate Diagonalization
Annan Yu (Cornell University), N. Benjamin Erichson (Lawrence Berkeley National Laboratory)
Time SeriesSequential
🎯 What it does: A perturbation-re-diagonalization (PTD) method is proposed, which re-diagonalizes the HiPPO matrix after adding a small perturbation, thereby achieving a robust diagonal state-space model (S4-PTD/S5-PTD) and improving performance on long sequence tasks.
Robustness of AI-Image Detectors: Fundamental Limits and Practical Attacks
Mehrdad Saberi (University of Maryland), Soheil Feizi (University of Maryland)
ClassificationRecognitionAdversarial AttackDiffusion modelGenerative Adversarial NetworkImageVideo
🎯 What it does: This study investigates the robustness of AI image detection methods, proposing diffusion purification attacks for low-disturbance watermarks and model substitution adversarial attacks for high-disturbance watermarks, demonstrating an irreconcilable trade-off between the robustness and reliability of classifier-based deepfake detectors.
RobustTSF: Towards Theory and Design of Robust Time Series Forecasting with Anomalies
Hao Cheng (University of California), Liang Sun (Alibaba Group)
Anomaly DetectionTime Series
🎯 What it does: A unified theory and method for time series forecasting with anomalies (TSFA) is proposed, and the RobustTSF algorithm is constructed.
Role of Locality and Weight Sharing in Image-Based Tasks: A Sample Complexity Separation between CNNs, LCNs, and FCNs
Aakash Lahoti (Carnegie Mellon University), Yuanzhi Li (Carnegie Mellon University)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a new dynamic signal distribution (DSD) classification task aimed at quantifying the differences in sample complexity among convolutional neural networks (CNN), locally connected networks (LCN), and fully connected networks (FCN) in image tasks.
Rotation Has Two Sides: Evaluating Data Augmentation for Deep One-class Classification
Guodong Wang (Beihang University), Di Huang (Beihang University)
ClassificationAnomaly DetectionRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper experimentally finds a strong linear correlation between the accuracy of rotation prediction and the detection performance of one-class classification (OCC). Based on this, a two-stage unsupervised method is proposed to estimate the distribution of rotation transformations in the dataset, thereby distinguishing between samples that maintain semantics (RAI) and those that transfer semantics (non-RAI), and improving the training process of OCC on this basis.
RT-Trajectory: Robotic Task Generalization via Hindsight Trajectory Sketches
Jiayuan Gu (Google DeepMind), Ted Xiao (Google DeepMind)
Robotic IntelligenceTransformerVision Language ModelVideo
🎯 What it does: This paper proposes a method for conditioning robot control strategies using rough two-dimensional trajectory sketches (RT-Trajectory), significantly enhancing the robot's generalization ability for new tasks and movements.
RTFS-Net: Recurrent Time-Frequency Modelling for Efficient Audio-Visual Speech Separation
Samuel Pegg (Tsinghua University), Xiaolin Hu (Tsinghua University)
RecognitionComputational EfficiencyRecurrent Neural NetworkVideoAudio
🎯 What it does: This paper proposes a time-frequency domain-based audio-video speech separation network called RTFS-Net, which utilizes complex time-frequency spectrograms obtained from STFT and employs bidirectional SRU, attention fusion, and spectral source separation modules to achieve high-quality separation.
S$2$AC: Energy-Based Reinforcement Learning with Stein Soft Actor Critic
Safa Messaoud (Qatar Computing Research Institute), Sanjay Chawla
Reinforcement Learning
🎯 What it does: An energy-based reinforcement learning algorithm S²AC based on Stein Variational Gradient Descent (SVGD) has been developed, which can learn expressive and multimodal policies within the maximum entropy RL framework and provide closed-form entropy estimates.
