These 1682 ICLR 2025 papers come with a code repository. Each shows an AI one-line summary below β get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every ICLR 2025 paper, 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)
CodeTransformerTextReview/Survey Paper
π― 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.
π― What it does: A new gradient-based attack method, Ο-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)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelText
π― 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.
Lingwei Zhu (University of Tokyo), Martha White (University of Alberta)
CodeOptimizationReinforcement LearningTabular
π― 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.
π― 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.
π― 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.
3DMolFormer: A Dual-channel Framework for Structure-based Drug Discovery
Xiuyuan Hu (Tsinghua University), Xue Liu (McGill University)
CodeDrug DiscoveryTransformerSupervised Fine-TuningReinforcement LearningBiomedical Data
π― 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.
A Causal Lens for Learning Long-term Fair Policies
Jacob Lear (University of Arkansas), Lu Zhang (University of Arkansas)
CodeReinforcement LearningTabularFinance Related
π― 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 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)
CodeOptimizationTransformerLarge Language ModelText
π― 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 Common Pitfall of Margin-based Language Model Alignment: Gradient Entanglement
Hui Yuan (Princeton University), Liu Leqi (University of Texas at Austin)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― 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 Distributional Approach to Uncertainty-Aware Preference Alignment Using Offline Demonstrations
Sheng Xu (Chinese University of Hong Kong), Guiliang Liu (Chinese University of Hong Kong)
CodeOptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningAgentic AITabular
π― 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)
CodeTransformerSequential
π― 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 Producing Interpretable Semantic Text Embeddings
Yiqun Sun (National University of Singapore), Jun Yu (Harbin Institute of Technology)
CodeRetrievalExplainability and InterpretabilityComputational EfficiencyLarge Language ModelContrastive LearningText
π― 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.
π― 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.
π― 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 Multi-Power Law for Loss Curve Prediction Across Learning Rate Schedules
Kairong Luo (Tsinghua University), Wenguang Chen (Tsinghua University)
CodeOptimizationLarge Language ModelText
π― 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 new framework for evaluating model out-of-distribution generalisation for the biochemical domain
Raul Fernandez-Diaz, Denis C. Shields (University College Dublin)
CodeDrug DiscoveryBiomedical Data
π― What it does: This paper proposes and implements an OOD generalization evaluation framework for the biochemical field called HESTIAβGOOD. 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 AUβGOOD metric to estimate expected performance. A CCPart partitioning algorithm that meets three major conditions is developed, along with a statistical testing method to compare AUβGOOD. This framework can be applied to any biochemistry entity that can define a similarity function.
π― 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.
CodeGenerationData SynthesisExplainability and InterpretabilityTransformerLarge Language ModelText
π― 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 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)
CodeOptimizationReinforcement LearningSequential
π― 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.
π― 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.
π― 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.
CodeGenerationData SynthesisTransformerVision Language ModelImageTextMultimodality
π― 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)
CodeTransformerLarge Language ModelText
π― 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);
π― 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)
CodeGenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― 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.
π― 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.
CodeRecognitionGenerationTransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
π― 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).
ACC-Collab: An Actor-Critic Approach to Multi-Agent LLM Collaboration
Andrew Estornell (ByteDance Research), Yang Liu (University of California)
CodeTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― 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.
π― 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.
π― What it does: A NODE training method based on the variational formula (VF) is proposedβVF-NODE, which uses global integration to replace numerical ODE solvers, significantly reducing the number of function evaluations and suppressing autoregressive errors.
π― What it does: This paper proposes the NiNo (Neuron Interaction and Nowcasting Networks) method, which accelerates training by periodically predicting network parameters, significantly reducing the training steps of Adam.
ACES: Automatic Cohort Extraction System for Event-Stream Datasets
Justin Xu (University of Oxford), Matthew B.A. McDermott
CodeData-Centric LearningDrug DiscoveryTabularBiomedical DataElectronic Health Records
π― What it does: An automated queue extraction system called ACES has been developed, simplifying the definition and reproduction of machine learning tasks and queues in electronic health records (EHR).
