ICLR 2025 Papers — Page 27
International Conference on Learning Representations · 3704 papers
Progressive Compositionality in Text-to-Image Generative Models
Xu Han, Paul Pu Liang (Massachusetts Institute of Technology)
GenerationLarge Language ModelDiffusion modelContrastive LearningImageText
🎯 What it does: A text-to-image generation framework based on curriculum contrastive learning (EVOGEN) is proposed, enhancing the compositional ability of the diffusion model through a gradual training phase from simple to complex.
Progressive Compression with Universally Quantized Diffusion Models
Yibo Yang (University of California), Stephan Mandt (University of California)
CompressionDiffusion modelImage
🎯 What it does: This paper proposes an unconditional progressive compression framework based on diffusion models called Universally Quantized Diffusion Model (UQDM), which enables stepwise decoding from low bit rates to lossless compression.
Progressive distillation induces an implicit curriculum
Abhishek Panigrahi (Princeton University), Surbhi Goel (University of Pennsylvania)
OptimizationKnowledge DistillationTransformerSupervised Fine-TuningText
🎯 What it does: This paper studies progressive distillation, which accelerates the learning of student models through implicit curriculum provided by intermediate checkpoints. It demonstrates that on sparse stochastic tasks, probabilistic context-free grammars (PCFG), and real-world Wikipedia and book data, the student outperforms both one-shot distillation and direct training in terms of training steps and sample efficiency.
Progressive Mixed-Precision Decoding for Efficient LLM Inference
Hao Mark Chen, Stylianos Venieris
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper studies a stage-aware quantization strategy and Progressive Mixed Precision Decoding (PMPD) for efficient inference of large language models on mobile devices.
Progressive Parameter Efficient Transfer Learning for Semantic Segmentation
Nan Zhou (Beihang University), Di Huang (Beihang University)
SegmentationSupervised Fine-TuningImage
🎯 What it does: This paper proposes a staged parameter-efficient transfer learning framework called ProPETL, designed to quickly adapt large-scale pre-trained models to semantic segmentation tasks.
Progressive Token Length Scaling in Transformer Encoders for Efficient Universal Segmentation
Abhishek Aich (NEC Laboratories America), Manmohan Chandraker (University of California San Diego)
Object DetectionSegmentationComputational EfficiencyTransformerImage
🎯 What it does: PRO-SCALE is proposed, an efficient implementation of hierarchical token length extension for the Transformer encoder within the Mask2Former universal segmentation framework, and introduces a lightweight pixel embedding (LPE) module.
Projection Head is Secretly an Information Bottleneck
Zhuo Ouyang (Peking University), Yisen Wang (Peking University)
CompressionRepresentation LearningConvolutional Neural NetworkAuto EncoderContrastive LearningImage
🎯 What it does: This paper conducts an in-depth theoretical analysis of the projection head in contrastive learning from an information theory perspective. Based on the conclusion that the projection head should act as an information bottleneck, it further proposes two types of methods to improve the projection head: training regularization (mutual information constraint) and structural regularization (discretized projection head, sparse autoencoder).
Prompt as Knowledge Bank: Boost Vision-language model via Structural Representation for zero-shot medical detection
Yuguang Yang (Beihang University), Baochang Zhang (Beihang University)
Object DetectionTransformerPrompt EngineeringVision Language ModelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposes the StructuralGLIP framework, which utilizes a structured prompt knowledge base to achieve zero-shot object detection in medical images, supporting category-level prompts and instance-level dynamic selection;
Prompting Fairness: Integrating Causality to Debias Large Language Models
Jingling Li (Google DeepMind), Yang Liu (University of California)
Large Language ModelPrompt EngineeringText
🎯 What it does: A causal reasoning-based prompt debiasing framework is proposed to eliminate social biases in the decision-making process of large language models (LLMs);
Promptriever: Instruction-Trained Retrievers Can Be Prompted Like Language Models
Orion Weller (Johns Hopkins University), Jack Hessel (Samaya AI)
RetrievalTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study proposes a model called Promptriever that can perform retrieval based on natural language prompts, and constructs nearly 500k training data instances with instruction-level guidance.
ProtComposer: Compositional Protein Structure Generation with 3D Ellipsoids
Hannes Stark (NVIDIA), Karsten Kreis (NVIDIA)
GenerationProtein Structure PredictionTransformerFlow-based ModelGraph
🎯 What it does: ProtComposer has been developed, a protein structure generation method based on 3D ellipsoid layout, which can control the spatial layout and secondary structure of proteins through manually specified or statistically modeled ellipsoids, and implements conditional generation on the MultiFlow generation framework; it also introduces techniques such as Invariant Cross Attention, classifier-independent guidance, and self-conditioning to enhance generation quality, controllability, and diversity.
Protecting against simultaneous data poisoning attacks
Neel Alex (University of Cambridge), David Krueger (University of Montreal)
Adversarial AttackData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: This paper studies the scenario of multiple simultaneous data poisoning attacks, evaluates the robustness of existing anti-poisoning methods against multiple attacks, and proposes a new defense based on loss dynamics called BaDLoss for detection and removal.
