ICLR 2025 Papers — Page 29
International Conference on Learning Representations · 3704 papers
Revealing and Reducing Gender Biases in Vision and Language Assistants (VLAs)
Leander Girrbach (Technical University of Munich), Zeynep Akata (Technical University of Munich)
Prompt EngineeringImageMultimodality
🎯 What it does: Evaluated the performance of 22 open-source visual language assistants in terms of gender bias and attempted various debiasing techniques.
Revealing the 3D Cosmic Web through Gravitationally Constrained Neural Fields
Brandon Zhao (California Institute of Technology), Katherine Bouman
OptimizationNeural Radiance FieldPoint CloudPhysics Related
🎯 What it does: This paper proposes a method for three-dimensional dark matter reconstruction from weak gravitational lensing observations using a differentiable gravitational constraint neural field.
RevisEval: Improving LLM-as-a-Judge via Response-Adapted References
Qiyuan Zhang (City University of Hong Kong), Chen Ma (City University of Hong Kong)
GenerationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes the REVISEVAL evaluation paradigm, which enhances the reliability of LLM-as-a-Judge and traditional reference-based evaluation methods by utilizing the revision capabilities of large language models to generate 'response-adaptive references' that are highly relevant to the responses being evaluated.
Revisit Large-Scale Image-Caption Data in Pre-training Multimodal Foundation Models
Zhengfeng Lai (Apple), Yinfei Yang (Apple)
GenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelImageTextMultimodality
🎯 What it does: A controllable and scalable image description pipeline is proposed, generating various formats of synthetic subtitles (SSC, DSC, DSC+, AFC), and systematically studying their interaction with the original AltText and its impact on CLIP, cross-modal LLMs, and diffusion model pre-training.
Revisit Micro-batch Clipping: Adaptive Data Pruning via Gradient Manipulation
Lun Wang (Google)
ClassificationRecognitionOptimizationTransformerSupervised Fine-TuningImageTextAudio
🎯 What it does: This paper proposes and deeply studies adaptive micro-batch clipping as a data pruning method, explaining its mechanism in improving model convergence speed and performance.
Revisit the Open Nature of Open Vocabulary Semantic Segmentation
Qiming Huang (University of Birmingham), Jianbo Jiao (University of Birmingham)
SegmentationImage
🎯 What it does: A mask-matching-based evaluation protocol is proposed for open vocabulary semantic segmentation (OVS), revealing ambiguous labels in the dataset by constructing a fuzzy vocabulary graph, and further enhancing model generalization by randomly dropping non-target vocabulary.
Revisiting a Design Choice in Gradient Temporal Difference Learning
Xiaochi Qian (University of Oxford), Shangtong Zhang (University of Virginia)
Reinforcement Learning
🎯 What it does: The A^⊤ t TD algorithm is proposed, which improves GTD by reusing the A^⊤ t TD idea, solving the fatal triangle problem of off-policy and function approximation with only a single set of parameters and a single learning rate.
Revisiting Convolution Architecture in the Realm of DNA Foundation Models
Yu Bo (Zhejiang University), Chunhua Shen (Zhejiang University of Technology)
ClassificationConvolutional Neural NetworkBiomedical DataBenchmark
🎯 What it does: This paper proposes a DNA foundational model called ConvNova based on convolutional neural networks, achieving optimal or near-optimal performance on multiple DNA prediction benchmarks, demonstrating the competitiveness of CNNs in DNA modeling.
Revisiting In-context Learning Inference Circuit in Large Language Models
Hakaze Cho (Japan Advanced Institute of Science and Technology), Naoya Inoue (Japan Advanced Institute of Science and Technology)
ClassificationTransformerLarge Language ModelText
🎯 What it does: This study proposes and validates a three-step reasoning circuit for context learning in large language models (input encoding → semantic merging → feature retrieval and copying), demonstrating its existence and dominance through fine-grained measurements and ablation experiments, revealing bypass mechanisms and explanations for observed phenomena.
Revisiting Large-Scale Non-convex Distributionally Robust Optimization
Qi Zhang (Arizona State University), Shaofeng Zou (Arizona State University)
OptimizationTabular
🎯 What it does: This paper studies distributionally robust optimization (DRO) under non-convex smooth loss functions, proposing a more precise partial generalization smoothness and partial affine variance noise theory. Based on this, it designs the Dual Stochastic Gradient Descent with Pruning (D‑SGD‑C) and Dual Spider Pruning (D‑Spider‑C) algorithms.
Revisiting Mode Connectivity in Neural Networks with Bezier Surface
Jie Ren (Illinois Institute of Technology), Ren Wang (IBM Research)
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes the use of nonlinear Bézier surfaces to achieve neural network mode connectivity, constructing a multi-dimensional low-loss high-precision parameter space.
REVISITING MULTI-PERMUTATION EQUIVARIANCE THROUGH THE LENS OF IRREDUCIBLE REPRESENTATIONS
Yonatan Sverdlov (Technion Israel Institute of Technology), Nadav Dym (Technion Israel Institute of Technology)
Anomaly DetectionImagePoint CloudGraph
🎯 What it does: This paper uses irreducible representations and Schur's lemma to re-derive and extend the complete characterization of linear equivariant layers under permutation groups and their wreath products, covering DeepSets, 2-IGN, DWS networks, and equivariant layers for unaligned symmetric sets.