Safe and Robust Watermark Injection with a Single OoD Image
Shuyang Yu (Michigan State University), Jiayu Zhou (Michigan State University)
Anomaly DetectionConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: Using a single out-of-distribution (OoD) image to generate proxy data, the model is fine-tuned through backpropagation without using the original training data, injecting a verifiable backdoor watermark to ensure that the model maintains performance on the original task while achieving intellectual property protection.
Safe Collaborative Filtering
Riku Togashi (CyberAgent), Tetsuro Morimura (CyberAgent)
Recommendation SystemTabular
🎯 What it does: A CVaR-based safe collaborative filtering algorithm, SAFER2, is proposed, significantly improving the recommendation quality for low-satisfaction (tail) users.
Safe Offline Reinforcement Learning with Feasibility-Guided Diffusion Model
Yinan Zheng (Tsinghua University), Jingjing Liu (Tsinghua University)
Safty and PrivacyReinforcement LearningDiffusion modelTabularBenchmark
🎯 What it does: A secure offline reinforcement learning framework FISOR is proposed, which transforms hard safety constraints into feasibility-dependent objectives, and utilizes HJ reachability analysis, expectile regression, weighted advantage learning, and diffusion models to achieve the decoupled training of three objectives: safety, reward maximization, and offline data regularization.
Safe RLHF: Safe Reinforcement Learning from Human Feedback
Josef Dai, Yaodong Yang (Peking University)
Safty and PrivacyReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes Safe RLHF, an improved RLHF training framework that decouples human preferences for 'usefulness' and 'harmlessness', and dynamically balances the two through safety constraints and Lagrangian methods to achieve a safer and more useful LLM.
SafeDreamer: Safe Reinforcement Learning with World Models
Weidong Huang (Peking University), Yaodong Yang (Peking University)
Safty and PrivacyReinforcement LearningWorld ModelImage
🎯 What it does: This paper presents SafeDreamer, a safe reinforcement learning framework that combines Lagrangian methods with world model planning, achieving near-zero cost performance in visual and low-dimensional input tasks.
Safety-Tuned LLaMAs: Lessons From Improving the Safety of Large Language Models that Follow Instructions
Federico Bianchi (Stanford University), James Zou (Stanford University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This study investigates the impact of adding a small number of safety examples in instruction fine-tuning on the safety of large language models, and validates its effectiveness on various safety evaluation datasets.
SAFLEX: Self-Adaptive Augmentation via Feature Label Extrapolation
Mucong Ding (University of Maryland), Furong Huang (University of Maryland)
ClassificationDomain AdaptationData-Centric LearningContrastive LearningImageTabularBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposes the SAFLEX method, which automatically learns sample weights and soft labels for any augmented samples to enhance the overall model's generalization performance.
SALMON: Self-Alignment with Instructable Reward Models
Zhiqing Sun (Carnegie Mellon University), Chuang Gan (UMass Amherst)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: By using only a small number of human-defined principles and 6 examples, the SALMON method achieves self-alignment on large language models, resulting in the Dromedary-2 model;
SALMONN: Towards Generic Hearing Abilities for Large Language Models
Changli Tang (Tsinghua University), Chao Zhang (Tsinghua University)
TransformerLarge Language ModelSupervised Fine-TuningMultimodalityAudio
🎯 What it does: This paper presents SALMONN, a multimodal large language model that integrates speech, audio events, and music to achieve general auditory capabilities.
SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation
Chongyu Fan (Michigan State University), Sijia Liu (Michigan State University)
ClassificationGenerationDiffusion modelImage
🎯 What it does: A machine model forgetting method based on gradient weight saliency (SalUn) is proposed and implemented, applicable to image classification and generation tasks.
Sample Efficient Myopic Exploration Through Multitask Reinforcement Learning with Diverse Tasks
Ziping Xu (Harvard University), Ambuj Tewari (University of Michigan)
Reinforcement Learning
🎯 What it does: This paper proposes a multi-task reinforcement learning framework that utilizes a diverse set of tasks to achieve sample efficiency for greedy exploration (such as ε-greedy) in multi-task environments.