Protein Language Model Fitness is a Matter of Preference
Cade W Gordon, Pieter Abbeel (University of California)
Protein Structure PredictionTransformerLarge Language ModelSupervised Fine-TuningBiomedical Data
🎯 What it does: This paper studies the performance of protein language models (pLM) in zero-shot protein adaptability prediction, explains performance fluctuations through sequence log-likelihood, and proposes a causal analysis based on influence functions and an unsupervised fine-tuning method for low-likelihood sequences.
Proteina: Scaling Flow-based Protein Structure Generative Models
Tomas Geffner (NVIDIA), Karsten Kreis (NVIDIA)
GenerationProtein Structure PredictionTransformerFlow-based ModelBiomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Developed Prote'INA, a flow-matching generative model for protein backbones, supporting long chains (up to 800 residues) and hierarchical fold-class conditional control, capable of autoregression and gradient guidance.
ProteinBench: A Holistic Evaluation of Protein Foundation Models
Fei YE, Quanquan Gu (ByteDance Research)
Protein Structure PredictionDiffusion modelMultimodalityBiomedical DataBenchmark
🎯 What it does: This paper presents ProteinBench, a unified multi-dimensional evaluation framework for systematically assessing the performance of protein foundation models on eight sub-tasks related to design and conformation prediction.
ProtoSnap: Prototype Alignment For Cuneiform Signs
Rachel Mikulinsky (Tel Aviv University), Hadar Averbuch-Elor (Tel Aviv University)
RecognitionGenerationData SynthesisOptimizationSupervised Fine-TuningDiffusion modelImage
🎯 What it does: Using the unsupervised ProtoSnap method, global alignment and local stroke refinement of photographed cuneiform images are performed to restore their internal structure and generate structured synthetic data.
Prototype antithesis for biological few-shot class-incremental learning
Binghao Liu (Alibaba Group), Fei Gu (Alibaba Group)
ClassificationRecognitionConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: For few-shot incremental learning in the biological field, this paper proposes the Prototype Antithesis (PA) method, which constructs unique species prototypes and shared family residual prototypes, and generates synthetic samples through residual mixing.
ProtPainter: Draw or Drag Protein via Topology-guided Diffusion
Zhengxi Lu (Zhejiang University), Min Zhang (Zhejiang University)
Protein Structure PredictionConvolutional Neural NetworkGraph Neural NetworkDiffusion modelGraph
🎯 What it does: Proposes the ProtPainter method, which uses 3D curves as topological constraints to generate protein backbones through curve encoding and achieve precise topological control;
Provable Benefit of Annealed Langevin Monte Carlo for Non-log-concave Sampling
Wei Guo (Georgia Institute of Technology), Yongxin Chen (Georgia Institute of Technology)
OptimizationStochastic Differential Equation
🎯 What it does: This paper proposes and analyzes an annealing-based Langevin Monte Carlo (ALMC) algorithm for efficient sampling under conditions that do not satisfy log-convexity or isothermality.
Provable Convergence and Limitations of Geometric Tempering for Langevin Dynamics
Omar Chehab (CREST ENSAE IP Paris), Adrien Vacher (CREST ENSAE IP Paris)
OptimizationStochastic Differential Equation
🎯 What it does: This paper studies the convergence of geometric temperature scheduling in Langevin dynamics, providing upper and lower bounds for KL and TV in both continuous and discrete time, and deriving the optimal scheduling formula.
Provable Convergence Bounds for Hybrid Dynamical Sampling and Optimization
Matthew X. Burns (University of Rochester), Michael Huang
OptimizationTabularStochastic Differential Equation
🎯 What it does: This paper reduces Large-Neighborhood Local Search (LNLS) to Block Langevin Diffusion (BLD) and provides a proof of non-asymptotic convergence for random and cyclic block selection strategies under non-ideal devices. Based on this, it derives the Wasserstein error bounds for devices with finite variations.
Provable Robust Overfitting Mitigation in Wasserstein Distributionally Robust Optimization
Shuang Liu (Academy of Mathematics and Systems Science Chinese Academy of Sciences), Xiao-Shan Gao (Academy of Mathematics and Systems Science Chinese Academy of Sciences)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A Statistically Robust Wasserstein Distributionally Robust Optimization (SR-WDRO) framework is proposed, which incorporates KL divergence on top of WDRO to simultaneously control statistical errors and adversarial noise, thereby alleviating robust overfitting.
Provable Uncertainty Decomposition via Higher-Order Calibration
Gustaf Ahdritz (Harvard University), Udi Wieder (Apple)
Image
🎯 What it does: Proposes higher-order calibration and k-times calibration theory, utilizing k-snapshot data to achieve model uncertainty decomposition.
Provable unlearning in topic modeling and downstream tasks
Stanley Wei (Princeton University), Amartya Sanyal (University of Copenhagen)
Text
🎯 What it does: This paper proposes a provable unlearning algorithm for topic models (bag-of-words language models) within the pre-training-fine-tuning framework, and provides its deletion capacity and empirical performance.