π― What it does: The DeComFL algorithm is proposed, achieving dimension-independent communication in federated learning through zero-order optimization, reducing the communication volume per round from O(d) to O(1);
Oussama Boussif (Mila - Quebec AI Institute), Yoshua Bengio (Mila - Quebec AI Institute)
CodeReinforcement LearningSequential
π― What it does: This paper proposes an online action abstraction method called ACTIONPIECE, which can automatically extract and incorporate high-level actions ('chunks') during the training process, thereby reducing the sampling path length and enhancing the sample efficiency and pattern discovery capabilities of reinforcement learning and GFlowNet.
Activation Gradient based Poisoned Sample Detection Against Backdoor Attacks
Danni Yuan (Chinese University of Hong Kong), Baoyuan Wu (Chinese University of Hong Kong)
CodeAnomaly DetectionImage
π― What it does: This paper proposes a method for detecting contaminated samples in backdoor attacks based on the Gradient Circular Distribution (GCD), called AGPD.
Daniel Musekamp (University of Stuttgart), Mathias Niepert (University of Stuttgart)
CodeConvolutional Neural NetworkTime SeriesBenchmarkPhysics Related
π― What it does: This paper proposes AL4PDE, the first active learning benchmark framework for neural PDE solvers, aimed at automating the collection of high-quality training data and evaluating active learning methods.
Ada-K Routing: Boosting the Efficiency of MoE-based LLMs
Tongtian Yue (Institute of Automation Chinese Academy of Sciences), Jing Liu (Institute of Automation Chinese Academy of Sciences)
CodeOptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningMixture of ExpertsText
π― What it does: This paper proposes the Ada-K routing strategy, allowing each token to dynamically decide how many experts to activate, thereby achieving more efficient MoE inference.
π― What it does: A new adaptive second-order optimizer, AdaFisher, is proposed, which preprocesses gradients using a diagonal block Kronecker approximation of the Fisher information matrix, significantly improving the convergence speed and generalization performance of deep network training.
AdaIR: Adaptive All-in-One Image Restoration via Frequency Mining and Modulation
Yuning Cui (Technical University of Munich), Fahad Shahbaz Khan (Mohammed Bin Zayed University of AI)
CodeRestorationTransformerImage
π― What it does: A single model named AdaIR for full-scene image denoising/dehazing/rain removal/deblurring/low-light enhancement is proposed, capable of adaptively recognizing and processing different types of image distortions.
Adam Exploits $\ell_\infty$-geometry of Loss Landscape via Coordinate-wise Adaptivity
Shuo Xie (Toyota Technological Institute at Chicago), Zhiyuan Li (Toyota Technological Institute at Chicago)
CodeOptimizationTransformerLarge Language ModelText
π― What it does: This paper proposes a new convergence analysis method, exploring the advantages of the Adam optimization algorithm over SGD when training language models, particularly its performance under β β geometry.
Yushun Zhang (Chinese University of Hong Kong), Ruoyu Sun (Shenzhen Research Institute of Big Data)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: This paper proposes Adam-mini, an optimizer that reduces memory usage of the optimizer state by approximately 50% while maintaining performance close to AdamW. It achieves a significant reduction in the number of learning rates by dividing model parameters into several blocks based on the Hessian structure and using a single learning rate for each block (taking the mean of Adam's v).
Adapt-$\infty$: Scalable Continual Multimodal Instruction Tuning via Dynamic Data Selection
Adyasha Maharana (University of North Carolina), Mohit Bansal (University of North Carolina)
CodeOptimizationTransformerSupervised Fine-TuningVision Language ModelMultimodality
π― What it does: Proposes Adaptββ, which achieves lifelong multimodal instruction fine-tuning through dynamic data selection, significantly reducing forgetting and enhancing forward transfer.
π― What it does: The ADAPT framework is proposed for the task of continuous learning in panoramic segmentation, where only the decoder is updated at each step while the other modules remain frozen, achieving efficient and sustainable knowledge retention and learning of new tasks.
