ICLR 2025 Papers with AI Summaries
International Conference on Learning Representations · 3704 papers
→ ICLR 2025 papers with code (1682)
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 3704 ICLR 2025 papers by keyword, author or institution —
free trial on arXivSub.
(Mis)Fitting Scaling Laws: A Survey of Scaling Law Fitting Techniques in Deep Learning
Margaret Li (University of Washington), Luke Zettlemoyer (University of Washington)
TransformerTextReview/Survey Paper
🎯 What it does: This paper provides a systematic review of 50 papers on scaling laws, summarizing the differences in formulas, training setups, evaluation methods, curve fitting, and more. It also explores the impact of these differences on the conclusions of scaling laws through experiments conducted on both self-trained data and public datasets.
{$\tau$}-bench: A Benchmark for \underline{T}ool-\underline{A}gent-\underline{U}ser Interaction in Real-World Domains
Shunyu Yao, Karthik R Narasimhan
TransformerLarge Language ModelAgentic AITextTabularBenchmark
🎯 What it does: A τ-bench benchmark is proposed to evaluate the reliability of language agents when interacting with human users and tools, and adhering to domain rules.
$\gamma-$MoD: Exploring Mixture-of-Depth Adaptation for Multimodal Large Language Models
Yaxin Luo (MBZUAI), Rongrong Ji (MBZUAI)
Domain AdaptationComputational EfficiencyTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: This paper proposes a hybrid depth (MoD) adaptation strategy named γ-MoD, which can convert most dense layers of existing multimodal large language models (MLLMs) into sparse MoD layers, significantly reducing computational overhead while maintaining performance.
$\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs
Vlad Sobal (Meta), Yann LeCun (Meta)
Representation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Proposes X-Sample Contrastive Loss (X-CLR), which enhances the modeling of sample relationships through a soft similarity graph.
$\phi$-Update: A Class of Policy Update Methods with Policy Convergence Guarantee
Wenye Li (Fudan University), Ke Wei (Fudan University)
OptimizationReinforcement LearningTabular
🎯 What it does: A general strategy update framework, ϕ-update, is proposed and analyzed, covering methods such as softmax NPG and Hadamard PG. It is proven to have global asymptotic convergence, policy convergence, precise asymptotic linear convergence rates, and global linear convergence.
$\sigma$-zero: Gradient-based Optimization of $\ell_0$-norm Adversarial Examples
Antonio Emanuele Cinà (University of Genoa), Marcello Pelillo (Ca' Foscari University of Venice)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A new gradient-based attack method, σ-zero, is proposed to find the smallest sparse ($l_0$-norm) adversarial examples to evaluate the robustness of deep networks under sparse perturbations.
$\text{D}_{2}\text{O}$: Dynamic Discriminative Operations for Efficient Long-Context Inference of Large Language Models
Zhongwei Wan (Ohio State University), Mi Zhang (Ohio State University)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: The D2O method is proposed, which compresses the KV cache of large language models using dynamic hierarchy and token-level discrimination operations, significantly reducing memory usage and improving throughput during long text inference.
$\text{I}^2\text{AM}$: Interpreting Image-to-Image Latent Diffusion Models via Bi-Attribution Maps
Junseo Park (Dongguk University), Hyeryung Jang (Dongguk University)
Image TranslationObject DetectionGenerationSuper ResolutionDiffusion modelImage
🎯 What it does: The I2AM method is proposed to generate bidirectional image-to-image attribution maps by aggregating cross-attention across time steps, heads, and layers to explain and debug I2I generative models.
$F^3Set$: Towards Analyzing Fast, Frequent, and Fine-grained Events from Videos
Zhaoyu Liu (Shanghai Jiao Tong University), Jin Song Dong (National University of Singapore)
ClassificationRecognitionObject DetectionRecurrent Neural NetworkTransformerVideoBenchmark
🎯 What it does: This paper studies the detection of fast, frequent, and fine-grained (F3) events in videos, and proposes a new benchmark dataset called F3 Set along with a corresponding end-to-end detection model named F3 ED.
$InterLCM$: Low-Quality Images as Intermediate States of Latent Consistency Models for Effective Blind Face Restoration
Senmao Li (Nankai University), Ming-Ming Cheng (Nankai University)
RestorationDiffusion modelImage
🎯 What it does: Treating low-quality facial images as intermediate states in the Latent Consistency Model (LCM), blind face restoration is achieved by utilizing LCM's few-step inference and semantic consistency, combined with a visual encoder and a spatial encoder.
$q$-exponential family for policy optimization
Lingwei Zhu (University of Tokyo), Martha White (University of Alberta)
OptimizationReinforcement LearningTabular
🎯 What it does: This study investigates and implements the q-exponential family (including Student's t, light-tailed and heavy-tailed q-Gaussian) as a feasible strategy parameterization for continuous action spaces, systematically evaluating its online and offline performance across various actor-critic algorithms.
$R^2$-Guard: Robust Reasoning Enabled LLM Guardrail via Knowledge-Enhanced Logical Reasoning
Mintong Kang (University of Illinois), Bo Li (University of Illinois)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: R-Guard 2 is proposed, a security protection framework for LLM that combines data-driven classification prediction with knowledge-based probabilistic graphical model inference.