Provable weak-to-strong generalization via benign overfitting
David Xing Wu, Anant Sahai (University of California Berkeley)
ClassificationOptimizationKnowledge DistillationTabular
🎯 What it does: This paper theoretically analyzes the generalization problem from weak to strong supervision in an over-parameterized Gaussian covariance model, revealing two phases that a strong student may experience under weak teacher supervision: random guessing and perfect generalization.
Provably Accurate Shapley Value Estimation via Leverage Score Sampling
Christopher Musco (New York University), R. Teal Witter (New York University)
Explainability and InterpretabilityTabular
🎯 What it does: Proposes the Leverage SHAP algorithm, which improves Kernel SHAP using leverage score sampling to achieve near-linear model evaluation times for Shapley value estimation.
Provably Reliable Conformal Prediction Sets in the Presence of Data Poisoning
Yan Scholten (Technical University of Munich), Stephan Günnemann (Technical University of Munich)
Adversarial AttackData-Centric LearningConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A Reliable Prediction Set (RPS) is proposed, which significantly enhances the pointwise reliability of conformal prediction under data poisoning attacks by partitioning the data during the training and calibration phases and using a smoothed voting score function combined with a majority voting mechanism.
Provably Robust Explainable Graph Neural Networks against Graph Perturbation Attacks
Jiate Li (Illinois Institute of Technology), Binghui Wang (Illinois Institute of Technology)
Explainability and InterpretabilityAdversarial AttackDrug DiscoveryGraph Neural NetworkGraph
🎯 What it does: A provably robust and interpretable graph neural network (XGNN) defense framework called XGNNCert is proposed, which ensures the consistency and accuracy of explanation results when facing graph structure perturbation attacks.
Provably Safeguarding a Classifier from OOD and Adversarial Samples
Nicolas Atienza (Thales Group), Michele Sebag (LISN CNRS-INRIA Paris-Saclay University)
ClassificationAnomaly DetectionConvolutional Neural NetworkTransformerImage
🎯 What it does: Transforming the trained classifier into a self-rejecting classifier that can refuse to make predictions when encountering OOD (out-of-distribution) or adversarial samples.
Provence: efficient and robust context pruning for retrieval-augmented generation
Nadezhda Chirkova (Naver Labs Europe), Stéphane CLINCHANT
GenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: A context pruner called Provence is proposed, which is based on sentence-level sequence labeling and can automatically detect and remove irrelevant sentences in retrieval-augmented generation (RAG), compatible with multi-domain and varying lengths of retrieval results.
Proving Olympiad Inequalities by Synergizing LLMs and Symbolic Reasoning
Zenan Li (Nanjing University), Xiaoxing Ma (Nanjing University)
Large Language ModelPrompt EngineeringChain-of-Thought
🎯 What it does: A neural symbolic framework called LIPS has been designed and implemented for the automatic derivation of inequality proofs at the level of mathematical Olympiads, ultimately generating formal proofs that are both readable and verifiable in Lean 4.
Proximal Mapping Loss: Understanding Loss Functions in Crowd Counting & Localization
Wei Lin (Harbin Institute of Technology), Antoni B. Chan (City University of Hong Kong)
Object DetectionOptimizationImage
🎯 What it does: A new loss function called Proximal Mapping Loss (PML) is proposed for training crowd counting models without the assumption of intersection hypothesis;
Proxy Denoising for Source-Free Domain Adaptation
Song Tang (University of Shanghai for Science and Technology), Xiatian Zhu (University of Surrey)
Domain AdaptationVision Language ModelContrastive LearningImage
🎯 What it does: This paper proposes a Proxy DeNoise method for source-free domain adaptation (SFDA), which generates and corrects pseudo-labels using visual-language models (such as CLIP) and guides the target model through mutual information distillation.
PseDet: Revisiting the Power of Pseudo Label in Incremental Object Detection
Qiuchen Wang (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)
Object DetectionImage
🎯 What it does: This study explores the use of pseudo-labels in incremental object detection and proposes the PseDet framework to enhance the quality of pseudo-labels from old models and mitigate forgetting.
PT-T2I/V: An Efficient Proxy-Tokenized Diffusion Transformer for Text-to-Image/Video-Task
Jing Wang (Shenzhen Campus of Sun Yat-Sen University), Xiaodan Liang (360 AI Research)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImageVideoText
🎯 What it does: Proposed a Proxy Token-based Diffusion Transformer (PT-DiT) for efficient generation from text to image/video.
Pursuing Better Decision Boundaries for Long-Tailed Object Detection via Category Information Amount
Yanbiao Ma (Xidian University), Jiayi Chen (Xidian University)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: Proposes and quantifies the Information Amount of categories, and based on this, designs the Information Amount Guided Angle Cosine Margin (IGAM) loss, which dynamically adjusts the decision space of each category to alleviate category bias in long-tail detection.
Pursuing Feature Separation based on Neural Collapse for Out-of-Distribution Detection
Yingwen Wu (Shanghai Jiao Tong University), Xiaolin Huang (Shanghai Jiao Tong University)
Anomaly DetectionConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes utilizing the property of Neural Collapse to achieve dimensional separation of ID and OOD features through a feature separation loss during the model fine-tuning phase, thereby enhancing OOD detection performance.