Adapters for Altering LLM Vocabularies: What Languages Benefit the Most?
HyoJung Han (University of Maryland), Huda Khayrallah (Amazon)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A vocabulary adaptation method based on adapters, VocADT, is proposed, which can integrate new vocabulary sets into LLMs without modifying the weights of the pre-trained model.
Adapting Multi-modal Large Language Model to Concept Drift From Pre-training Onwards
Xiaoyu Yang (Australian Artificial Intelligence Institute), En Yu (Australian Artificial Intelligence Institute)
CodeClassificationAnomaly DetectionTransformerLarge Language ModelMixture of ExpertsContrastive LearningImageTextMultimodality
π― What it does: This paper studies the robustness of multimodal large language models in long-tail and anomalous distribution drift environments, proposing a unified drift theory framework and designing a T-distribution adapter to mitigate the negative effects of tail drift and OOD drift.
π― What it does: An adaptive Q-network, AdaQN, is proposed, which maintains multiple sets of Q-networks with different hyperparameters during training and selects the network with the smallest error as the shared target during each target network update, achieving online hyperparameter tuning without increasing additional environment interactions.
Adaptive Pruning of Pretrained Transformer via Differential Inclusions
Yizhuo Ding (Fudan University), Yanwei Fu (Fudan University)
CodeRetrievalCompressionOptimizationTransformerLarge Language ModelImageText
π― What it does: This paper proposes a dynamic sparsification method based on differential inclusionβSolution Path Pruning (SPP), which can generate a family of Transformer compressed models with different sparsity levels within a single search phase.
π― What it does: Proposed the Adaptive Rank Allocation framework and RaNA adapter, which compresses the MLP and QKV linear layers of the Transformer using low-rank decomposition and adaptive routers to improve inference speed;
Adaptive Shrinkage Estimation for Personalized Deep Kernel Regression in Modeling Brain Trajectories
Vasiliki Tassopoulou (University of Pennsylvania), Christos Davatzikos (University of Pennsylvania)
CodeTime SeriesBiomedical DataAlzheimer's Disease
π― What it does: This paper proposes an adaptive shrinkage deep kernel regression framework that can predict the long-term trajectories of brain biomarkers in real-time based on a small amount of individual follow-up data.
CodeOptimizationData-Centric LearningReinforcement LearningMultimodalityBiomedical Data
π― What it does: This paper proposes a 'teacher-student' framework, where the teacher model generates sampling trajectories in high-loss areas to guide the student (GFlowNet) for incremental sampling, significantly improving exploration efficiency and mode coverage in multimodal distributions.
π― What it does: Proposed and implemented the Adversarial Diffusion Bridge Model (ADBM), which constructs an inverse bridge in the diffusion model to perform denoising defense against adversarial samples.
π― What it does: Fine-tuning a pre-trained diffusion model through reinforcement learning, adding new conditional controls to enable the model to generate samples based on additional labels.
Addressing Label Shift in Distributed Learning via Entropy Regularizationβ
Zhiyuan Wu (University of Oslo), Ali Ramezani-Kebrya (University of Oslo)
CodeOptimizationFederated LearningImage
π― What it does: This paper proposes the VRLS method, which improves the maximum likelihood estimation of label transfer through entropy regularization in distributed learning, allowing for a more accurate estimation of the importance ratio of test/training labels; it also combines IW-ERM in a multi-node environment to robustly handle label shifts for internal/external nodes.
π― What it does: A parameter-efficient fine-tuning method called ADePT is proposed, which utilizes short soft prompts and a shared shallow feedforward network to generate input embedding offsets.
π― What it does: A graph generation framework based on Beta diffusion, Graph Beta Diffusion (GBD), is proposed to address the challenge of generating graph data with mixed discrete and continuous features.
π― What it does: A method called MISA is proposed to enhance prompt-based models in general continuous learning (GCL) scenarios, primarily including forgetting-aware adaptation of prompt parameters (ISA-FAM) during the initial session and the use of a parameter-free logit mask at the output layer to mitigate catastrophic forgetting.