Sample-Efficiency in Multi-Batch Reinforcement Learning: The Need for Dimension-Dependent Adaptivity
Emmeran Johnson (Imperial College London), Patrick Rebeschini (University of Oxford)
Reinforcement Learning
🎯 What it does: This paper studies the relationship between sample efficiency and the degree of adaptability (number of batches K) under a multi-batch reinforcement learning framework. It proves that under d-dimensional linear function approximation, if the total number of queries is polynomially bounded, at least Ω(log log d) batches are required to achieve sample-efficient policy evaluation or optimal policy identification.
Sample-efficient Learning of Infinite-horizon Average-reward MDPs with General Function Approximation
Jianliang He (Fudan University), Zhuoran Yang (Yale University)
OptimizationReinforcement Learning
🎯 What it does: A unified theoretical framework is proposed to study sample-efficient learning of infinite average reward MDPs under general function approximation;
Sample-Efficient Learning of POMDPs with Multiple Observations In Hindsight
Jiacheng Guo (Princeton University), Yu Bai (Salesforce Research)
Reinforcement Learning
🎯 What it does: A new enhanced feedback model k-MOMDP is proposed, where after each round of interaction, the learner can obtain k-1 additional observations corresponding to the visited hidden states. Under this model, a sample-efficient learning subclass of POMDP is studied, and corresponding algorithms and theoretical sample complexities are provided.
Sample-Efficient Linear Representation Learning from Non-IID Non-Isotropic Data
Thomas TCK Zhang, Nikolai Matni (University of Pennsylvania)
Federated LearningRepresentation Learning
🎯 What it does: A linear representation learning framework is proposed for non-i.i.d. and non-homoscedastic data, aiming to recover shared linear operators or feature spaces from multi-task data.
Sample-Efficient Multi-Agent RL: An Optimization Perspective
Nuoya Xiong (Tsinghua University), Zhuoran Yang (Yale University)
OptimizationReinforcement Learning
🎯 What it does: A unified algorithmic framework MAMEX is proposed for sample-efficient multi-agent reinforcement learning in general summation Markov games (MG) through universal function approximation, capable of simultaneously addressing Nash, CCE, and CE equilibrium points.
Sample-Efficient Quality-Diversity by Cooperative Coevolution
Ke Xue (Nanjing University), Chao Qian (Nanjing University)
OptimizationReinforcement LearningSequential
🎯 What it does: A Cooperative Coevolution Quality-Diversity (CCQD) framework is designed, which significantly improves the sampling efficiency of QD algorithms by splitting the policy network into representation and decision layers, maintaining two subpopulations that co-evolve.
Sampling Multimodal Distributions with the Vanilla Score: Benefits of Data-Based Initialization
Frederic Koehler (University of Chicago), Thuy-Duong Vuong (Stanford University)
GenerationData SynthesisOptimizationScore-based ModelTabularStochastic Differential Equation
🎯 What it does: This study investigates the use of 'vanilla score' for generative modeling and demonstrates that effective sampling can be achieved on multi-modal mixtures of log-convex distributions through data-driven initialization and early stopping of Langevin dynamics.
SAN: Inducing Metrizability of GAN with Discriminative Normalized Linear Layer
Yuhta Takida (Sony AI), Yuki Mitsufuji (Sony Group Corporation)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a new GAN training framework called Slicing Adversarial Network (SAN), which ensures the metric consistency between the generated distribution and the real distribution by enforcing directional optimality in the discriminator, and can transform various existing GANs into SAN with two lightweight modifications to the discriminator.
