GDrag:Towards General-Purpose Interactive Editing with Anti-ambiguity Point Diffusion
Xiaojian Lin (Sun Yat-Sen University), Xiaodan Liang
CodeDiffusion modelImage
π― What it does: A general task-aware point drag image editing framework GDrag has been developed, achieving more precise interactive editing by defining atomic tasks and utilizing ambiguity-resistant dense trajectories and adaptive motion supervision.
π― What it does: A generative agent called GenDataAgent is proposed to generate synthetic data in real-time during the training process, aimed at enhancing image classification training data and improving model generalization and fairness.
General Scene Adaptation for Vision-and-Language Navigation
Haodong Hong (University of Queensland), Qi Wu (University of Adelaide)
CodeDomain AdaptationRobotic IntelligenceTransformerLarge Language ModelVision Language ModelTextMultimodality
π― What it does: This paper proposes the GSA-VLN task and its evaluation dataset GSA-R2R, exploring how visual-language navigation agents achieve adaptation in persistent scenes through continuous memory and parameter updates.
π― What it does: A method is proposed to generate a high-quality 3D human Gaussian model (HGM) that can be rendered from any viewpoint using a single portrait photo.
Generalizable Motion Planning via Operator Learning
Sharath Matada (University of California San Diego), Nikolay Atanasov (University of California San Diego)
CodeOptimizationRobotic IntelligenceGraph
π― What it does: This paper proposes the Planning Neural Operator (PNO), which achieves generalized motion planning for different environments and target locations by learning the solution operator of the Eikonal PDE.
Generalization v.s. Memorization: Tracing Language Modelsβ Capabilities Back to Pretraining Data
Xinyi Wang (University of California), William Yang Wang (University of California)
CodeTransformerLarge Language ModelText
π― What it does: This paper quantifies the distributed memory and generalization degree of large language models in different tasks by constructing task grammar tables and task grammar language models, and tracks the source of their capabilities from pre-training corpora.
Generalization, Expressivity, and Universality of Graph Neural Networks on Attributed Graphs
Levi Rauchwerger (Technion - Israel Institute of Technology), Ron Levie (Technion - Israel Institute of Technology)
CodeGraph Neural NetworkGraph
π― What it does: A pseudo-metric for attributed graphs is proposed, making the graph space compact under this metric, and proving that Graph Neural Networks (MPNN) are continuous and can separate different graphs under this metric, thereby demonstrating the universal approximation and generalization bounds of MPNN.
π― What it does: This paper proposes Generalized Consistency Trajectory Models (GCTM), which enables one-time conversion between any two distributions, extending the original Consistency Trajectory Models (CTM) framework that could only convert from Gaussian noise to data.
Generating CAD Code with Vision-Language Models for 3D Designs
Kamel Alrashedy (Georgia Institute of Technology), Matthew Gombolay (Georgia Institute of Technology)
CodeGenerationOptimizationAI Code AssistantTransformerLarge Language ModelVision Language ModelTextPoint CloudBenchmarkChain-of-Thought
π― What it does: The CADCodeVerify method is proposed, which utilizes Vision-Language models to automatically generate and iteratively refine CAD scripts, enabling the generation and improvement of 3D designs without human intervention.
Generating Likely Counterfactuals Using Sum-Product Networks
JiΕΓ NΔmeΔek (Czech Technical University), Jakub Marecek
CodeOptimizationExplainability and InterpretabilityTabularFinance Related
π― What it does: By embedding the Sum-Product Network (SPN) into a Mixed Integer Optimization (MIO) framework, a method for generating counterfactual explanations (LiCE) that can simultaneously optimize interpretability, similarity, sparsity, executability, and feasibility (high likelihood) is proposed.
Generative Adapter: Contextualizing Language Models in Parameters with A Single Forward Pass
Tong Chen (University of Washington), Hao Cheng (Microsoft)
CodeGenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: An adaptive method based on a generative adapter is proposed, which can encode any context into parameter updates with just one forward pass during inference, thereby quickly customizing large language models.
Generative Flows on Synthetic Pathway for Drug Design
Seonghwan Seo (Korea Advanced Institute of Science and Technology), Woo Youn Kim (Korea Advanced Institute of Science and Technology)
CodeDrug DiscoveryFlow-based ModelTabular
π― What it does: This work proposes RXNFLOW, a synthetic pathway generation framework based on generative flow networks, which utilizes a massive chemical building block module and reaction templates to construct synthesizable drug molecules.