Revisiting Nearest Neighbor for Tabular Data: A Deep Tabular Baseline Two Decades Later
Han-Jia Ye (Nanjing University), Wei-Lun Chao (Ohio State University)
ClassificationOptimizationTabular
🎯 What it does: Improved the classic nearest neighbor method NCA by incorporating deep learning techniques, constructing MODERNNCA as a powerful deep learning baseline for tabular data.
Revisiting Prefix-tuning: Statistical Benefits of Reparameterization among Prompts
Minh Le (Qualcomm AI Research), Nhat Ho (University of Texas at Austin)
TransformerPrompt EngineeringMixture of ExpertsImageText
🎯 What it does: This study investigates and proves the theoretical foundation of the reparameterization strategy of prefix-tuning, revealing its implicit key/value sharing structure, and explains its improvement in sample efficiency through the MoE framework.
Revisiting Random Walks for Learning on Graphs
Jinwoo Kim (Korea Advanced Institute of Science and Technology), Seunghoon Hong (Korea Advanced Institute of Science and Technology)
Graph Neural NetworkTransformerLarge Language ModelTextGraph
🎯 What it does: This paper proposes and systematically analyzes Random Walk Neural Networks (RWNNs), which generate readable sequences through random walks and are read by deep networks to achieve graph/vertex-level predictions.
Revisiting Source-Free Domain Adaptation: a New Perspective via Uncertainty Control
Gezheng Xu (University of Western Ontario), Grace Yi
Domain AdaptationContrastive LearningImage
🎯 What it does: The UCon-SFDA method is proposed, which enhances the robustness of model self-supervised learning in source-free domain adaptation tasks through theoretical analysis of negative sample sampling errors and positive sample prediction uncertainties, designing two uncertainty control strategies: negative sample dispersion control and partial label learning for positive samples.
Revisiting text-to-image evaluation with Gecko: on metrics, prompts, and human rating
Olivia Wiles (Google DeepMind), Aida Nematzadeh (Google DeepMind)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: A large-scale, fine-grained text-to-image (T2I) evaluation suite called Gecko has been constructed, covering 2,000 diverse prompts, four annotation templates, four mainstream models, and collecting approximately 100,000 human ratings.
Revisiting Zeroth-Order Optimization: Minimum-Variance Two-Point Estimators and Directionally Aligned Perturbations
Shaocong Ma (University of Maryland), Heng Huang (University of Maryland)
OptimizationLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper studies two-point zero-order gradient estimators, proposes and analyzes a perturbation distribution that achieves minimum variance as the perturbation step size approaches zero, and designs a disturbance scheme based on gradient direction alignment (DAP).
Revolutionizing EMCCD Denoising through a Novel Physics-Based Learning Framework for Noise Modeling
Haiyang Jiang (University of Tokyo), Yinqiang Zheng (University of Tokyo)
RestorationTransformerSupervised Fine-TuningImage
🎯 What it does: A physical noise model for EMCCD cameras was designed and implemented, and this model was used to synthesize training data to train Uformer, completing the EMCCD image denoising task in low-light environments.
REvolve: Reward Evolution with Large Language Models using Human Feedback
RISHI HAZRA, Pedro Zuidberg Dos Martires (Orebro University)
Autonomous DrivingOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVideo
🎯 What it does: A reward function design framework REvolve based on evolutionary algorithms and large language models is proposed, which utilizes human feedback to guide the evolution of the reward function and generates interpretable Python code rewards.
Reward Dimension Reduction for Scalable Multi-Objective Reinforcement Learning
Giseung Park (Korea Advanced Institute of Science and Technology), Youngchul Sung (Korea Advanced Institute of Science and Technology)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes an online reward dimension reduction method that can reduce the reward space in multi-objective reinforcement learning while maintaining Pareto optimality.
Reward Learning from Multiple Feedback Types
Yannick Metz (University of Konstanz), Mennatallah El-Assady (ETH Zurich)
Autonomous DrivingReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: This study investigates various types of human feedback (evaluation, comparison, demonstration, correction, description, and descriptive preference) for reward learning and implements a unified simulation generation framework.
Rewarding Progress: Scaling Automated Process Verifiers for LLM Reasoning
Amrith Setlur (Carnegie Mellon University), Aviral Kumar (Google DeepMind)
Computational EfficiencyReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a Process Advantage Validator (PAV) that utilizes advantage as a process reward, enhancing the reasoning performance of large language models through this reward.
RFMamba: Frequency-Aware State Space Model for RF-Based Human-Centric Perception
Rui Zhang (University of Science and Technology of China), Yan Chen (University of Science and Technology of China)
Pose EstimationConvolutional Neural NetworkSupervised Fine-TuningMultimodality
🎯 What it does: A frequency-aware state space model RFMamba is proposed for through-wall human perception using radio frequency signals, supporting tasks such as posture estimation, action recognition, and identity recognition.
RFWave: Multi-band Rectified Flow for Audio Waveform Reconstruction
Peng Liu (Transsion), Zhiyong Wu (Tsinghua University)
RestorationGenerationData SynthesisFlow-based ModelRectified FlowGenerative Adversarial NetworkAudio
🎯 What it does: Proposes RFWave, a multi-band Rectified Flow model for high-quality audio waveform reconstruction.