Pushing the Limits of All-Atom Geometric Graph Neural Networks: Pre-Training, Scaling, and Zero-Shot Transfer
Zihan Pengmei (University of Chicago), Huzefa Rangwala (Amazon Web Services)
Protein Structure PredictionGraph Neural NetworkTransformerSupervised Fine-TuningGraphBiomedical Data
🎯 What it does: This paper constructs a universal descriptor for zero-shot transfer directly in protein systems through denoising pre-training of all atomic geometric graph neural networks (Geom-GNN), and systematically evaluates its scaling behavior in self-supervised, supervised, and unsupervised tasks; it also demonstrates the effectiveness of combining pre-trained graph embeddings with other architectures (such as VAMPnet and ProNet) to enhance protein dynamics modeling and folding classification.
PuzzleFusion++: Auto-agglomerative 3D Fracture Assembly by Denoise and Verify
Zhengqing Wang (Simon Fraser University), Yasutaka Furukawa (Simon Fraser University)
Object DetectionGenerationPose EstimationTransformerDiffusion modelPoint Cloud
🎯 What it does: An automatic agglomerative 3D fragment reconstruction framework called PuzzleFusion++ is proposed, which iteratively aligns and merges fragments to ultimately complete the reconstruction of the entire object.
PvNeXt: Rethinking Network Design and Temporal Motion for Point Cloud Video Recognition
Jie Wang (Beijing Institute of Technology), Jianan Li
RecognitionVideoPoint Cloud
🎯 What it does: The PvNeXt architecture is proposed, which effectively recognizes point cloud videos using personalized single-step queries.
PWM: Policy Learning with Multi-Task World Models
Ignat Georgiev (Georgia Institute of Technology), Animesh Garg (Georgia Institute of Technology)
OptimizationRobotic IntelligenceReinforcement LearningWorld ModelSequential
🎯 What it does: The PWM method is proposed, which utilizes a pre-trained multi-task world model to quickly learn continuous control strategies for various tasks through first-order gradient optimization.
Pyramidal Flow Matching for Efficient Video Generative Modeling
Yang Jin (Peking University), Zhouchen Lin (Peking University)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelFlow-based ModelAuto EncoderVideo
🎯 What it does: A unified pyramid flow matching framework is designed for efficient video generation, combining spatial pyramid flow and temporal pyramid conditions, and end-to-end training of a single Diffusion Transformer.
Q-Adapter: Customizing Pre-trained LLMs to New Preferences with Forgetting Mitigation
Yi-Chen Li (Nanjing University), Bo An (Nanyang Technological University)
Recommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes Q-Adapter, which utilizes residual Q learning to directly learn adaptive modules from new human preference data, customizing pre-trained LLMs.
Q-SFT: Q-Learning for Language Models via Supervised Fine-Tuning
Joey Hong (University of California Berkeley), Sergey Levine (University of California Berkeley)
Robotic IntelligenceTransformerSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A Q-learning method based on offline reinforcement learning, called Q-SFT, is proposed. It achieves value learning and policy extraction in tasks such as multi-turn dialogue and robot control by directly optimizing weighted cross-entropy on pre-trained language/visual models.
QA-Calibration of Language Model Confidence Scores
Putra Manggala (University of Amsterdam), Aaditya Ramdas (Amazon and Carnegie Mellon University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A new generative question answering (QA) confidence calibration method called QA-calibration is proposed, addressing the interpretability issues of traditional average calibration.
QERA: an Analytical Framework for Quantization Error Reconstruction
Cheng Zhang (Imperial College London), Yiren Zhao (Imperial College London)
CompressionOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a parsing framework named QERA to solve the quantization error reconstruction problem, and based on this, provides exact solutions (QERA-exact) and approximate solutions (QERA-approx).
Qinco2: Vector Compression and Search with Improved Implicit Neural Codebooks
Théophane Vallaeys (Meta), Matthijs Douze (Meta)
RetrievalCompressionTabular
🎯 What it does: Proposes QINCO2, an improved implicit neural residual quantization method that balances vector compression and trillion-scale approximate nearest neighbor search;
QMP: Q-switch Mixture of Policies for Multi-Task Behavior Sharing
Grace Zhang (University of Southern California), Joseph J Lim
Robotic IntelligenceReinforcement LearningMixture of Experts
🎯 What it does: Proposes the Q-switch Mixture of Policies (QMP) framework, which utilizes task-specific Q-functions to select optimal behaviors, thereby achieving unbiased behavior sharing in multi-task reinforcement learning and improving sample efficiency.
qNBO: quasi-Newton Meets Bilevel Optimization
Sheng Fang (Fujian Medical University), Jin Zhang (Southern University of Science and Technology)
OptimizationHyperparameter SearchMeta LearningImageText
🎯 What it does: A dual-layer optimization algorithm based on the quasi-Newton (qNBO) framework is proposed to efficiently estimate supergradients and accelerate the solution of the lower-level problem.