3D StreetUnveiler with Semantic-aware 2DGS - a simple baseline
Jingwei Xu (ShanghaiTech University), Shenghua Gao (Fudan University)
RestorationAutonomous DrivingGaussian SplattingVideo
🎯 What it does: Using vehicle camera videos and LiDAR data, reconstruct crowded street scenes by removing temporarily stationary vehicles and pedestrians, ultimately generating a 3D representation of an empty street that supports rendering from any viewpoint.
3D Vision-Language Gaussian Splatting
Qucheng Peng (University of Central Florida), Ziyan Wu (United Imaging Intelligence)
SegmentationGenerationVision Language ModelGaussian SplattingMultimodalityPoint Cloud
🎯 What it does: Proposes a 3D vision-language Gaussian scattering framework that utilizes cross-modal rasterization and camera viewpoint mixing techniques to achieve open vocabulary semantic scene understanding.
3D-AffordanceLLM: Harnessing Large Language Models for Open-Vocabulary Affordance Detection in 3D Worlds
Hengshuo Chu (Harbin Institute of Technology), Liqiang Nie (Huawei Noah's Ark Lab)
Object DetectionSegmentationTransformerLarge Language ModelSupervised Fine-TuningMultimodalityPoint Cloud
🎯 What it does: The IRAS task is proposed and the 3D-AffordanceLLM is implemented, utilizing LLM to directly generate 3D point cloud segmentation masks through the <AFF> token, achieving open vocabulary instruction reasoning for affordance detection.
3D-MolT5: Leveraging Discrete Structural Information for Molecule-Text Modeling
Qizhi Pei (Renmin University of China), Lijun Wu (Shanghai AI Laboratory)
GenerationDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningMultimodalityGraphBiomedical Data
🎯 What it does: Proposes the 3D-MolT5 framework, which unifies the modeling of molecular sequences, 3D structures, and text;
3D-Properties: Identifying Challenges in DPO and Charting a Path Forward
Yuzi Yan (Tsinghua University), Dong Yan (Baichuan AI)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Analyzed and explained the three major defects (3D attributes) of DPO in LLM alignment, and proposed regularization methods to enhance training stability and performance.
3D-SPATIAL MULTIMODAL MEMORY
Xueyan Zou (University of California San Diego), Xiaolong Wang (NVIDIA)
RetrievalCompressionGaussian SplattingMultimodality
🎯 What it does: A 3D spatial multimodal memory system M3 is proposed, which combines Gaussian splatting and various foundational models to store multimodal features of static scenes.
3DGS-Drag: Dragging Gaussians for Intuitive Point-Based 3D Editing
Jiahua Dong (University of Illinois), Yu-Xiong Wang (University of Illinois)
GenerationData SynthesisDiffusion modelGaussian SplattingPoint Cloud
🎯 What it does: The 3DGS-Drag framework is proposed, allowing users to intuitively edit real 3D scenes by specifying drag points (handle and target) in three-dimensional space, enabling various editing tasks such as geometric displacement, shape adjustment, completion, and content extension.
3DIS: Depth-Driven Decoupled Image Synthesis for Universal Multi-Instance Generation
Dewei Zhou (Zhejiang University), Yi Yang (Zhejiang University)
SegmentationGenerationData SynthesisDiffusion modelImageText
🎯 What it does: A 3DIS framework is proposed, which splits multi-instance generation into two stages: first generating layout-controllable scene depth maps, and then using a ControlNet-based training-independent detail renderer to generate high-quality RGB images, thereby achieving fine control over instance positions and attributes.
3DitScene: Editing Any Scene via Language-guided Disentangled Gaussian Splatting
Qihang Zhang (Chinese University of Hong Kong), Ceyuan Yang (ByteDance)
Image TranslationSegmentationGenerationDepth EstimationKnowledge DistillationDiffusion modelGaussian SplattingImage
🎯 What it does: A unified 3D visual editing framework called 3DitScene is proposed, utilizing a language-based separable Gaussian expansion technique to achieve seamless editing from 2D images to 3D scenes, supporting precise manipulation of both panoramas and individual objects.
3DMolFormer: A Dual-channel Framework for Structure-based Drug Discovery
Xiuyuan Hu (Tsinghua University), Xue Liu (McGill University)
Drug DiscoveryTransformerSupervised Fine-TuningReinforcement LearningBiomedical Data
🎯 What it does: A unified dual-channel Transformer framework called 3DMolFormer is proposed, which can perform protein-ligand docking and pocket-aware 3D drug design.
3DTrajMaster: Mastering 3D Trajectory for Multi-Entity Motion in Video Generation
Xiao FU, Dahua Lin (Chinese University of Hong Kong)
GenerationData SynthesisTransformerDiffusion modelVideoText
🎯 What it does: A model named 3DTrajMaster has been developed, which can precisely control the 3D motion of multiple entities during video generation based on text prompts and multi-entity 6DoF pose sequences.
4K4DGen: Panoramic 4D Generation at 4K Resolution
Renjie Li (Bytedance), Zhiwen Fan (University of Texas at Austin)
GenerationData SynthesisDiffusion modelGaussian SplattingImageVideo
🎯 What it does: The 4K4DGen framework is proposed, which can transform a single 4K panoramic image into a 360° high-resolution (4096×2048) and interactive 4D scene, supporting real-time rendering from any point in time and perspective.