RGB-Event ISP: The Dataset and Benchmark
Yunfan LU, Hui Xiong (Hong Kong University of Science and Technology)
Data SynthesisConvolutional Neural NetworkTransformerImageBenchmark
🎯 What it does: This paper first constructs a pixel-level paired dataset of events and RAW images, and evaluates and validates the event-guided image signal processing (ISP) pipeline based on this dataset.
Risk-Sensitive Diffusion: Robustly Optimizing Diffusion Models with Noisy Samples
Yangming Li (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
GenerationOptimizationDiffusion modelScore-based ModelTabularTime SeriesBiomedical DataElectronic Health RecordsStochastic Differential Equation
🎯 What it does: This paper proposes a robust method for training diffusion models on non-image data in the presence of noise, which pairs each noisy sample with its risk vector and introduces a risk-sensitive SDE. By incorporating risk information into the diffusion process, it mitigates the negative impact of noise, achieving robust optimization and sampling of the diffusion model.
Risk-Sensitive Variational Actor-Critic: A Model-Based Approach
Alonso Granados (University of Arizona), Jason Pacheco (University of Arizona)
Reinforcement Learning
🎯 What it does: A risk-sensitive actor-critic algorithm based on variational inference, rsV AC, has been developed, capable of simultaneously learning risk-averse and risk-seeking strategies in environments with unknown transition dynamics and stochastic rewards.
RM-Bench: Benchmarking Reward Models of Language Models with Subtlety and Style
Yantao Liu (Fudan University), Juanzi Li (Tsinghua University)
Large Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: The RM-BENCH benchmark is proposed to evaluate the ability of reward models to distinguish subtle content differences and resist style bias, demonstrating a high correlation with the performance of policy models.
RMB: Comprehensively benchmarking reward models in LLM alignment
Enyu Zhou (Fudan University), Xuanjing Huang (Fudan University)
TransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: The RMB benchmark is proposed to evaluate the effectiveness of reward models in LLM alignment, covering 49 fine-grained real-world scenarios, and providing pairwise and Best-of-N evaluations.
RMP-SAM: Towards Real-Time Multi-Purpose Segment Anything
Shilin Xu (Peking University), Ming-Hsuan Yang (Google Research)
SegmentationConvolutional Neural NetworkImageVideo
🎯 What it does: This paper proposes a real-time multipurpose segmentation model RMP-SAM, capable of performing image panoptic segmentation, video instance segmentation, and interactive segmentation all at once;
RNNs are not Transformers (Yet): The Key Bottleneck on In-Context Retrieval
Kaiyue Wen (Stanford University), Kaifeng Lyu (University of California)
RetrievalRecurrent Neural NetworkTransformerTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This study investigates the gap in representational power between Recurrent Neural Networks (RNNs) and Transformers in algorithmic problems, and explores the impact of Chain of Thought (CoT) and Retrieval-Augmented Generation (RAG) on RNNs.
Robotouille: An Asynchronous Planning Benchmark for LLM Agents
Gonzalo Gonzalez-Pumariega (Cornell University), Sanjiban Choudhury (Cornell University)
Robotic IntelligenceTransformerLarge Language ModelAgentic AITextBenchmarkChain-of-Thought
🎯 What it does: A kitchen cooking simulation benchmark called ROBOTOUILLE is proposed to evaluate the asynchronous planning capabilities of large language models, along with the construction of synchronous, asynchronous, and multi-agent datasets.
Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets
Guangqi Jiang (University of California), Huazhe Xu (Tsinghua University)
Representation LearningRobotic IntelligenceConvolutional Neural NetworkContrastive LearningImageVideo
🎯 What it does: This paper proposes a metric for evaluating the effectiveness of machine vision representations—manipulation centrality—and designs the MCR pre-training framework based on this to learn robot vision representations with higher manipulation centrality.
RobuRCDet: Enhancing Robustness of Radar-Camera Fusion in Bird's Eye View for 3D Object Detection
Jingtong Yue (Peking University), Ming-Hsuan Yang (UC Merced)
Object DetectionAutonomous DrivingGaussian SplattingMultimodalityPoint Cloud
🎯 What it does: A robust radar-camera fusion 3D object detection framework, RobuRCDet, is proposed, focusing on improving detection performance in harsh environments and noisy conditions.
Robust Barycenter Estimation using Semi-Unbalanced Neural Optimal Transport
Milena Gazdieva (Skolkovo Institute of Science and Technology), Alexander Korotin (Skolkovo Institute of Science and Technology)
Anomaly DetectionOptimizationGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a robust barycenter estimation method based on Semi-Unbalanced Optimal Transport (SUOT), which maintains good performance even when the data contains outliers or class imbalance.
Robust Conformal Prediction with a Single Binary Certificate
Soroush H. Zargarbashi (CISPA Helmholtz Center for Information Security), Aleksandar Bojchevski (University of Cologne)
ClassificationObject DetectionGraph Neural NetworkSupervised Fine-TuningImageGraph
🎯 What it does: BinCP is proposed, a robust consistency prediction method based on random smoothing, which guarantees coverage probability under worst-case inputs by binarizing the smoothed scores and using only a single binary certificate, while significantly reducing the amount of Monte Carlo sampling.