Advantage-Guided Distillation for Preference Alignment in Small Language Models
Shiping Gao (Sun Yat-sen University), Qifan Wang (Meta AI)
CodeKnowledge DistillationTransformerLarge Language ModelReinforcement LearningText
π― What it does: Two methods for preference alignment of small language models using aligned teacher models are proposed: Dual-Constrained Knowledge Distillation (DCKD) and Advantage-Guided Distillation for Preference Alignment (ADPA);
Xinhe Wang (University of Michigan), Jiaqi W. Ma (University of Illinois Urbana-Champaign)
CodeAdversarial AttackImageText
π― What it does: This study investigates the vulnerability of data attribution methods in value assessment and compensation, proposing two attack strategiesβShadow Attack (based on data distribution and shadow training) and Outlier Attack (based on black-box queries and outlier bias)βto make subtle perturbations to training data in order to enhance their own compensation share.
π― What it does: A framework for Vehicle Routing Problem (VRP) using Adversarial Generative Flow Networks (AGFN) is proposed, which directly constructs high-quality paths through the generator;
Zonglin Di (University of California), Yang Liu (Washington University in St. Louis)
CodeSafty and PrivacyAdversarial AttackConvolutional Neural NetworkTransformerImageText
π― What it does: A machine unlearning framework based on Stackelberg game is proposed, actively embedding Membership Inference Attacks (MIA) into the design of unlearning algorithms to eliminate the influence of specified training data.
Zhuoyi PENG, Yi Yang (Hong Kong University of Science and Technology)
CodeClassificationSafty and PrivacyData-Centric LearningTransformerGenerative Adversarial NetworkContrastive LearningImage
π― What it does: A machine forgetting framework called MixUnlearn based on adversarial mixup is proposed, which utilizes a generator to synthesize challenging samples by mixing forgotten samples with remaining samples, driving the model to achieve complete forgetting of sensitive data while retaining other knowledge without complete retraining, thus addressing the catastrophic forgetting problem.
π― What it does: Evaluated and demonstrated that adversarial perturbation-based art style protection tools cannot prevent style imitation by generative models, proposing and validating several low-cost robust imitation methods.
π― What it does: This paper proposes the SiDA framework, which incorporates adversarial loss into single-step diffusion model distillation, combining the Fisher divergence of SiD with the discriminator of Diffusion GAN.
π― What it does: This paper proposes a defense method based on Support Vector Machines (SVM) called FLORAL, designed to resist label poisoning attacks.
π― What it does: The COBRA method is proposed, which enhances the robustness of anomaly detection under adversarial attacks by self-supervising the generation of pseudo-anomalous samples and combining adversarial training with contrastive learning loss.
π― What it does: Proposes the AROS framework, which utilizes Lyapunov stability theory to improve the robustness of OOD detection under adversarial attacks.
π― What it does: This paper researches and implements a strategy robust learning method based on opponent incentives, aiming to replace traditional adversarial training.
π― What it does: A defense framework named ADVPAINT has been designed and implemented, which directly interferes with the self-attention and cross-attention mechanisms of the Stable Diffusion Inpainting model in image inpainting tasks using adversarial perturbations, thereby preventing malicious actors from replacing or inserting content in published images.
π― What it does: A steerable EquivarLayer is proposed to achieve equivariance for arbitrary input-output feature types on affine groups and their continuous subgroups, and it is applied as a canonicalization function in image classification tasks.
Jiayi Zhang (The Hong Kong University of Science and Technology), Chenglin Wu (DeepWisdom)
CodeOptimizationAI Code AssistantLarge Language ModelAgentic AITextBenchmark
π― What it does: By combining the code-represented LLM invocation nodes and edges that form a workflow with predefined operators, and utilizing MCTS with LLM as the optimizer, an agentic workflow is automatically generated and optimized, achieving fully automated workflow generation.