6D Object Pose Tracking in Internet Videos for Robotic Manipulation
Georgy Ponimatkin (Czech Technical University in Prague), Josef Sivic (Czech Technical University in Prague)
Object TrackingPose EstimationRobotic IntelligenceLarge Language ModelVideoRetrieval-Augmented Generation
🎯 What it does: A method is proposed to estimate and track the 6D pose of a target object in real-time from online teaching videos without prior knowledge of the 3D model of the target object, and to convert the obtained trajectory into robot execution trajectories.
6DGS: Enhanced Direction-Aware Gaussian Splatting for Volumetric Rendering
Zhongpai Gao (United Imaging Intelligence), Ziyan Wu (United Imaging Intelligence)
Gaussian SplattingPoint Cloud
🎯 What it does: The 6D Gaussian Splatting (6DGS) method is proposed, improving the perspective-dependent volumetric rendering performance of 3DGS.
A Benchmark for Semantic Sensitive Information in LLMs Outputs
Qingjie Zhang (Tsinghua University), Chao Zhang (Tsinghua University)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes a definition and evaluation framework for 'Semantically Sensitive Information (SemSI)' and constructs the SemSI-Set (10,830 natural question-answer pairs) and the compressed SemSI-cSet (1,000 pairs). It conducts SemSI benchmark testing on 25 mainstream LLMs, revealing that they generally produce SemSI under simple natural questioning.
A Black Swan Hypothesis: The Role of Human Irrationality in AI Safety
Hyunin Lee (University of California Berkeley), Ming Jin (Virginia Tech)
OptimizationSafty and Privacy
🎯 What it does: This paper introduces the concept of S-BLACK SWAN, clarifying that even in unchanged environments, high-risk rare events can occur due to human illusions regarding the value and probability of events.
A Causal Lens for Learning Long-term Fair Policies
Jacob Lear (University of Arkansas), Lu Zhang (University of Arkansas)
Reinforcement LearningTabularFinance Related
🎯 What it does: This paper proposes a reinforcement learning framework based on a causal perspective, measuring long-term fairness through the qualification gain gap and conducting three causal decompositions: direct, indirect, and spurious effects.
A CLIP-Powered Framework for Robust and Generalizable Data Selection
Suorong Yang (Nanjing University), Furao Shen (Nanjing University)
OptimizationData-Centric LearningConvolutional Neural NetworkTransformerContrastive LearningImageMultimodality
🎯 What it does: This paper proposes a CLIP-based multimodal data selection framework that can select the most representative and diverse samples from large-scale datasets before training, significantly reducing storage and computational costs while improving model performance.
A Closer Look at Machine Unlearning for Large Language Models
Xiaojian Yuan (University of Science and Technology of China), Min Lin (University of Science and Technology of China)
OptimizationTransformerLarge Language ModelText
🎯 What it does: This paper studies the machine forgetting techniques of large language models (LLMs), proposes new evaluation metrics, and provides improvement plans for both targetless and targeted forgetting methods, validating their effectiveness in various scenarios.
A Coefficient Makes SVRG Effective
Yida Yin (University of California Berkeley), Zhuang Liu (Meta AI Research)
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper studies how to make SVRG effective in training deep neural networks, proposing a new optimization method α-SVRG by adding a tunable multiplier α to the variance reduction term of SVRG and providing a linear decay scheduling scheme, with empirical validation on various network architectures and datasets.
A Common Pitfall of Margin-based Language Model Alignment: Gradient Entanglement
Hui Yuan (Princeton University), Liu Leqi (University of Texas at Austin)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: The paper addresses the commonly used marginal preference optimization loss in human feedback in reinforcement learning (RLHF), revealing the fundamental issue of gradient entanglement that causes the model to simultaneously increase or decrease the probabilities of selected and rejected responses when improving the margin.
A Computational Framework for Modeling Emergence of Color Vision in the Human Brain
Atsunobu Kotani (University of California), Ren Ng (University of California)
RecognitionRepresentation LearningConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes a complete computational framework that simulates the generation of retinal signals by the eyes and reconstructs color perception in the brain solely through self-supervised prediction of neural signals, naturally inferring color dimensions.
A Conditional Independence Test in the Presence of Discretization
Boyang Sun (Mohamed bin Zayed University of Artificial Intelligence), Kun Zhang (Carnegie Mellon University)
TabularFinance Related
🎯 What it does: A conditional independence (CI) test method for discretized observational data (DCT) is proposed, which can accurately recover the CI relationships of potential continuous variables even after the variables have been discretized.
A Decade's Battle on Dataset Bias: Are We There Yet?
Zhuang Liu (Meta AI Research), Kaiming He (Meta AI Research)
ClassificationData-Centric LearningConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: Conducted systematic experiments on the 'Name That Dataset' task using modern deep networks on large-scale, diverse datasets, verifying that the network can identify the dataset to which an image belongs with high accuracy and analyzing the learned bias features.
A Deep Generative Learning Approach for Two-stage Adaptive Robust Optimization
Aron Brenner (Massachusetts Institute of Technology), Saurabh Amin (Massachusetts Institute of Technology)
OptimizationAuto EncoderTabularTime Series
🎯 What it does: The AGRO algorithm is proposed, embedding the Variational Autoencoder (VAE) into a two-stage Adaptive Robust Optimization (ARO) framework, achieving refinement and realism of uncertainty sets through adversarial generation in the latent space.