Robust Feature Learning for Multi-Index Models in High Dimensions
Alireza Mousavi-Hosseini (University of Toronto), Murat A Erdogdu
OptimizationRepresentation LearningAdversarial AttackTabular
🎯 What it does: This study investigates adversarial robust feature learning in high-dimensional environments using multi-exponential models, and proves that under ℓ₂ constrained attacks, performing standard feature learning first and then robust training on the second layer can achieve the Bayes optimal robust risk.
Robust Function-Calling for On-Device Language Model via Function Masking
Qiqiang Lin (OPPO Research Institute), Weinan Zhang (Shanghai Jiao Tong University)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: To address the robustness issue of language models when executing external function calls, a series of lightweight models called Hammer has been proposed, specifically designed for on-device function calls.
Robust Gymnasium: A Unified Modular Benchmark for Robust Reinforcement Learning
Shangding Gu (University of California), Costas Spanos (California Institute of Technology)
Robotic IntelligenceLarge Language ModelReinforcement LearningBenchmark
🎯 What it does: This paper proposes Robust-Gymnasium—a unified, modular benchmark for robust reinforcement learning;
Robust LLM safeguarding via refusal feature adversarial training
Lei Yu (University of Toronto), Nicola Cancedda (Meta)
Computational EfficiencyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper presents a novel adversarial training algorithm called ReFAT, based on Reject Feature Ablation (RF Ablation), through an explanatory analysis of the attack resistance mechanisms of large language models (LLMs). It significantly enhances the robustness of LLMs against various attacks while maintaining general capabilities.
Robust Representation Consistency Model via Contrastive Denoising
Jiachen Lei (Zhejiang University), Anima Anandkumar (Caltech)
ClassificationComputational EfficiencyRepresentation LearningTransformerDiffusion modelContrastive LearningImageOrdinary Differential Equation
🎯 What it does: A robust model for one-time denoising and classification is designed by aligning the representation of noisy and clean samples on the probabilistic flow trajectory.
Robust Root Cause Diagnosis using In-Distribution Interventions
Lokesh Nagalapatti (Indian Institute of Technology Bombay), Amit Sharma (Microsoft Research India)
Anomaly DetectionTabular
🎯 What it does: The In-Distribution Interventions (IDI) method is proposed, which uses interventions conducted only within the training distribution to evaluate the fix conditions of root cause nodes, thereby achieving root cause diagnosis for cloud services and industrial system anomalies.
Robust Simulation-Based Inference under Missing Data via Neural Processes
Yogesh Verma (Aalto University), Vikas Garg (Aalto University)
Meta LearningDrug DiscoveryAuto EncoderBiomedical Data
🎯 What it does: In the scenario of simulation-based inference (SBI) with missing data, a model called RISE is proposed to jointly learn missing value imputation and posterior estimation, achieving robust inference against varying missing rates.
Robust System Identification: Finite-sample Guarantees and Connection to Regularization
Hyuk Park (University of Illinois Urbana-Champaign), Yingying Li (University of Illinois Urbana-Champaign)
OptimizationReinforcement LearningTime Series
🎯 What it does: A robust least squares framework is studied for learning the parameters of nonlinear dynamical systems using a single trajectory.
Robust Transfer of Safety-Constrained Reinforcement Learning Agents
Markel Zubia (Ruhr University Bochum), Nils Jansen (Radboud University Nijmegen)
Safty and PrivacyReinforcement Learning
🎯 What it does: This paper studies a method for training safe agents in the source environment and robustly transferring them to the target environment, ensuring that safety constraints are met in the target environment.
Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances
Shilin Lu (Nanyang Technological University), Adams Wai-Kin Kong (Nanyang Technological University)
RestorationGenerationGenerative Adversarial NetworkImageVideoBenchmark
🎯 What it does: A new benchmark W-Bench is proposed for four types of editing tasks: image reconstruction, global editing, local editing, and image-to-video generation, and a robust watermark model VINE is designed based on spectral analysis and generative priors.
Robust Weight Initialization for Tanh Neural Networks with Fixed Point Analysis
Hyun woo Lee, Hyunju Kim (Korea Institute of Energy Technology)
OptimizationConvolutional Neural NetworkImagePhysics Related
🎯 What it does: A weight initialization method based on fixed point analysis of tanh(ax) is proposed, aimed at keeping the activation values of deep tanh networks from vanishing or saturating, without the need for batch normalization or layer normalization.
Robust-PIFu: Robust Pixel-aligned Implicit Function for 3D Human Digitalization from a Single Image
Kennard Chan, Weisi Lin (Nanyang Technological University)
SegmentationGenerationConvolutional Neural NetworkDiffusion modelImageMesh
🎯 What it does: Using a large-scale pre-trained implicit diffusion model, we propose Robust-PIFu, which addresses external occlusions (such as mutual occlusions between multiple people) and internal occlusions (such as self-occlusions or non-frontal poses) in human images from a single view, generating complete and structurally accurate clothing human meshes.