SaNN: Simple Yet Powerful Simplicial-aware Neural Networks
Sravanthi Gurugubelli (Indian Institute of Science), Sundeep Prabhakar Chepuri (Indian Institute of Science)
Computational EfficiencyRepresentation LearningGraph Neural NetworkSpiking Neural NetworkMeshGraph
🎯 What it does: A simplified high-order graph network (SaNN) based on pre-aggregated simplex features is proposed, achieving almost constant complexity in memory and time during training while maintaining or surpassing the expressive power of existing SNN models.
SaProt: Protein Language Modeling with Structure-aware Vocabulary
Jin Su (Zhejiang University), Fajie Yuan (Westlake University)
Protein Structure PredictionTransformerLarge Language ModelSupervised Fine-TuningBiomedical Data
🎯 What it does: A structure-aware vocabulary is proposed, combining amino acid residues with Foldseek 3D structure tokens into a unified token, and training a 650M parameter self-supervised Transformer model (SaProt) on approximately 40 million AlphaFold2 predicted structures.
SAS: Structured Activation Sparsification
Yusuke Sekikawa (DENSO IT Lab Inc), Shingo Yashima (DENSO IT Lab Inc)
CompressionOptimizationComputational EfficiencyConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A method is proposed to compress activation maps into wide activations with local sparse structures through Structured Activation Sparsification (SAS), and to achieve efficient matrix multiplication using NVIDIA Sparse Tensor Core, improving model accuracy without increasing the multiplication count.
Scalable and Effective Implicit Graph Neural Networks on Large Graphs
Juncheng Liu (National University of Singapore), Xiaokui Xiao (Amazon.com Inc.)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes a scalable and efficient implicit graph neural network (SEIGNN) that combines mini-batch training with hierarchical sampling, coarse-grained nodes, and an unbiased random solver to efficiently train implicit GNNs on large graphs.
Scalable Diffusion for Materials Generation
Sherry Yang (Google DeepMind), Ekin Dogus Cubuk (University of Alberta)
GenerationData SynthesisDiffusion modelGraphPhysics Related
🎯 What it does: A unified crystal representation called UniMat is proposed, and based on this, a diffusion probabilistic model is trained for material generation; simultaneously, the stability of the generated materials is evaluated through DFT calculations.
Scalable Language Model with Generalized Continual Learning
Bohao PENG (Chinese University of Hong Kong), Jiaya Jia
ClassificationRetrievalOptimizationTransformerLarge Language ModelTextSequentialRetrieval-Augmented Generation
🎯 What it does: A scalable language model (SLM) is proposed to achieve continuous learning for sequence tasks, avoiding the use of experience replay, optimization constraints, or task IDs during inference.
Scalable Modular Network: A Framework for Adaptive Learning via Agreement Routing
Minyang Hu (Institute of Computing Technology Chinese Academy of Sciences), Xilin CHEN
ClassificationMeta LearningTransformerMixture of ExpertsImage
🎯 What it does: A scalable modular network (SMN) is proposed, which dynamically selects and combines specialized modules based on different inputs through an agreement router, and supports the introduction of new modules after pre-training to enhance adaptability.
Scalable Monotonic Neural Networks
Hyunho Kim (Sungkyunkwan University), Jong-Seok Lee (Sungkyunkwan University)
Tabular
🎯 What it does: This paper studies a scalable monotonic neural network (SMNN) that ensures monotonicity for specified inputs and achieves efficient training.
Scalable Neural Network Kernels
Arijit Sehanobish (Independent Researcher), Valerii Likhosherstov (Waymo)
ClassificationOptimizationComputational EfficiencyTransformerImageTabular
🎯 What it does: Proposes a Scalable Neural Network Kernel (SNNK) as an alternative to traditional Feedforward Layers (FFL), achieving more efficient forward computation by using only dot product kernels after decoupling inputs and parameters;
Scale-Adaptive Diffusion Model for Complex Sketch Synthesis
Jijin Hu (Beijing University of Posts and Telecommunications), Yi-Zhe Song (University of Surrey)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A diffusion model for pixel-level drawing is proposed, which dynamically adjusts the guidance scale to generate complex and easily recognizable hand-drawn sketches.