A deep inverse-mapping model for a flapping robotic wing
Hadar Sharvit (Hebrew University of Jerusalem), Tsevi Beatus (Hebrew University of Jerusalem)
Robotic IntelligenceRecurrent Neural NetworkTransformerTime SeriesSequential
🎯 What it does: A deep inverse mapping model for wing flaps of flapping robots has been developed, capable of predicting the required wing kinematic inputs based on the desired aerodynamic forces;
A Differentiable Rank-Based Objective for Better Feature Learning
Krunoslav Lehman Pavasovic (Meta), Levent Sagun (Meta)
Representation LearningImageTextTabular
🎯 What it does: A differentiable rank-based objective, difFOCI, is proposed for feature learning, feature selection, regularization, and fairness constraints.
A Distributional Approach to Uncertainty-Aware Preference Alignment Using Offline Demonstrations
Sheng Xu (Chinese University of Hong Kong), Guiliang Liu (Chinese University of Hong Kong)
OptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningAgentic AITabular
🎯 What it does: A MAP objective learning distributed reward model based on Beta prior is proposed on an offline preference-labeled dataset, utilizing a distributed Bellman operator and CVaR optimization for risk-sensitive strategies, achieving enhanced safety in robot control and LLM alignment.
A Formal Framework for Understanding Length Generalization in Transformers
Xinting Huang (Saarland University), Michael Hahn (Saarland University)
TransformerSequential
🎯 What it does: This paper proposes a formal framework for the theoretical analysis of the length generalization ability of Transformers when faced with longer sequences that were not seen during training. It introduces the 'Limit Transformer' and uses C-RASP to formally prove that under an idealized inference procedure, when the target function can be expressed by these models, Transformers can achieve length generalization.
A General Framework for Off-Policy Learning with Partially-Observed Reward
Rikiya Takehi (Waseda University), Yuta Saito (Cornell University)
Recommendation SystemReinforcement LearningVideoTabular
🎯 What it does: This paper proposes a hybrid strategy optimization method called HyPeR, which utilizes auxiliary rewards from dense observations to reduce the variance of offline policy gradient estimates and achieve effective offline policy learning in situations where the target reward is under-observed.
A General Framework for Producing Interpretable Semantic Text Embeddings
Yiqun Sun (National University of Singapore), Jun Yu (Harbin Institute of Technology)
RetrievalExplainability and InterpretabilityComputational EfficiencyLarge Language ModelContrastive LearningText
🎯 What it does: A general framework called CQG-MBQA is proposed, which generates yes/no questions through Conversational Question Generation (CQG) and produces interpretable text embeddings using a Multi-task Binary Question Answering (MBQA) model.
A Generalist Hanabi Agent
Arjun V Sudhakar, Sarath Chandar (Chandar Research Lab)
Recurrent Neural NetworkLarge Language ModelReinforcement LearningText
🎯 What it does: A general Hanabi agent R3D2 was constructed and trained, which can self-learn in games with 2 to 5 players and collaborate with partners using different algorithms.
A Generic Framework for Conformal Fairness
Aditya T. Vadlamani (Ohio State University), srinivasan parthasarathy
ClassificationGraph Neural NetworkGraphTabular
🎯 What it does: This paper proposes a general Conformal Fairness (CF) framework that utilizes hierarchical calibration and threshold search to control the disparity in confidence set coverage among different sensitive groups, thereby achieving fairness without requiring sensitive group information during inference.
A Geometric Framework for Understanding Memorization in Generative Models
Brendan Leigh Ross (Layer 6 AI), Gabriel Loaiza-Ganem (Layer 6 AI)
GenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: The 'Memory-Mapping Hypothesis (MMH) under a Geometric Framework' is proposed, which quantifies the degree of memorization of generative models using Local Intrinsic Dimension (LID) and distinguishes between memorization caused by overfitting and that caused by the data distribution itself.
A Graph Enhanced Symbolic Discovery Framework For Efficient Logic Optimization
Yinqi Bai (University of Science and Technology of China), Feng Wu
OptimizationExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationGraph Neural NetworkGraph
🎯 What it does: A CMO framework is proposed, which uses symbolic learning methods to learn lightweight, interpretable, and generalizable scoring functions in logic optimization, significantly improving the execution efficiency of LO.
A Large-scale Dataset and Benchmark for Commuting Origin-Destination Flow Generation
Can Rong (Tsinghua University), Yong Li (Tsinghua University)
GenerationData SynthesisRecommendation SystemGraph Neural NetworkGraphTabularBenchmark
🎯 What it does: This paper presents a large-scale commuting origin-destination (OD) flow dataset (LargeCommuingOD) covering 1,333 counties and 100 metropolitan areas in the United States, totaling 3,333 regions. It benchmarks existing physical models, classical machine learning models, graph neural network models, and graph generation models on this dataset, exploring the advantages of the graph generation-based network model (WEDAN) in the commuting flow generation task.
A Large-scale Training Paradigm for Graph Generative Models
Yu Wang (University of Oregon), Tyler Derr (Vanderbilt University)
GenerationData SynthesisGraph Neural NetworkTransformerDiffusion modelGraph
🎯 What it does: A large-scale training framework is proposed, training Large Graph Generative Models (LGGMs) that cover 13 domains and over 5000 graphs;
A Little Goes a Long Way: Efficient Long Context Training and Inference with Partial Contexts
Suyu Ge (University of Illinois Urbana-Champaign), Hao Peng (University of Illinois Urbana-Champaign)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: By integrating sparse attention with a full attention hybrid architecture on Llama-2 7B/70B, the context length is extended from 4K to 128K, while significantly reducing computational and KV cache costs during both training and inference phases.