Fan Wu (University of Illinois Urbana-Champaign), Varun Chandrasekaran (University of Illinois Urbana-Champaign)
CodeGenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: This paper defines and systematically evaluates the phenomenon of 'generative monoculture' in large language models, experimentally testing the lack of diversity in book reviews and code generation tasks, and attempts various generation parameters and prompts to mitigate this issue.
CodeGenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsTextRetrieval-Augmented Generation
π― What it does: This paper proposes a unified model for text embedding and generationβGRIT (Generative Representational Instruction Tuning), which addresses the previous issue of focusing solely on a single task by simultaneously learning embedding tasks and generation tasks on the same large language model.
CodeGenerationRobotic IntelligenceLarge Language ModelDiffusion modelVideo
π― What it does: A generative world exploration framework called GenEx is constructed, based on a panoramic video diffusion model, enabling agents to generate continuous 360Β° perspectives through imagination and cognitive updates;
π― What it does: A language model-based speech enhancement framework called GenSE is proposed, which includes a single quantizer speech encoder SimCodec and a hierarchical denoising-generating two-stage process, maintaining speaker timbre consistency through token chain prompts.
GeoILP: A Synthetic Dataset to Guide Large-Scale Rule Induction
Si Chen (Beihang University), Xu Zhang (National Computer Network Emergency Response Technical Team Coordination Center of China)
CodeImage
π― What it does: This paper presents GeoILP, a large-scale synthetic geometric reasoning dataset designed to evaluate and advance research in large-scale ILP methods.
π― What it does: This study investigates the geometric evolution of neural networks in the input space, proposes the Geometric Invariance Hypothesis (GIH), and explores the impact of architecture and data on geometry and generalization.
π― What it does: This paper proposes Geometry Image Diffusion (GIMDiffusion), a method that combines geometric images with collaborative control to generate 3D models from text descriptions, capable of producing re-lightable and editable 3D assets while also outputting UV maps.
π― What it does: Analyzed the function space geometry of unnormalized (lightning) self-attention networks, described the fiber structure of parameter mappings, and derived the dimensionality formula for deep networks.
CodeTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelTextMultimodality
π― What it does: A multimodal large model named GeoX has been developed specifically for automatic geometric problem solving, achieving a unified input-output format through formalized visual-language pre-training.
GeSubNet: Gene Interaction Inference for Disease Subtype Network Generation
Ziwei Yang (Kyoto University), Jimeng Sun (University of Illinois Urbana-Champaign)
CodeGenerationData SynthesisDrug DiscoveryGraph Neural NetworkAuto EncoderGraphBiomedical Data
π― What it does: Proposed the GeSubNet framework, which combines patient gene expression with prior gene network learning to infer disease subtype-specific gene networks.
GETS: Ensemble Temperature Scaling for Calibration in Graph Neural Networks
Dingyi Zhuang (Massachusetts Institute of Technology), Jinhua Zhao (Massachusetts Institute of Technology)
CodeGraph Neural NetworkMixture of ExpertsGraph
π― What it does: A graph mixture of experts (MoE) based ensemble temperature scaling framework GETS is proposed for node-level uncertainty calibration of graph neural networks (GNNs).
π― What it does: A plugin method called GIFT is proposed in dataset distillation, which fully utilizes the complete information of soft labels and hard labels, significantly improving the training effectiveness of synthetic data.
Glimpse: Enabling White-Box Methods to Use Proprietary Models for Zero-Shot LLM-Generated Text Detection
Guangsheng Bao (Zhejiang University), Yue Zhang (Westlake University)
CodeGenerationOptimizationTransformerLarge Language ModelText
π― What it does: This paper proposes the Glimpse method, which estimates the complete word distribution using only the Top-K word probabilities, enabling white-box detection techniques to operate on private LLMs.
π― What it does: A benchmark called GLYCANML is proposed for evaluating machine learning models on various tasks related to glycan molecules, including classification, immunogenicity, glycosylation types, and protein-glycan interactions, covering both sequence and graph representations.