ScaleCrafter: Tuning-free Higher-Resolution Visual Generation with Diffusion Models
Yingqing He (Hong Kong University of Science and Technology), Ying Shan (Tencent AI Lab)
GenerationSuper ResolutionDiffusion modelImageVideoText
🎯 What it does: A method for generating ultra-high-resolution images and videos is proposed, achieved through noise-damped classifier-free guidance (CFG) using dynamic dilation convolution, convolution dispersion, and noise suppression, without any training or fine-tuning of the pre-trained diffusion model.
Scaling Convex Neural Networks with Burer-Monteiro Factorization
Arda Sahiner (Arcus Inc. Stanford University), Mert Pilanci (Stanford University)
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes the application of Burer-Monteiro factorization to convex neural networks (especially two-layer MLPs, CNNs, and self-attention networks with ReLU activation) to achieve efficient and scalable training, and provides relative optimality bounds.
Scaling for Training Time and Post-hoc Out-of-distribution Detection Enhancement
Kai Xu (National University of Singapore), Angela Yao (National University of Singapore)
Anomaly DetectionOptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper first conducts theoretical and experimental analysis of the existing Activation Shaping method (ASH), proving that its core enhancement of OOD detection is due to scaling the activations rather than pruning; it then proposes a post-hoc method called SCALE that only uses scaling, further improving OOD detection performance without reducing ID accuracy; based on this, the scaling concept is transferred to the training phase, designing Intermediate Tensor Shaping (ISH) for activation weighting optimization during training, thus achieving higher OOD detection performance with lower training costs.
Scaling Laws for Associative Memories
Vivien Cabannes (Meta), Alberto Bietti (Flatiron Institute)
Large Language ModelTabular
🎯 What it does: This paper studies the associative memory mechanism and proposes a model based on high-dimensional matrices, which is composed of embedded outer products, aimed at understanding the memory behavior of large language models.
Scaling Laws for Sparsely-Connected Foundation Models
Elias Frantar (Google DeepMind), Utku Evci (Google DeepMind)
TransformerImageText
🎯 What it does: This study investigates the scaling behavior of parameter sparsity in large-scale Transformers (for vision and language) during training on big data, and provides the corresponding joint scaling law.
Scaling Laws of RoPE-based Extrapolation
Xiaoran Liu (Fudan University), Dahua Lin (Shanghai AI Lab)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study investigates the extrapolation performance of LLMs based on RoPE when exceeding the pre-training context length and proposes a scaling law that can uniformly describe the extrapolation capability of RoPE.
Scaling physics-informed hard constraints with mixture-of-experts
Nithin Chalapathi (University of California), Aditi S. Krishnapriyan (University of California)
Mixture of ExpertsTime SeriesPhysics Related
🎯 What it does: A hard physical constraint framework based on Mixture-of-Experts is proposed, achieving strict adherence to physical laws in neural PDE solvers through local optimization.
Scaling Supervised Local Learning with Augmented Auxiliary Networks
Chenxiang Ma (Hong Kong Polytechnic University), KC Tan
ClassificationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A supervised local learning method named AugLocal is proposed, which enhances the collaboration between local layers and subsequent layers by uniformly sampling subsets of the subsequent layers in the auxiliary network of each hidden layer, achieving performance close to BP on large-scale networks.
SCHEMA: State CHangEs MAtter for Procedure Planning in Instructional Videos
Yulei Niu (Columbia University), Shih-Fu Chang (Columbia University)
TransformerLarge Language ModelPrompt EngineeringContrastive LearningVideoTextMultimodality
🎯 What it does: Utilize large language models to generate descriptions of state changes corresponding to steps, and learn a structured state space through visual-language alignment, thereby completing process planning in instructional videos.