QP-SNN: Quantized and Pruned Spiking Neural Networks
Wenjie Wei (University of Electronic Science and Technology of China), Yang Yang (University of Electronic Science and Technology of China)
ClassificationObject DetectionSpiking Neural NetworkReinforcement LearningImage
🎯 What it does: A hardware-friendly lightweight pulse neural network (QP-SNN) is constructed, obtaining the base model through unified quantization and structured pruning, and further enhancing performance through weight rescaling (ReScaW) and singular value pruning (SVS).
QPM: Discrete Optimization for Globally Interpretable Image Classification
Thomas Norrenbrock (Institute for Information Processing), Bodo Rosenhahn (Institute for Information Processing)
ClassificationOptimizationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: A discrete optimization method based on quadratic programming, QPM, is proposed, which achieves a globally interpretable image classification model using optimal binary feature allocation with only 5 common features for each category.
QuaDiM: A Conditional Diffusion Model For Quantum State Property Estimation
Yehui Tang (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
TransformerDiffusion modelGraphPhysics Related
🎯 What it does: This study proposes QuaDiM, a non-autoregressive quantum state property estimation method based on diffusion models.
Quality Measures for Dynamic Graph Generative Models
Ryien Hosseini (University of Chicago), Henry Hoffmann (University of Chicago)
GenerationData SynthesisComputational EfficiencyGraph Neural NetworkGraphTime SeriesSequential
🎯 What it does: A continuous-time dynamic graph generation model evaluation metric based on the Johnson-Lindenstrauss transform is proposed, which directly performs random projection on event sequences without the need for discrete snapshots.
Quality over Quantity in Attention Layers: When Adding More Heads Hurts
Noah Amsel (New York University), Joan Bruna (Flatiron Institute)
Transformer
🎯 What it does: This study investigates the impact of the rank of attention layers and the number of heads on representational capacity, proving that low-rank attention cannot effectively approximate tasks such as nearest neighbor search, and experimentally validating the theoretical conclusions.
Quamba: A Post-Training Quantization Recipe for Selective State Space Models
Hung-Yueh Chiang (University of Texas at Austin), Diana Marculescu (University of Texas at Austin)
Large Language ModelText
🎯 What it does: Designed and implemented a post-training 8-bit quantization scheme for selective state space models (SSM) - Quamba, supporting the deployment of large models like Mamba and Jamba on low-resource devices.
Quantifying Generalization Complexity for Large Language Models
Zhenting Qi (Harvard University), James R. Glass
GenerationOptimizationTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: The SCYLLA evaluation framework is proposed, which dynamically generates task instances and quantifies the generalization ability of large language models by comparing ID and OOD data.
Quantitative Approximation for Neural Operators in Nonlinear Parabolic Equations
Takashi Furuya (Doshisha University), Satoshi Okuda (Rikkyo University)
Ordinary Differential Equation
🎯 What it does: This paper presents a quantitative error upper bound for the neural operator approximation of nonlinear parabolic equation solvers and proves that under certain conditions, exponential growth in model complexity can be avoided.
Quantized Spike-driven Transformer
Xuerui Qiu (University of Electronic Science and Technology of China), Haizhou Li (Chinese University of Hong Kong)
Object DetectionSegmentationComputational EfficiencyKnowledge DistillationSpiking Neural NetworkTransformerImage
🎯 What it does: A low-bit-weight quantized pulse-driven Transformer (QSD-Transformer) is proposed, which enhances performance while maintaining energy efficiency through information-enhanced LIF neurons and a fine-grained distillation mechanism.
Quantum (Inspired) $D^2$-sampling with Applications
Poojan Chetan Shah (Indian Institute of Technology Delhi), Ragesh Jaiswal (Indian Institute of Technology Delhi)
ImagePhysics Related
🎯 What it does: Designed a quantum and quantum-inspired k-means++ based on D₂ sampling, achieving O(log k) approximation under the QRAM model;
Quantum-PEFT: Ultra parameter-efficient fine-tuning
Toshiaki Koike-Akino (Mitsubishi Electric Research Laboratories), Volkan Cevher (École Polytechnique Fédérale de Lausanne)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes Quantum-PEFT, which utilizes Pauli parameterization and full-rank unit reparameterization in quantum circuits to achieve extreme parameter-efficient fine-tuning of large pre-trained models.
Query-based Knowledge Transfer for Heterogeneous Learning Environments
Norah Alballa (King Abdullah University of Science and Technology), Marco Canini (Queen Mary University of London)
Federated LearningKnowledge DistillationImageBiomedical Data
🎯 What it does: A query-driven knowledge transfer framework (QKT) is proposed, allowing offline clients to efficiently acquire knowledge from other client models based on their own query needs without sharing raw data.
Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Chaochen Gao (Institute of Information Engineering, Chinese Academy of Sciences), Songlin Hu (Xiaohongshu Inc)
GenerationData SynthesisTransformerLarge Language ModelText
🎯 What it does: A query-based long text synthesis method called Quest has been developed to generate diverse and semantically relevant long-context training data.
R-Sparse: Rank-Aware Activation Sparsity for Efficient LLM Inference
Zhenyu Zhang (University of Texas at Austin), Steven Li (Meta AI)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes a training-free, input activation sparse-based Rank-Aware Activation Sparsity (R-Sparse) method for efficient inference of large language models (LLMs).