RobustKV: Defending Large Language Models against Jailbreak Attacks via KV Eviction
Tanqiu Jiang (Stony Brook University), Ting Wang (Stony Brook University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes RobustKV, a method to defend against jailbreak attacks by removing low-importance tokens from the KV cache;
Robustness Auditing for Linear Regression: To Singularity and Beyond
Ittai Rubinstein (Massachusetts Institute of Technology), Samuel B. Hopkins (Massachusetts Institute of Technology)
TabularFinance Related
🎯 What it does: Two algorithms, ACRE and OHARE, are proposed to provide verifiable upper and lower bounds for ordinary least squares (OLS) regression after deleting several samples, thereby achieving robustness auditing.
Robustness Inspired Graph Backdoor Defense
Zhiwei Zhang (Pennsylvania State University), Suhang Wang (Pennsylvania State University)
Anomaly DetectionOptimizationGraph Neural NetworkGraph
🎯 What it does: A backdoor defense framework for graph neural networks based on random edge deletion and robust training, called RIGBD, is proposed, which can identify and defend against attacks under various backdoor triggers.
Robustness of Quantum Algorithms for Nonconvex Optimization
Weiyuan Gong (Harvard University), Tongyang Li (Peking University)
Optimization
🎯 What it does: This paper systematically studies the query complexity of quantum algorithms for finding ε-second-order stationary points (ε-SOSP) of non-convex functions under noisy zero-order or first-order quantum oracles, providing upper and lower bounds for different noise levels, and making theoretical comparisons with classical algorithms.
Robustness Reprogramming for Representation Learning
Zhichao Hou (North Carolina State University), Xiaorui Liu (North Carolina State University)
Representation LearningAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a reprogramming method that does not change the parameters of the trained deep model to enhance its robustness against adversarial or noise perturbations.
RocketEval: Efficient automated LLM evaluation via grading checklist
Tianjun Wei (City University of Hong Kong), Jianghong Ma (Harbin Institute of Technology Shenzhen)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: RocketEval is proposed, an automated evaluation framework that utilizes lightweight LLMs as judges, transforming assessments into multi-dimensional question-and-answer formats using instance-specific checklists.
Rodimus*: Breaking the Accuracy-Efficiency Trade-Off with Efficient Attentions
Zhihao He (Shanghai Jiao Tong University), Weiyao Lin (Ant Group)
GenerationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper introduces Rodimus (a pure recursive linear attention model) and Rodimus+ (a hybrid model combining sliding window shared key attention), achieving O(1) word generation complexity and outperforming existing LLMs on multiple benchmarks.
Root Cause Analysis of Anomalies in Multivariate Time Series through Granger Causal Discovery
Xiao Han (Utah State University), Shuhan Yuan (Utah State University)
Anomaly DetectionAuto EncoderTime Series
🎯 What it does: This paper proposes an end-to-end framework named AERCA, which combines Granger causality discovery with root cause analysis by inferring the distribution of exogenous variables through an encoder and detecting anomalous root causes during deployment.
Rotated Runtime Smooth: Training-Free Activation Smoother for accurate INT4 inference
Ke Yi (South China University of Technology), Jingren Zhou (Alibaba Group)
TransformerLarge Language ModelText
🎯 What it does: A method for smoothing activations of large language models at runtime (Rotated Runtime Smooth) is proposed, achieving high precision and low latency for INT4 inference;
Round and Round We Go! What makes Rotary Positional Encodings useful?
Federico Barbero (University of Oxford), Petar Veličković (Google DeepMind)
TransformerLarge Language ModelText
🎯 What it does: Analyze and experimentally verify the true mechanism of RoPE on Transformer attention, exploring the role of different frequencies and their contributions to positional and semantic attention;
ROUTE: Robust Multitask Tuning and Collaboration for Text-to-SQL
Yang Qin (Sichuan University), Jieping Ye (Independent Researcher)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: The ROUTE framework is proposed, which first conducts multi-task supervised fine-tuning (including Text2SQL, Schema Linking, Noise Correction, and Continuation Writing), and then uses Multi-task Collaborative Prompting (MCP) to gradually generate and correct SQL, significantly improving the performance of open-source LLMs on the Text2SQL task.
RouteLLM: Learning to Route LLMs from Preference Data
Isaac Ong (University of California Berkeley), Ion Stoica (University of California Berkeley)
Recommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A routing framework for LLM based on human preference data is proposed, which trains the router to dynamically select strong or weak models during inference to balance quality and cost.
Routing Experts: Learning to Route Dynamic Experts in Existing Multi-modal Large Language Models
Qiong Wu (Xiamen University), Rongrong Ji (Xiamen University)
Computational EfficiencyTransformerLarge Language ModelMixture of ExpertsVision Language ModelMultimodality
🎯 What it does: This paper proposes the RoE (Routing Experts) method, treating the pre-trained multimodal large language model (MLLM) as a sparse expert network, achieving example-dependent computational paths through dynamic hierarchical routing, significantly improving inference efficiency.
RRM: Robust Reward Model Training Mitigates Reward Hacking
Tianqi Liu (Google DeepMind), Mohammad Saleh
Reinforcement Learning from Human FeedbackReinforcement LearningTextBenchmark
🎯 What it does: A robust reward model (RRM) is proposed, which eliminates context-independent artifacts (such as length, format, etc.) in responses through a causal framework and cross-example random permutation data augmentation method, thereby enhancing the robustness of the reward model and RLHF strategy.