Score Models for Offline Goal-Conditioned Reinforcement Learning
Harshit Sikchi (University of Texas at Austin), Scott Niekum (University of Texas at Austin)
Robotic IntelligenceReinforcement LearningContrastive LearningImageBenchmark
🎯 What it does: An offline goal-conditioned reinforcement learning method named SMORe is proposed, which directly learns the target achievement policy from offline data using a discriminator-free mixed dominance matching framework.
Score Regularized Policy Optimization through Diffusion Behavior
Huayu Chen (Tsinghua University), Jun Zhu (Tsinghua University)
OptimizationReinforcement LearningDiffusion modelTabularBenchmark
🎯 What it does: A novel offline reinforcement learning algorithm named SRPO (Score Regularized Policy Optimization) is proposed, which utilizes the score function of a pre-trained diffusion behavior model to regularize the policy at the gradient level, enabling deterministic policy extraction without diffusion sampling.
Score-based generative models break the curse of dimensionality in learning a family of sub-Gaussian distributions
Frank Cole (University of Minnesota), Yulong Lu (University of Minnesota)
GenerationData SynthesisScore-based Model
🎯 What it does: This paper studies the approximation and generalization performance of the Score Generative Model (SGM) when learning a family of sub-Gaussian probability distributions. It proposes a complexity measure based on the relative density against the standard Gaussian measure and provides an upper bound on the sample complexity for the curse of dimensionality.
SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis
Dustin Podell, Robin Rombach
GenerationData SynthesisTransformerDiffusion modelImageStochastic Differential Equation
🎯 What it does: We propose and train Stable Diffusion XL (SDXL), a latent diffusion model aimed at generating high-resolution images from text, equipped with a larger UNet, dual text encoders, and fine-grained control;
SE(3)-Stochastic Flow Matching for Protein Backbone Generation
Joey Bose (McGill University), Alexander Tong (Université de Montréal)
Protein Structure PredictionFlow-based ModelBiomedical DataStochastic Differential Equation
🎯 What it does: The FoldFlow series model is proposed, utilizing flow matching to learn continuous dynamics on SE(3), generating protein backbone structures from prior distributions, covering high-quality, designable, and diverse samples of 300 amino acids.
SEA: Sparse Linear Attention with Estimated Attention Mask
Heejun Lee (Korea Advanced Institute of Science and Technology), Sung Ju Hwang (Korea Advanced Institute of Science and Technology)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Using the existing pre-trained Transformers (BERT, OPT, etc.) with full quadratic complexity attention, we first estimate a compressed size attention matrix using Kernel-Based linear attention, and then generate a sparse attention mask through grouped Top-k selection, achieving O(T) inference complexity; at the same time, knowledge distillation is used to transfer the complete attention matrix of the teacher model to the sparse model.
SEABO: A Simple Search-Based Method for Offline Imitation Learning
Jiafei Lyu (Tsinghua University), Zongqing Lu (Peking University)
Robotic IntelligenceReinforcement LearningTabular
🎯 What it does: A search-based offline imitation learning method called SEABO is proposed, which automatically annotates rewards using expert demonstrations and unlabeled data, and integrates with offline RL algorithms.
SEAL: A Framework for Systematic Evaluation of Real-World Super-Resolution
Wenlong Zhang (Shanghai Jiao Tong University), Chao Dong (Shenzhen Institute of Advanced Technology)
RestorationSuper ResolutionGenerative Adversarial NetworkImageBenchmark
🎯 What it does: The SEAL framework is proposed, which constructs a representative degradation set using clustering methods to systematically evaluate real super-resolution methods.
Searching for High-Value Molecules Using Reinforcement Learning and Transformers
Raj Ghugare (Universite de Montreal), Glen Berseth (Universite de Montreal)
Drug DiscoveryRecurrent Neural NetworkTransformerReinforcement LearningText
🎯 What it does: This paper presents ChemRLformer, a Transformer-based text reinforcement learning framework for generating molecules with high-value attributes.