R2Det: Exploring Relaxed Rotation Equivariance in 2D Object Detection
Zhiqiang Wu (East China Normal University), Xian Wei
ClassificationObject DetectionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a Relaxed Rotation-Equivariant Group (R_n) and its corresponding R2GConv operation, and constructs a lightweight backbone network R2Net and an improved 2D object detector R2Det, aiming to address the issue of rotation symmetry disruption in practical scenarios.
RA-TTA: Retrieval-Augmented Test-Time Adaptation for Vision-Language Models
Youngjun Lee (KAIST), Jae-Gil Lee (KAIST)
RetrievalDomain AdaptationTransformerLarge Language ModelVision Language ModelImageTextRetrieval-Augmented Generation
🎯 What it does: A retrieval-enhanced test-time adaptive method RA-TTA is designed to improve zero-shot inference of VLM under distribution shift by utilizing an external image database.
Radar: Fast Long-Context Decoding for Any Transformer
Yongchang Hao (University of Alberta), Frederick Tung (RBC Borealis)
GenerationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes Radar, a training-free dynamic context selection method that significantly accelerates inference while maintaining the original attention structure of Transformers.
RAG-DDR: Optimizing Retrieval-Augmented Generation Using Differentiable Data Rewards
Xinze Li (Northeastern University), Chenyan Xiong (Carnegie Mellon University)
RetrievalOptimizationTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the Differentiable Data Rewards (DDR) method, which optimizes retrieval-augmented generation (RAG) systems end-to-end by collecting system rewards and aligning the data preferences of various modules, enabling the retrieval and generation modules to work together to produce more accurate answers.
RAG-SR: Retrieval-Augmented Generation for Neural Symbolic Regression
Hengzhe Zhang (Victoria University of Wellington), Mengjie Zhang (Victoria University of Wellington)
GenerationRetrievalOptimizationTransformerContrastive LearningTabularRetrieval-Augmented Generation
🎯 What it does: A feature construction symbolic regression framework RAG-SR is proposed, which combines retrieval-augmented neural semantic libraries.
RainbowPO: A Unified Framework for Combining Improvements in Preference Optimization
Hanyang Zhao (Columbia University), Sambit Sahu (Capital One)
Recommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A unified preference optimization framework RAINBOWPO is proposed, which disassembles and combines seven key components of the DPO series algorithms to enhance the alignment effect of LLM.
RandLoRA: Full rank parameter-efficient fine-tuning of large models
Paul Albert (Australian Institute for Machine Learning), Ehsan Abbasnejad (Australian Institute for Machine Learning)
Supervised Fine-TuningImageText
🎯 What it does: This paper proposes RandLoRA, a parameter-efficient and memory-friendly full-rank fine-tuning method that achieves this by learning a linear combination of random low-rank bases and only training diagonal scaling matrices.
Random Is All You Need: Random Noise Injection on Feature Statistics for Generalizable Deep Image Denoising
Zhengwei Yin (University of Tokyo), Yinqiang Zheng (University of Tokyo)
RestorationConvolutional Neural NetworkImage
🎯 What it does: Designed and implemented RNINet, an image denoising network based on a simplified encoder-decoder structure, incorporating a noise injection block during the encoding phase, allowing effective denoising of various unseen noise types using only Gaussian noise for training.
Random-Set Neural Networks
Shireen Kudukkil Manchingal (Oxford Brookes University), Fabio Cuzzolin (Oxford Brookes University)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: A Random Set Neural Network (RS-NN) is proposed, whose output is a belief function of a set of categories, capable of explicitly modeling epistemic uncertainty in classification tasks.
Range, not Independence, Drives Modularity in Biologically Inspired Representations
Will Dorrell (University College London), James C. R. Whittington (Stanford University)
OptimizationRepresentation LearningReinforcement Learning from Human FeedbackRecurrent Neural NetworkAuto EncoderImageSequential
🎯 What it does: A modular theory for linear autoencoders with biological constraints (non-negativity, energy efficiency) is proposed, providing necessary and sufficient conditions. This theory is extended to nonlinear feedforward networks, recurrent networks, and neuroscience experiments, explaining how neurons exhibit modular or mixed selection under different conditions.
Ranking-aware adapter for text-driven image ordering with CLIP
Wei-Hsiang Yu (National Yang Ming Chiao Tung University), Yi-Hsuan Tsai (Atmanity Inc.)
RetrievalRecommendation SystemTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: A lightweight Ranking-Aware Adapter is proposed, reconstructing the pre-trained CLIP for learning to rank (LTR) tasks, utilizing text-driven visual differences to achieve multi-image ranking.
RankSHAP: Shapley Value Based Feature Attributions for Learning to Rank
Tanya Chowdhury (University of Massachusetts Amherst), James Allan (University of Massachusetts Amherst)
RetrievalExplainability and InterpretabilityLarge Language ModelText
🎯 What it does: This paper proposes RankSHAP, a feature importance explanation method for ranking tasks.