A Meta-Learning Approach to Bayesian Causal Discovery
Anish Dhir (Imperial College London), Mark van der Wilk (University of Oxford)
Meta LearningTransformerGraphTabular
🎯 What it does: Using a Bayesian meta-learning approach, the BCNP model is constructed to directly learn the DAG posterior distribution from observational data and perform sampling.
A Multi-Power Law for Loss Curve Prediction Across Learning Rate Schedules
Kairong Luo (Tsinghua University), Wenguang Chen (Tsinghua University)
OptimizationLarge Language ModelText
🎯 What it does: A Multi-Power Law model is proposed to predict the loss curve during the pre-training process of large language models, and this model is used to automatically search for learning rate schedules that outperform traditional cosine scheduling.
A Multiscale Frequency Domain Causal Framework for Enhanced Pathological Analysis
Xiaoyu Cui (Shenzhen Institute of Advanced Technology), Jiandong Su (Shenzhen Institute of Advanced Technology)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: A multi-scale frequency domain causal framework (MFC-MIL) is proposed for whole slide image (WSI) classification, combining multi-scale spatial representation, frequency domain structural representation, and a causal memory intervention module.
A new framework for evaluating model out-of-distribution generalisation for the biochemical domain
Raul Fernandez-Diaz, Denis C. Shields (University College Dublin)
Drug DiscoveryBiomedical Data
🎯 What it does: This paper proposes and implements an OOD generalization evaluation framework for the biochemical field called HESTIA‑GOOD. It quantifies the model's performance under the target deployment distribution by constructing training/testing splits with different similarity thresholds and plotting the GOOD curve. Additionally, it introduces the AU‑GOOD metric to estimate expected performance. A CCPart partitioning algorithm that meets three major conditions is developed, along with a statistical testing method to compare AU‑GOOD. This framework can be applied to any biochemistry entity that can define a similarity function.
A New Perspective on Shampoo's Preconditioner
Depen Morwani (Harvard University), Lucas Janson (Harvard University)
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper systematically analyzes the preconditioner of the Shampoo optimizer from a theoretical perspective, proving that its square approximation is equivalent to the optimal Kronecker decomposition obtained from a single power iteration of the gradient covariance or Gauss-Newton matrix, and experimentally validates this approximation on various datasets and network architectures.
A Non-Contrastive Learning Framework for Sequential Recommendation with Preference-Preserving Profile Generation
Huimin Zeng (University of Illinois Urbana-Champaign), Dong Wang (University of Illinois Urbana-Champaign)
Recommendation SystemContrastive LearningSequential
🎯 What it does: This paper proposes the first Non-Contrastive Learning (NCL) framework, NCL-SR, which utilizes preferred maintained user profiles to train sequence recommendation models.
A Percolation Model of Emergence: Analyzing Transformers Trained on a Formal Language
Ekdeep Singh Lubana (Harvard University), Hidenori Tanaka (NTT Research)
GenerationData SynthesisExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper proposes a phenomenological definition of the concept of 'emergence' and demonstrates the stage-wise performance leaps that occur with increasing data scale by training a Transformer on a formal language based on PCFG and context-sensitive constraints.
A Periodic Bayesian Flow for Material Generation
Hanlin Wu (Tsinghua University), Jingjing Liu (Tsinghua University)
GenerationData SynthesisFlow-based ModelGraph
🎯 What it does: This paper proposes a periodic Bayesian flow model CrysBFN for the generation and prediction of periodic crystal structures.
A Policy-Gradient Approach to Solving Imperfect-Information Games with Best-Iterate Convergence
Mingyang Liu (Massachusetts Institute of Technology), Asuman E. Ozdaglar (Massachusetts Institute of Technology)
OptimizationReinforcement LearningSequential
🎯 What it does: This paper studies a strategy gradient-based algorithm (QFR) for solving two-player zero-sum games with incomplete information and proves that it can achieve optimal iterative convergence in self-play.
A Probabilistic Perspective on Unlearning and Alignment for Large Language Models
Yan Scholten (Technical University of Munich), Leo Schwinn (Technical University of Munich)
OptimizationTransformerLarge Language ModelText
🎯 What it does: A probabilistic evaluation framework for large language models is proposed, which directly assesses the output distribution rather than relying solely on point estimates from greedy decoding, and this framework is applied to the evaluation of machine forgetting and model alignment.
A Quantum Circuit-Based Compression Perspective for Parameter-Efficient Learning
Chen-Yu Liu (National Taiwan University), Min-Hsiu Hsieh (Foxconn Research Institute)
CompressionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a Quantum Circuit-based Parameter Adaptation (QPA) method to generate the trainable parameters required for parameter-efficient fine-tuning (PEFT) of large language models, significantly reducing the amount of trainable parameters during fine-tuning.
A Riemannian Framework for Learning Reduced-order Lagrangian Dynamics
Katharina Friedl (KTH Royal Institute of Technology), Danica Kragic (KTH Royal Institute of Technology)
Auto EncoderPhysics Related
🎯 What it does: This study investigates a method for learning low-order dynamic models of high-dimensional Lagrangian systems using Riemannian geometric networks, capable of achieving dimensionality reduction and long-term prediction of high-dimensional systems while maintaining physical consistency.