Agent S: An Open Agentic Framework that Uses Computers Like a Human
Saaket Agashe (Simular Research), Xin Eric Wang (Simular Research)
CodeTransformerLarge Language ModelAgentic AITextMultimodalityBenchmarkRetrieval-Augmented Generation
π― What it does: This study investigates a GUI agent, Agent S, based on a multimodal large language model, capable of autonomously completing multi-step computer tasks.
Agent Skill Acquisition for Large Language Models via CycleQD
So Kuroki (Sakana AI), Yujin Tang (Sakana AI)
CodeTransformerLarge Language ModelReinforcement LearningAgentic AIImageText
π― What it does: A CycleQD framework based on Quality Diversity (QD) and evolutionary algorithms is proposed to achieve multi-skill acquisition in LLMs through expert model fusion and SVD mutation.
Ao Li (Hong Kong University of Science and Technology), Yaliang Li (Alibaba Group)
CodeTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringText
π― What it does: Designed and implemented an AOP framework that uses meta-proxies to split, allocate, and evaluate tasks in a multi-agent system, achieving more efficient agent-oriented planning.
AgentOccam: A Simple Yet Strong Baseline for LLM-Based Web Agents
Ke Yang (University of Illinois Urbana-Champaign), Huzefa Rangwala (Amazon)
CodeTransformerLarge Language ModelAgentic AIText
π― What it does: A web task execution agent AGENTOCCAM based on a large language model has been designed, achieving strong performance in zero-shot scenarios by optimizing only the observation space and action space.
AgentRefine: Enhancing Agent Generalization through Refinement Tuning
Dayuan Fu (Beijing University of Posts and Telecommunications), Weiran Xu (Beijing University of Posts and Telecommunications)
CodeRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningAgentic AIText
π― What it does: The AgentRefine framework is proposed, which enhances the generalization ability of open-source LLMs by allowing them to self-correct erroneous actions through environmental feedback in generated diverse environments.
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: A multi-agent collaborative writing framework called AGENTS' ROOM has been developed, which gradually generates long stories using specialized planning and writing agents.
AgentSquare: Automatic LLM Agent Search in Modular Design Space
Yu Shang (Tsinghua University), Yong Li (Tsinghua University)
CodeTransformerLarge Language ModelAgentic AIPrompt EngineeringText
π― What it does: Designed and implemented an automated LLM agent search framework called AgentSquare based on a modular design space, which automatically discovers and optimizes LLM agents across six different benchmarks using module evolution, module recombination, and performance predictors.
AI2TALE: An Innovative Information Theory-based Approach for Learning to Localize Phishing Attacks
Van Nguyen (Monash University), Carsten Rudolph (Monash University)
CodeAnomaly DetectionExplainability and InterpretabilityText
π― What it does: This paper proposes a deep learning framework called AI2TALE, based on information theory and the information bottleneck, for locating and explaining phishing attacks in emails under weak supervision;
π― What it does: This paper proposes Air-DualODE, which utilizes a physics-guided dual neural ODE model to simultaneously model physical (diffusion-convection) dynamics and data-driven unknown dynamics in an open air quality system, achieving accurate PM2.5 predictions through fusion and alignment in the latent space.
CodeLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
π― What it does: By performing symbolic mutation on existing theorems in the Lean theorem proving environment (using the tactics rw and apply), approximately 6.3 million new formal theorems were generated, and these theorems were constructed into a large-scale corpus along with their corresponding proofs for continuous pre-training and supervised fine-tuning of LLMs, thereby enhancing their performance in theorem proving tasks.
π― What it does: This paper analyzes the algorithm stability of adversarial training and provides a general upper bound on the generalization error based on a perturbation mapping's scalability parameter.
π― What it does: In response to the application of diffusion models in visual perception tasks, a method is proposed to align the generative denoising process through learning objectives, training data, and interactive user interfaces, significantly improving the performance of depth estimation, referential image segmentation, and general perception tasks.