Rapid Selection and Ordering of In-Context Demonstrations via Prompt Embedding Clustering
Kha Pham (Deakin University), Truyen Tran (Deakin University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study investigates the sensitivity of performance to the demonstration order in in-context learning (ICL) with large language models, discovering a 'first/last example' clustering phenomenon in the prompt embedding space. It proposes an efficient demonstration selection and ranking method based on this clustering (Cluster-based Search), reducing the search complexity from factorial to quadratic.
RAPID: Retrieval Augmented Training of Differentially Private Diffusion Models
Tanqiu Jiang (Stony Brook University), Ting Wang (Stony Brook University)
GenerationData SynthesisSafty and PrivacyComputational EfficiencyDiffusion modelContrastive LearningImageRetrieval-Augmented Generation
🎯 What it does: By integrating Retrieval-Augmented Generation (RAG) technology into the training of differential privacy diffusion models, a trajectory knowledge base is constructed using public data. During training on private data, only the later sampling steps are subjected to differential privacy processing, significantly improving generation quality, reducing memory usage, and lowering inference costs.
Rapidly Adapting Policies to the Real-World via Simulation-Guided Fine-Tuning
Patrick Yin (University of Washington), Abhishek Gupta (University of Washington)
Robotic IntelligenceReinforcement Learning
🎯 What it does: After pre-training the policy and value function in simulation, online fine-tuning of the real robot is performed using the SGFT framework (reward reshaping based on simulation value function and short horizon search).
Rare event modeling with self-regularized normalizing flows: what can we learn from a single failure?
Charles Dawson (Massachusetts Institute of Technology), Chuchu Fan (Massachusetts Institute of Technology)
Data SynthesisAnomaly DetectionFlow-based ModelImageTime Series
🎯 What it does: This paper proposes CALNF—a self-regularizing normalization flow framework for learning posterior distributions from a very small amount of failure event data.
Rare-to-Frequent: Unlocking Compositional Generation Power of Diffusion Models on Rare Concepts with LLM Guidance
Dongmin Park (KRAFTON), Jaewoong Cho (KRAFTON)
GenerationLarge Language ModelDiffusion modelImageBenchmarkChain-of-Thought
🎯 What it does: Proposes the R2F framework, which utilizes LLM to extract the mapping of rare concepts to frequent concepts, and alternately guides during the sampling process of diffusion models to significantly enhance the image generation quality of rare concepts.
RaSA: Rank-Sharing Low-Rank Adaptation
Zhiwei He (Shanghai Jiao Tong University), Rui Wang (Tencent AI Lab)
OptimizationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the RaSA (Rank‑Sharing Low‑Rank Adaptation) method, which enhances the model's expressive capability by sharing a portion of the rank across layers based on LoRA's low-rank parameter updates, while maintaining the same number of trainable parameters.
Rational Decision-Making Agent with Learning Internal Utility Judgment
Yining Ye (Tsinghua University), Maosong Sun (Tsinghua University)
Large Language ModelReinforcement LearningAgentic AITabular
🎯 What it does: Proposes RaDAgent, which constructs an internal utility judgment mechanism based on LLM, enabling the agent to make autonomous decisions without external evaluation.
Rationalizing and Augmenting Dynamic Graph Neural Networks
Guibin Zhang (Tongji University), Jian Guo (International Digital Economy Academy)
Domain AdaptationAnomaly DetectionGraph Neural NetworkContrastive LearningGraphTime Series
🎯 What it does: A graph data augmentation framework for dynamic graphs, DyAug, is proposed, which maintains temporal consistency through temporal conditional graph rationalization to enhance the performance, robustness, and generalization ability of dynamic GNNs.
RazorAttention: Efficient KV Cache Compression Through Retrieval Heads
Hanlin Tang (Huawei Technologies Co), Gongyi Wang (Huawei Technologies Co)
RetrievalCompressionTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes RazorAttention, a KV cache compression method for large models with long contexts that retains all semantic information while significantly reducing memory usage.
RB-Modulation: Training-Free Stylization using Reference-Based Modulation
Litu Rout (Google), Wen-Sheng Chu (Google)
Image TranslationGenerationOptimizationDiffusion modelImage
🎯 What it does: A training-free reference tone modulation framework named RB-Modulation is proposed for achieving stylization and content-style combination in diffusion models.
RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation
Songming Liu (Tsinghua University), Jun Zhu (Tsinghua University)
Robotic IntelligenceTransformerVision Language ModelDiffusion modelMultimodality
🎯 What it does: A foundational model RDT based on a diffusion model for dual-arm visual language control is proposed and trained for language-conditioned dual-hand operations.
Re-Aligning Language to Visual Objects with an Agentic Workflow
Yuming Chen (Nankai University), Yibing Song
Object DetectionTransformerLarge Language ModelAgentic AIVision Language ModelImageTextMultimodality
🎯 What it does: A workflow based on LLM agents (Real-LOD) has been designed and implemented to automatically correct visual object descriptions generated by VLM, aligning language expressions with target objects and generating high-quality language-visual paired data for training language-based object detection models.