RTDiff: Reverse Trajectory Synthesis via Diffusion for Offline Reinforcement Learning
Qianlan Yang (University of Illinois Urbana Champaign), Yu-Xiong Wang (University of Illinois Urbana Champaign)
Data SynthesisReinforcement LearningDiffusion modelTabular
🎯 What it does: This paper proposes RTDiff, an offline reinforcement learning data augmentation method that utilizes diffusion models for reverse trajectory synthesis.
RTop-K: Ultra-Fast Row-Wise Top-K Selection for Neural Network Acceleration on GPUs
Xi Xie (University of Connecticut), Caiwen Ding (University of Minnesota)
OptimizationComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: A row-level Top-k selection algorithm specifically designed for GPUs, RTop-K, is proposed, which implements efficient parallel row-level Top-k selection using binary search and supports an early stopping mechanism.
RuAG: Learned-rule-augmented Generation for Large Language Models
Yudi Zhang (Eindhoven University of Technology), Qi Zhang (Microsoft)
GenerationAnomaly DetectionTransformerLarge Language ModelPrompt EngineeringTextTime SeriesRetrieval-Augmented Generation
🎯 What it does: This paper proposes a method for automatically generating and refining first-order logic rules (RuAG) using large language models (LLMs). It compresses offline data into interpretable rules and then injects these rules into LLM prompts in natural language to enhance the reasoning and generation performance of LLMs across multiple tasks.
S4M: S4 for multivariate time series forecasting with Missing values
Peng Jing, Xiaoxiao Li (University of British Columbia)
Recurrent Neural NetworkTransformerTime Series
🎯 What it does: An end-to-end temporal prediction framework S4M is proposed, which can directly handle block missing values in multivariate time series.
SafeDiffuser: Safe Planning with Diffusion Probabilistic Models
Wei Xiao (Massachusetts Institute of Technology), Daniela Rus (Massachusetts Institute of Technology)
OptimizationSafty and PrivacyRobotic IntelligenceReinforcement LearningDiffusion modelTabular
🎯 What it does: In safety-critical trajectory planning, the SafeDiffuser method is proposed, embedding control barrier functions into the denoising process of diffusion models to achieve finite-time diffusion invariance, thereby providing safety guarantees when generating trajectories.
Safety Alignment Should be Made More Than Just a Few Tokens Deep
Xiangyu Qi (Princeton University), Peter Henderson (Princeton University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study reveals that the current safety alignment of LLMs primarily relies on the first few tokens, termed shallow safety alignment, and discusses the various attack vulnerabilities it leads to; it proposes two countermeasures: deep safety alignment (data augmentation) and token-level constraint fine-tuning.
Safety Layers in Aligned Large Language Models: The Key to LLM Security
Shen Li (University of Science and Technology of China), Yaliang Li (University of Science and Technology of China)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper discovers and locates the 'safety layer', which is crucial for safety alignment, for the first time through hierarchical vector similarity, parameter scaling, and the phenomenon of excessive rejection. Based on this, it proposes freezing the safety layer for partial parameter fine-tuning (SPPFT), effectively preventing safety degradation caused by fine-tuning attacks.
Safety Representations for Safer Policy Learning
Kaustubh Mani (Mila), Liam Paull
Safty and PrivacyReinforcement Learning
🎯 What it does: The SRPL framework is proposed, which learns state-conditioned safety representations in reinforcement learning and combines it with existing algorithms to achieve safer and more efficient learning.
Safety-Prioritizing Curricula for Constrained Reinforcement Learning
Cevahir Koprulu (University of Texas at Austin), ufuk topcu
Safty and PrivacyReinforcement Learning
🎯 What it does: This paper proposes a safety-first automatic curriculum generation method (SCG) for multi-task learning in Constrained Reinforcement Learning (CRL), which can reduce constraint violations during training and quickly converge to the optimal policy.
SafeWatch: An Efficient Safety-Policy Following Video Guardrail Model with Transparent Explanations
Zhaorun Chen (University of Chicago), Bo Li (University of Chicago)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningVideoMultimodalityBenchmark
🎯 What it does: This paper proposes SAFEWATCH, a video guardian model based on a multimodal large language model, capable of generating multi-label alerts according to custom safety policies and providing interpretable explanations.
SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation
Jaehong Yoon (University of North Carolina at Chapel Hill), Mohit Bansal (University of North Carolina at Chapel Hill)
GenerationDiffusion modelImageVideoText
🎯 What it does: This paper proposes a training-free, adaptive safe text-image/video generation method called SAFREE, which can mask user-defined harmful concepts while maintaining generation quality without modifying model weights.
SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration
Jintao Zhang (Tsinghua University), Jianfei Chen (Tsinghua University)
Computational EfficiencyTransformerLarge Language ModelImageVideoTextMultimodality
🎯 What it does: This paper proposes SageAttention, an attention acceleration method based on INT8 quantization, achieving efficient inference through smoothing the K matrix, using FP16 accumulators, and adaptive quantization.
SAGEPhos: Sage Bio-Coupled and Augmented Fusion for Phosphorylation Site Detection
Jingjie Zhang (Chinese University of Hong Kong), Chunbin Gu (Chinese University of Hong Kong)
Drug DiscoveryGraph Neural NetworkTransformerBiomedical Data
🎯 What it does: A kinase-substrate dual-modal fusion framework SAGEPhos based on structural information is proposed for phosphorylation site prediction.