A Robust Method to Discover Causal or Anticausal Relation
Yu Yao (University of Sydney), Tongliang Liu (University of Sydney)
ClassificationAnomaly DetectionAuto EncoderImageTabular
🎯 What it does: This study investigates a robust method for determining whether a dataset exhibits a causal relationship (X→Y) or an inverse causal relationship (Y→X), particularly in the context of high-dimensional perceptual data and scenarios with label noise.
A Sanity Check for AI-generated Image Detection
Shilin Yan (Xiaohongshu Inc.), Weidi Xie (Shanghai Jiao Tong University)
ClassificationRecognitionConvolutional Neural NetworkMixture of ExpertsImageBenchmark
🎯 What it does: This paper presents the Chameleon dataset for authenticity verification in the task of detecting AI-generated images, and proposes the AIDE model for detecting AI-generated images.
A Second-Order Perspective on Model Compositionality and Incremental Learning
Angelo Porrello (University of Modena and Reggio Emilia), Rita Cucchiara (University of Modena and Reggio Emilia)
Representation LearningSupervised Fine-TuningImage
🎯 What it does: This study investigates the theory and algorithms for achieving modular composability and incremental learning through second-order Taylor expansion on pre-trained networks.
A Simple Approach to Unifying Diffusion-based Conditional Generation
Xirui Li (Shanghai Jiao Tong University), Ming-Hsuan Yang (Google DeepMind)
GenerationPose EstimationDepth EstimationDiffusion modelImage
🎯 What it does: A unified diffusion model called UniCon is proposed, capable of performing various image-conditioned tasks such as controllable generation, estimation, and joint generation within the same model.
A Simple Framework for Open-Vocabulary Zero-Shot Segmentation
Thomas Stegmüller (École Polytechnique Fédérale de Lausanne), Jean-Philippe Thiran (Centre Hospitalier Universitaire Vaudois)
Object DetectionSegmentationTransformerContrastive LearningImageText
🎯 What it does: Proposes the SimZSS framework, aligning a frozen self-supervised visual Transformer with a trained text encoder to achieve open vocabulary zero-shot segmentation.
A Simple yet Effective $\Delta\Delta G$ Predictor is An Unsupervised Antibody Optimizer and Explainer
Lirong Wu (Westlake University), Stan Z. Li (Westlake University)
OptimizationExplainability and InterpretabilityKnowledge DistillationDrug DiscoveryTransformerBiomedical Data
🎯 What it does: This paper proposes a lightweight ∆∆G prediction model called Light-DDG, which serves as an unsupervised antibody optimizer and interpreter, completing antibody mutation screening and optimization.
A Single Goal is All You Need: Skills and Exploration Emerge from Contrastive RL without Rewards, Demonstrations, or Subgoals
Grace Liu (Princeton University), Benjamin Eysenbach (Princeton University)
Robotic IntelligenceReinforcement LearningContrastive Learning
🎯 What it does: A reinforcement learning algorithm is proposed that explores using only a single target state, where the target state serves as the goal for all data collection and evaluation in Contrastive RL, demonstrating its ability to achieve complex manipulation and navigation tasks without rewards, examples, or sub-goals.
A Skewness-Based Criterion for Addressing Heteroscedastic Noise in Causal Discovery
Yingyu Lin (University of California San Diego), Biwei Huang (University of California San Diego)
Score-based Model
🎯 What it does: This paper proposes a discrimination criterion based on the skewness of scores for causal discovery under the heteroscedastic symmetric noise model (HSNM), and designs the SkewScore algorithm based on this criterion.
A Solvable Attention for Neural Scaling Laws
Bochen Lyu (University of Southampton), Zhanxing Zhu (University of Southampton)
TransformerOrdinary Differential Equation
🎯 What it does: This paper constructs a Multi-Task Sparse Feature Regression (MSFR) problem to study the training dynamics of linear self-attention blocks in context learning and provides an approximate analytical solution.
A Spark of Vision-Language Intelligence: 2-Dimensional Autoregressive Transformer for Efficient Finegrained Image Generation
Liang Chen (Peking University), Baobao Chang (Alibaba Group)
GenerationData SynthesisTransformerVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a 2D autoregressive Transformer (DnD-Transformer) that introduces a depth dimension beyond the spatial dimension, allowing for the prediction of multiple layers of discrete codes in a single forward inference, thus addressing the issues of information loss and high computational cost in VQ-VAE.
A Statistical Approach for Controlled Training Data Detection
Zirui Hu (Nanyang Technological University), Dacheng Tao (Nanyang Technological University)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes a new method for detecting training data leakage in large language models, called KTD, which can identify training samples while ensuring a preset false discovery rate (FDR);
A Statistical Framework for Ranking LLM-based Chatbots
Siavash Ameli (University of California), Michael W. Mahoney (University of California)
Large Language ModelText
🎯 What it does: A statistical framework for LLM chatbots is proposed, capable of fitting, ranking, and conducting association analysis on thousands of pairwise comparison data.