Aligning Language Models with Demonstrated Feedback
Omar Shaikh (Stanford University), Diyi Yang (Stanford University)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: The DITTO method is proposed, which directly generates preference comparison data for language models through a small number (<10) of user demonstrations, achieving personalized alignment.
π― What it does: This paper studies a continuous state space model capable of handling irregular time seriesβACSSM. It proposes efficient simulation and inference of posterior trajectories through multi-boundary Doobβs h-transformation and variational inference (SOC).
An Asynchronous Bundle Method for Distributed Learning Problems
Daniel Cederberg (Stanford University), Mikael Johansson (KTH)
CodeOptimizationFederated LearningTabular
π― What it does: An asynchronous Bundle method is proposed, utilizing the multi-tangent model of each worker node to solve subproblems on the parameter server, thereby achieving iteration updates in distributed learning without synchronization and without the need for maximum delay information.
An Auditing Test to Detect Behavioral Shift in Language Models
Leo Richter (University College London), Matt Kusner
CodeLarge Language ModelText
π― What it does: A sequential hypothesis testing method is proposed for monitoring changes in language model behavior, capable of detecting shifts in model behavior distribution over time under black-box access.
π― What it does: This paper proposes an efficient framework that estimates the Shapley values of contributors to diffusion models using sparse fine-tuning, aiming for fair attribution of data contributors.
π― What it does: This study proposes Neural Attention Memory Models (NAMMs), which adaptively prune the KV cache of the Transformer through evolutionary learning to enhance the performance and efficiency of long-context tasks.
An Information Criterion for Controlled Disentanglement of Multimodal Data
Chenyu Wang (Massachusetts Institute of Technology), Caroline Uhler (Massachusetts Institute of Technology)
CodeExplainability and InterpretabilityRepresentation LearningDrug DiscoveryContrastive LearningMultimodalityBenchmark
π― What it does: This paper proposes a self-supervised multimodal representation learning framework DISENTANGLEDSSL, aimed at decoupling shared information from modality-specific information to enhance interpretability and robustness.
An Intelligent Agentic System for Complex Image Restoration Problems
Kaiwen Zhu (Shanghai Jiao Tong University), Chao Dong (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
CodeRestorationTransformerLarge Language ModelSupervised Fine-TuningAgentic AIVision Language ModelImage
π― What it does: Proposes the AgenticIR agent system, which achieves perception-scheduling-execution-reflection-rescheduling in five stages through the interaction of LLM and VLM, dynamically calling a single degradation recovery tool to solve complex image restoration tasks.
An Undetectable Watermark for Generative Image Models
Sam Gunn (University of California Berkeley), Dawn Song (University of California Berkeley)
CodeGenerationData SynthesisDiffusion modelImage
π― What it does: The first undetectable watermark scheme (PRC watermark) is proposed, which can embed watermarks in the latent space of diffusion models without affecting image quality and achieve robust detection.
CodeOptimizationComputational EfficiencyTabularBenchmarkPhysics Related
π― What it does: ANaGRAM is proposed, a low-complexity natural gradient optimization method in PINNs, treating PINNs as least squares regression and proving that the natural gradient is equivalent to the Green's function;
AnalogGenie: A Generative Engine for Automatic Discovery of Analog Circuit Topologies
Jian Gao (Northeastern University), Xuan Zhang (Northeastern University)
CodeGenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningGraph
π― What it does: This paper presents AnalogGenieβa GPT-based generative engine for the automatic discovery of multi-type, scalable analog circuit topologies.
Analytic DAG Constraints for Differentiable DAG Learning
Zhen Zhang (Australian Institute for Machine Learning), Javen Qinfeng Shi (Australian Institute for Machine Learning)
CodeOptimizationGraph Neural NetworkGraphBiomedical Data
π― What it does: This paper proposes the use of analytic functions to construct differentiable DAG constraints, establishing a theoretical connection between analytic functions and DAG constraints. It utilizes the closure properties of analytic functions (differentiation, addition, multiplication) to design a series of high-order DAG constraints. Additionally, efficient algorithms for evaluating these constraints are provided and implemented within a path tracking optimization framework.