Sail into the Headwind: Alignment via Robust Rewards and Dynamic Labels against Reward Hacking
Paria Rashidinejad (Fundamental AI Research), Yuandong Tian (Fundamental AI Research)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This study addresses the reward hacking problem in offline preference optimization and proposes the POWER-DL method, which combines weighted entropy robust rewards with dynamic labeling to enhance the alignment performance of LLMs.
SaLoRA: Safety-Alignment Preserved Low-Rank Adaptation
Mingjie Li (CISPA Helmholtz Center for Information Security), Yisen Wang (Peking University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes SaLoRA, a fine-tuning method that maintains the safety alignment of large language models (LLMs) while performing parameter-efficient fine-tuning using LoRA.
Salvage: Shapley-distribution Approximation Learning Via Attribution Guided Exploration for Explainable Image Classification
Mehdi Naouar (University of Freiburg), Maria Kalweit (University of Freiburg)
ClassificationExplainability and InterpretabilityTransformerImage
🎯 What it does: Developed an interpretable image classification method called Salvage based on Shapley distribution approximation, which trains an interpreter to learn the prediction distribution of the classifier under masked images and extracts Shapley values to generate feature importance maps.
SAM 2: Segment Anything in Images and Videos
Nikhila Ravi (Meta), Christoph Feichtenhofer (Meta)
Object DetectionSegmentationTransformerPrompt EngineeringImageVideo
🎯 What it does: We propose SAM 2, a unified Promptable visual segmentation model that can perform interactive segmentation on single-frame images and real-time streaming processing in videos, supporting point, box, and mask prompts.
SAM-CP: Marrying SAM with Composable Prompts for Versatile Segmentation
Pengfei Chen (University of Chinese Academy of Sciences), Qi Tian (Huawei Inc.)
Object DetectionSegmentationConvolutional Neural NetworkPrompt EngineeringImage
🎯 What it does: This paper proposes SAM-CP, which achieves semantic, instance, and panoptic segmentation by applying two types of composable prompts on patches generated by Segment Anything (SAM) (Prompt I: determine whether a patch belongs to a certain text label; Prompt II: determine whether patches under the same text label belong to the same instance).
Samba: Simple Hybrid State Space Models for Efficient Unlimited Context Language Modeling
Liliang Ren (Microsoft), Weizhu Chen (Microsoft)
GenerationRetrievalComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes SAMBA, a linear-time long sequence language model that interleaves the Mamba state space model with sliding window attention (SWA) at a hierarchical level.
Samba: Synchronized Set-of-Sequences Modeling for Multiple Object Tracking
Mattia Segu (ETH Zurich), Bernt Schiele (Max Planck Institute for Informatics)
Object TrackingTransformerSupervised Fine-TuningVideo
🎯 What it does: Proposes SambaMOTR, which utilizes the synchronous sequence model Samba to achieve multi-object tracking;
SaMer: A Scenario-aware Multi-dimensional Evaluator for Large Language Models
Kehua Feng (Zhejiang University), Huajun Chen (Zhejiang University)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A LLM evaluator called SaMer has been developed, which can automatically identify and assess multi-dimensional quality based on query context, and has been validated on single scoring, comparative scoring, and fine-grained multi-dimensional evaluation tasks.
Sample then Identify: A General Framework for Risk Control and Assessment in Multimodal Large Language Models
Qingni Wang (University of Electronic Science and Technology of China), Feng Zheng (Southern University of Science and Technology)
TransformerLarge Language ModelVideoMultimodality
🎯 What it does: A general two-step framework TRON is proposed for risk control and assessment in the open VideoQA task of multimodal large language models.
SAMRefiner: Taming Segment Anything Model for Universal Mask Refinement
Yuqi Lin (Zhejiang University), Kaipeng Zhang (Shanghai AI Laboratory)
Object DetectionSegmentationImageVideoBenchmark
🎯 What it does: This paper proposes SAMRefiner, a general and efficient mask refinement framework that utilizes the SAM model and automatically generates prompts from rough masks through noise-robust multi-prompt strategies (distance-guided points, context elastic boxes, Gaussian masks) to improve mask quality. It also introduces a segmentation-then-merge (STM) process for semantic segmentation and enhances IoU prediction through LoRA adaptive steps.
SANA: Efficient High-Resolution Text-to-Image Synthesis with Linear Diffusion Transformers
Enze Xie (Tsinghua University), Song Han (NVIDIA)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelAuto EncoderImageText
🎯 What it does: SANA has proposed an efficient text-to-image generation framework that can quickly generate high-quality images at a resolution of 4096×4096 and can be deployed on a 16GB GPU laptop.
SANER: Annotation-free Societal Attribute Neutralizer for Debiasing CLIP
Yusuke Hirota (Osaka University), Ryo Hachiuma (NVIDIA)
GenerationRetrievalTransformerContrastive LearningImageText
🎯 What it does: A CLIP debiasing method named SANER is proposed, which utilizes text neutralization and an unannotated loss to debias text features, removing attribute information only when there is no attribute description while maintaining integrity when attributes are described.