A Stochastic Approach to the Subset Selection Problem via Mirror Descent
Dan Greenstein (Technion Israel Institute of Technology), Nadav Hallak (Technion Israel Institute of Technology)
OptimizationTransformerSupervised Fine-TuningImage
🎯 What it does: A subset selection algorithm based on random distribution and mirror descent is proposed, providing two distributions (fixed size and expected size) along with unbiased gradient estimation and convergence proof.
A Theoretical Analysis of Self-Supervised Learning for Vision Transformers
Yu Huang (University of Pennsylvania), Yingbin Liang (Ohio State University)
TransformerAuto EncoderContrastive Learning
🎯 What it does: This paper conducts a theoretical analysis to study the training dynamics of two self-supervised learning objectives, Masked Autoencoder (MAE) and Contrastive Learning (CL), on a single-layer softmax-based Vision Transformer. It provides convergence guarantees and qualitatively characterizes the final attention patterns.
A Theoretical Framework for Partially-Observed Reward States in RLHF
Chinmaya Kausik (University of Michigan), Ambuj Tewari (University of Michigan)
Recommendation SystemReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: A partially observable reward state (PORRL) framework is proposed to more realistically model the internal states and intermediate feedback in RLHF. Within this framework, two types of algorithms are designed: model-based optimistic algorithms POR-UCRL and POR-UCBVI, as well as a model-based GOLF that improves its Bellman-Eluder dimension to α-HABE dimension. Finally, a white-box dimensionality reduction method is provided to convert adversarial feedback (dueling) into cardinal feedback.
A Theoretical Perspective: How to Prevent Model Collapse in Self-consuming Training Loops
Shi Fu (Nanyang Technological University), Dacheng Tao (Nanyang Technological University)
GenerationData SynthesisOptimizationTransformer
🎯 What it does: This paper introduces the concept of recursive stability, provides an upper bound on the generalization error of self-consistent training loops (STL), and extends the theory to the context learning scenario of Transformers, analyzing the trade-offs of synthetic data augmentation.
A Theoretically-Principled Sparse, Connected, and Rigid Graph Representation of Molecules
Shih-Hsin Wang (University of Utah), Bao Wang (University of Utah)
Drug DiscoveryGraph Neural NetworkPoint CloudGraph
🎯 What it does: A SCHull graph construction method is proposed, which achieves sparse, connected, and rigid guarantees for molecular graphs by projecting onto the unit sphere and constructing a convex hull.
A Theory for Token-Level Harmonization in Retrieval-Augmented Generation
Shicheng Xu (University of Chinese Academy of Sciences), Xueqi Cheng (University of Chinese Academy of Sciences)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: A theoretical framework based on distribution fusion is proposed to explain the token-level benefits and risks in Retrieval-Augmented Generation (RAG), and based on this, the Tok-RAG method is introduced to achieve parallel token generation of pure LLM and RAG to retain benefits and avoid risks.
A Theory of Initialisation's Impact on Specialisation
Devon Jarvis (University of the Witwatersrand), Stefano Sarao Mannelli (Chalmers University of Technology and University of Gothenburg)
Auto EncoderGenerative Adversarial NetworkTabular
🎯 What it does: This paper studies the impact of network initialization on neuron specialization and further explores how specialization determines the performance of catastrophic forgetting and interference in continual learning.
A Tight Convergence Analysis of Inexact Stochastic Proximal Point Algorithm for Stochastic Composite Optimization Problems
Shulan Zhu (Tsinghua University), Yancheng Yuan (Hong Kong Polytechnic University)
OptimizationTabular
🎯 What it does: This paper studies the stochastic proximal point algorithm (isPPA) for approximately solving subproblems in stochastic composite optimization problems, providing its stability, almost sure convergence, and the convergence rate of the final iteration under local Lipschitz and quadratic growth conditions, proving the tightness of the convergence rate and validating it through experiments.
A Training-Free Sub-quadratic Cost Transformer Model Serving Framework with Hierarchically Pruned Attention
Heejun Lee (Korea Advanced Institute of Science and Technology), Sung Ju Hwang (Korea Advanced Institute of Science and Technology)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes a training-independent Hierarchical Pruning Attention (HiP) mechanism to accelerate the inference of pre-trained Transformer models on long contexts (millions of tokens).
A Transfer Attack to Image Watermarks
Yuepeng Hu (Duke University), Neil Zhenqiang Gong (Duke University)
Adversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: In the no-detection API (no-box) environment, a transfer attack based on a multi-template watermark model is proposed, successfully bypassing the watermark detector of AI-generated images by adding slight perturbations to the watermark image.
A transfer learning framework for weak to strong generalization
Seamus Somerstep (University of Michigan), Yuekai Sun (MIT IBM Watson AI Lab)
Large Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This study addresses the weak-to-strong generalization problem, modeling it as transfer learning, proving the limitations of traditional fine-tuning, and proposing a refinement method based on ICL resampling to achieve generalization.
A Truncated Newton Method for Optimal Transport
Mete Kemertas (University of Toronto), Allan Douglas Jepson
OptimizationReinforcement LearningImage
🎯 What it does: An efficient GPU parallel EOT solver based on the truncated Newton method is proposed.
A Unified Framework for Forward and Inverse Problems in Subsurface Imaging using Latent Space Translations
Naveen Gupta (Virginia Tech), Anuj Karpatne (UNC at Chapel Hill)
Image TranslationGenerationAuto EncoderImage
🎯 What it does: A unified GFI framework is proposed, under which two novel networks—Latent U-Net and Invertible X-Net—are designed to simultaneously address the forward and inverse problems of subsurface imaging.