Re-evaluating Open-ended Evaluation of Large Language Models
Siqi Liu (Google DeepMind), Marc Lanctot (Google DeepMind)
Large Language ModelPrompt EngineeringText
🎯 What it does: A framework for 'equilibrium assessment' based on N-player general games is proposed to replace the traditional Elo method for open evaluation of large language models (LLMs), focusing on addressing issues of prompt redundancy and bias.
Re-Evaluating the Impact of Unseen-Class Unlabeled Data on Semi-Supervised Learning Model
Rundong He (Shandong University), Tailin Wu (Westlake University)
ClassificationDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: Reassess the impact of unseen classes in semi-supervised learning on unlabeled data, and propose the RE-SSL evaluation framework along with a set of global and local robustness metrics.
Re-Imagining Multimodal Instruction Tuning: A Representation View
Yiyang Liu (University of Missouri Kansas City), Cheng Han (University of Missouri Kansas City)
Representation LearningTransformerVision Language ModelMultimodality
🎯 What it does: For instruction fine-tuning of large-scale multimodal models, a parameter-efficient representation tuning method (MRT) is proposed, which only edits the representations in the visual encoder, cross-modal projection layer, and language model while keeping the model parameters frozen.
Reading Your Heart: Learning ECG Words and Sentences via Pre-training ECG Language Model
Jiarui Jin (Peking University), Shenda Hong (Peking University)
Anomaly DetectionRepresentation LearningTransformerContrastive LearningTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: The HeartLang framework is proposed, utilizing QRS-Tokenizer to split electrocardiograms into heartbeat words and rhythm sentences, constructing an ECG vocabulary and learning multi-level representations of unlabeled ECG through vector quantization of heartbeat reconstruction and masked sentence pre-training.
Ready-to-React: Online Reaction Policy for Two-Character Interaction Generation
Zhi Cen (Zhejiang University), Ruizhen Hu (Shenzhen University)
GenerationTransformerDiffusion modelVideo
🎯 What it does: A reaction strategy-based online two-role interaction generation method called Ready-to-React is proposed.
Real-time design of architectural structures with differentiable mechanics and neural networks
Rafael Pastrana (Princeton University), Ryan P Adams
OptimizationMeshPhysics Related
🎯 What it does: This study investigates a model that combines neural networks with a differentiable mechanical simulator, capable of inferring the shape approximations of architectural structures (such as stone arch shells and cable towers) in real-time while maintaining mechanical integrity.
Real-Time Video Generation with Pyramid Attention Broadcast
Xuanlei Zhao (National University of Singapore), Yang You (National University of Singapore)
GenerationComputational EfficiencyTransformerVideoBenchmark
🎯 What it does: This paper proposes Pyramid Attention Broadcast (PAB), a real-time, training-free acceleration method that improves the efficiency of DiT video generation.
Real2Code: Reconstruct Articulated Objects via Code Generation
Zhao Mandi (Stanford University), Shuran Song (Stanford University)
GenerationPose EstimationTransformerLarge Language ModelSupervised Fine-TuningImagePoint Cloud
🎯 What it does: The Real2Code method is proposed, which utilizes visual input to reconstruct multi-joint objects through segmentation, shape completion, and code generation using a large language model.
Realistic Evaluation of Deep Partial-Label Learning Algorithms
Wei Wang (University of Tokyo), Masashi Sugiyama (University of Tokyo)
ClassificationData-Centric LearningConvolutional Neural NetworkImageTabularBenchmark
🎯 What it does: A unified Partial Label Learning (PLL) evaluation benchmark PLENCH has been constructed, new model selection criteria have been proposed, and a human-annotated image dataset PLCIFAR10 has been created.
Reasoning Elicitation in Language Models via Counterfactual Feedback
Alihan Hüyük, Javier Gonzalez (Microsoft Research)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a novel evaluation and fine-tuning framework for causal reasoning in large language models (LLMs), covering new metrics such as necessity/sufficiency inconsistency rate, and fine-tuning the model through 'Causal Consistency Feedback' (CCF);
Reasoning of Large Language Models over Knowledge Graphs with Super-Relations
Song Wang (University of Virginia), Yada Zhu (IBM)
RetrievalTransformerLarge Language ModelGraph
🎯 What it does: The ReKnoS framework is proposed, utilizing super-relations for multi-path parallel reasoning on knowledge graphs, significantly expanding the search space and improving retrieval success rates.
Reasoning with Latent Thoughts: On the Power of Looped Transformers
Nikunj Saunshi (Google Research), Sashank J. Reddi (Google Research)
TransformerTextChain-of-Thought
🎯 What it does: This paper studies the cyclic (weight-sharing) Transformer, which enhances reasoning capabilities by iterating the same network block multiple times under the same number of parameters, and applies it to synthetic reasoning tasks and downstream reasoning benchmarks for language models.
Reasoning-Enhanced Healthcare Predictions with Knowledge Graph Community Retrieval
Pengcheng Jiang (University of Illinois at Urbana-Champaign), Jiawei Han (University of Illinois at Urbana-Champaign)
ClassificationRecommendation SystemGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTabularBiomedical DataElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: Proposes the KARE framework, which combines knowledge graph community retrieval with LLM reasoning to enhance the accuracy of medical predictions.