SaRA: High-Efficient Diffusion Model Fine-tuning with Progressive Sparse Low-Rank Adaptation
Teng Hu (Shanghai Jiao Tong University), Lizhuang Ma (Shanghai Jiao Tong University)
GenerationComputational EfficiencySupervised Fine-TuningDiffusion modelImage
🎯 What it does: A parameter-efficient fine-tuning method named SaRA is designed, which identifies and utilizes the 'ineffective' parameters with the smallest absolute values in the pre-trained diffusion model to construct a sparse low-rank adapter for rapid transfer to downstream tasks and to enhance the performance of the main task.
Satisficing Regret Minimization in Bandits
Qing Feng (Cornell University), Ruihao Zhu (Cornell University)
OptimizationReinforcement Learning from Human FeedbackTabular
🎯 What it does: A general SELECT algorithm template is proposed for minimizing satisficing objectives in the multi-armed bandit (Bandit) problem; it achieves constant satisficing regret in realizable cases and maintains the same standard regret as learning oracles in non-realizable cases.
SAVA: Scalable Learning-Agnostic Data Valuation
Samuel Kessler (Microsoft), Vu Nguyen (Amazon)
Computational EfficiencyData-Centric LearningImage
🎯 What it does: A scalable data value assessment method called SAVA is proposed, which performs calculations based on hierarchical optimal transport (OT) at the batch level to address the memory bottleneck issue of the original LAVA on large-scale datasets.
SBSC: Step-by-Step Coding for Improving Mathematical Olympiad Performance
Kunal Singh (Fractal AI Research), Siva Kishore Gollapalli (Fractal AI Research)
AI Code AssistantTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: A multi-round code generation mathematical reasoning framework called SBSC is proposed, enabling LLMs to gradually decompose Olympiad-level problems and generate new programs at each step using the execution results of the previous step.
SC-OmniGS: Self-Calibrating Omnidirectional Gaussian Splatting
Huajian Huang (Hong Kong University of Science and Technology), Sai-Kit Yeung (Hong Kong University of Science and Technology)
GenerationData SynthesisOptimizationGaussian SplattingImage
🎯 What it does: A self-calibrating panoramic Gaussian scattering system SC-OmniGS is proposed for fast and high-quality 360° light field reconstruction.
Scalable and Certifiable Graph Unlearning: Overcoming the Approximation Error Barrier
Lu Yi (Renmin University of China), Zhewei Wei (Renmin University of China)
OptimizationComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: A scalable and verifiable graph unlearning method, ScaleGUN, is proposed, which can achieve efficient and verifiable graph neural network data deletion at the billion-edge level.
Scalable Bayesian Learning with posteriors
Samuel Duffield (Normal Computing), Daniel Simpson (Normal Computing)
Large Language ModelSupervised Fine-TuningText
🎯 What it does: This paper presents a scalable Bayesian learning library called posteriors based on PyTorch, demonstrating its practicality on large language models.
Scalable Benchmarking and Robust Learning for Noise-Free Ego-Motion and 3D Reconstruction from Noisy Video
Xiaohao Xu (University of Michigan), Xiaonan Huang (University of Michigan)
Data SynthesisPose EstimationDepth EstimationOptimizationGaussian SplattingSimultaneous Localization and MappingVideoBenchmark
🎯 What it does: This study proposes a scalable noise data synthesis pipeline, the Robust-Ego3D benchmark, and the corresponding CorrGS method for robust self-motion estimation and high-quality 3D reconstruction in noisy videos.
Scalable Decentralized Learning with Teleportation
Yuki Takezawa (Kyoto University), Sebastian U Stich
Federated LearningComputational EfficiencyHyperparameter SearchConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A decentralized learning method called TELEPORTATION is proposed, which activates only a subset of nodes in each round and performs gossip averaging among these nodes to avoid parameter drift, addressing the convergence rate decline caused by a large number of nodes.
Scalable Decision-Making in Stochastic Environments through Learned Temporal Abstraction
Baiting Luo (Vanderbilt University), Ayan Mukhopadhyay (Vanderbilt University)
Robotic IntelligenceTransformerReinforcement LearningAuto EncoderSequential
🎯 What it does: A discrete macro action planning framework named L-MAP is proposed, which utilizes VQ-VAE to learn state-conditioned macro actions and performs MCTS planning in a discrete latent space, addressing the discretization and planning bottlenecks in high-dimensional continuous stochastic environments.
Scalable Discrete Diffusion Samplers: Combinatorial Optimization and Statistical Physics
Sebastian Sanokowski (Johannes Kepler University Linz), Sebastian Lehner (Johannes Kepler University Linz)
OptimizationGraph Neural NetworkReinforcement LearningDiffusion modelGraphPhysics Related
🎯 What it does: Proposes a scalable discrete diffusion sampler (SDDS) to achieve higher step counts for discrete NPO and unbiased sampling;
Scalable Extraction of Training Data from Aligned, Production Language Models
Milad Nasr (Google DeepMind), Katherine Lee (Google DeepMind)
Adversarial AttackData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper studies the feasibility of training data leakage from large language models that have been aligned and deployed as chatbots, proposing two attack methods—divergence attack and finetuning attack—which successfully extracted thousands of training samples from closed-source models like ChatGPT and Gemini.