A Unified Theory of Quantum Neural Network Loss Landscapes
Eric Ricardo Anschuetz
Physics Related
🎯 What it does: This paper studies the loss landscape of quantum neural networks (QNN) at random initialization and proves its convergence to a Wishart process, thereby providing a unified theoretical framework.
A Watermark for Order-Agnostic Language Models
Ruibo Chen (University of Maryland), Heng Huang (University of Maryland)
GenerationData SynthesisProtein Structure PredictionLarge Language ModelTextSequentialBiomedical Data
🎯 What it does: The PATTERN-MARK watermarking framework is proposed, specifically designed for unordered decoding language models, generating key sequences through Markov chains and enhancing the probability of specific vocabulary during generation, followed by using statistical pattern detection to recover the watermark;
A-Bench: Are LMMs Masters at Evaluating AI-generated Images?
Zicheng Zhang (Shanghai Jiaotong University), Guangtao Zhai (Shanghai Jiaotong University)
RecognitionGenerationTransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: A diagnostic benchmark called A-Bench is proposed to evaluate the semantic understanding and visual quality perception capabilities of large multimodal models (LMMs) when assessing AI-generated images (AIGI).
A3D: Does Diffusion Dream about 3D Alignment?
Savva Victorovich Ignatyev, Evgeny Burnaev (Advanced Research Institute)
GenerationData SynthesisDiffusion modelScore-based ModelNeural Radiance FieldTextPoint Cloud
🎯 What it does: By embedding multiple 3D objects and their transitions in a common latent space and jointly training NeRF with text prompts, a set of structurally aligned 3D objects can be generated at once, supporting subsequent mixing and structure-preserving deformations.
ACC-Collab: An Actor-Critic Approach to Multi-Agent LLM Collaboration
Andrew Estornell (ByteDance Research), Yang Liu (University of California)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: The ACC-Collab framework is proposed, which jointly trains two roles of LLMs (Actor and Critic) to collaboratively solve problems through multi-turn dialogue.
Accelerated Over-Relaxation Heavy-Ball Method: Achieving Global Accelerated Convergence with Broad Generalization
Jingrong Wei (University of California), Long Chen (University of California)
OptimizationOrdinary Differential Equation
🎯 What it does: An accelerated over-relaxed heavy ball method (AOR-HB) is proposed, which achieves global accelerated convergence for general smooth strong convex problems.
Accelerated training through iterative gradient propagation along the residual path
Erwan Fagnou (Universite Paris Dauphine PSL), Alexandre Allauzen (ESPCI PSL)
OptimizationComputational EfficiencyRecurrent Neural NetworkTransformerImageText
🎯 What it does: A parallel iterative backpropagation algorithm named Highway-BP is proposed, which approximates standard backpropagation using residual paths to accelerate the training of deep sequential models.
Accelerating 3D Molecule Generation via Jointly Geometric Optimal Transport
Haokai Hong (Hong Kong Polytechnic University), KC Tan
GenerationDrug DiscoveryFlow-based ModelAuto EncoderGraphOrdinary Differential Equation
🎯 What it does: The GOAT framework is proposed, which achieves 3D molecular generation through joint geometric optimal transport (including atomic coordinates and atomic features). It constructs a straight transport path using flow matching and autoencoders, significantly accelerating sampling and improving generation quality.
Accelerating Auto-regressive Text-to-Image Generation with Training-free Speculative Jacobi Decoding
Yao Teng (University of Hong Kong), Xihui Liu (University of Hong Kong)
GenerationData SynthesisImageText
🎯 What it does: Proposes Speculative Jacobi Decoding (SJD), a training-free probabilistic parallel decoding algorithm designed to accelerate autoregressive text-to-image generation models.
Accelerating Diffusion Transformers with Token-wise Feature Caching
Chang Zou (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
GenerationComputational EfficiencyTransformerDiffusion modelImage
🎯 What it does: Implemented token-level feature caching for the diffusion Transformer to accelerate inference.
Accelerating Goal-Conditioned Reinforcement Learning Algorithms and Research
Michał Bortkiewicz (Warsaw University of Technology), Benjamin Eysenbach (Princeton University)
Robotic IntelligenceReinforcement LearningContrastive LearningSequentialBenchmark
🎯 What it does: We propose JaxGCRL, an extremely fast GPU-accelerated codebase and benchmark for self-supervised goal-conditioned reinforcement learning (GCRL), which quickly trains and evaluates algorithms across 8 continuous control environments.
Accelerating Inference of Retrieval-Augmented Generation via Sparse Context Selection
Yun Zhu (Google DeepMind), Jindong Chen (Google)
GenerationRetrievalComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: The Sparse RAG framework is proposed, which significantly reduces the input length and latency of retrieval-augmented generation by massively pre-filling the KV cache during a single forward inference and focusing only on highly relevant documents during decoding.
Accelerating neural network training: An analysis of the AlgoPerf competition
Priya Kasimbeg (Google DeepMind), George E. Dahl (Google DeepMind)
OptimizationComputational EfficiencyTransformerSupervised Fine-TuningTabularTime SeriesBenchmark
🎯 What it does: This paper organizes and analyzes the results of the ALGOPERF training algorithm competition, evaluating the acceleration effects of various training algorithms on fixed hardware.