GMValuator: Similarity-based Data Valuation for Generative Models
Jiaxi Yang (University of British Columbia), Xiaoxiao Li (University of British Columbia)
CodeGenerationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningImageText
π― What it does: GMVALUATOR is proposed, a data value assessment method that is training-free and universally applicable to any generative model, transforming the data value problem into a similarity matching between generated samples and training samples.
GOFA: A Generative One-For-All Model for Joint Graph Language Modeling
Lecheng Kong (Washington University in St. Louis), Muhan Zhang (Peking University)
CodeGraph Neural NetworkTransformerLarge Language ModelGraph
π― What it does: A generative integrated graph foundation model called GOFA has been constructed, which embeds graph neural networks into large language models and achieves zero-shot inference on various graph tasks.
π― What it does: A dataset distillation loss based on Conditional Mutual Information (CMI) is proposed, utilizing CMI constraints in the feature space of a pre-trained network to reduce the class-aware complexity of synthetic data, resulting in a more learnable synthetic dataset.
π― What it does: A graph OOD detection framework called GOLD is proposed, which utilizes only ID data to self-generate pseudo OOD embeddings and discriminates them through energy scores.
π― What it does: A novel three-dimensional equivariant graph neural network, GotenNet, has been designed and implemented to efficiently and accurately capture the spatial structure and symmetry of molecular graphs.
GOttack: Universal Adversarial Attacks on Graph Neural Networks via Graph Orbits Learning
Zulfikar Alom (University of Manitoba), Cuneyt Gurcan Akcora (University of Central Florida)
CodeAdversarial AttackGraph Neural NetworkGraph
π― What it does: This paper proposes a global adversarial attack framework called Gottack, based on graph orbit learning, to induce structural perturbations in Graph Neural Networks (GNNs) for node classification tasks, leading to misclassification.
π― What it does: A two-stage unsupervised 3D object segmentation framework called GrabS is proposed, which first learns a generative object prior on a single-object dataset and then discovers and segments multiple objects in complex scenes through an embodied agent.
π― What it does: This paper studies the acceleration effect of Stochastic Nesterov Accelerated Gradient (SNAG) compared to Stochastic Gradient Descent (SGD) under convex or strongly convex objective functions using Gradient Autocorrelation (RACOGA).
Gradient descent with generalized Newtonβs method
Zhiqi Bu, Shiyun Xu (University of Pennsylvania)
CodeRecommendation SystemOptimizationTransformerLarge Language ModelImageTextAudio
π― What it does: A general second-order optimization method is proposedβGeneralized Newton's Method (GeN), which can automatically and dynamically determine the optimal learning rate for any optimizer (such as SGD, Adam) and accelerate convergence without the need for manual hyperparameter tuning.
π― What it does: The ECI sampling framework is proposed, utilizing a pre-trained flow matching model to achieve precise generation and regression of hard constraints such as PDEs in a zero-shot, gradient-free manner.
π― What it does: Proposed and implemented the GRAIN attack in the context of federated learning, which accurately recovers the graph structure and node features of graph neural networks by observing client gradients.
π― What it does: A dynamic workflow scheduling method for cloud computing environments, GOODRL, is proposed, which can perform online/offline learning and scheduling of workflows on heterogeneous clusters with the goal of minimizing average flow time.
Guibin Zhang (Tongji University), Shirui Pan (Griffith University)
CodeOptimizationComputational EfficiencyGraph Neural NetworkMixture of ExpertsGraph
π― What it does: To address the computational bottleneck of large-scale graphs, a Mixture of Graphs (MoG) method is proposed, which can dynamically select the most suitable sparsification expert for each node and mix the sparse subgraphs generated by it on the Grassmann manifold to obtain high-quality sparse graphs, while improving the inference speed and model performance of GNNs.
Xiang Cheng (Duke University), Suvrit Sra (Technical University of Munich)
CodeGraph Neural NetworkTransformerGraph
π― What it does: This paper demonstrates that by precisely configuring the weights of linear Transformers, they can solve classic graph algorithms on graph data, such as current flow, Laplacian inverse, square root inverse, heat kernel, and graph Laplacian eigenvectors.
GraphArena: Evaluating and Exploring Large Language Models on Graph Computation
Jianheng Tang (Hong Kong University of Science and Technology), Jia Li (Hong Kong University of Science and Technology)
CodeGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringGraphBenchmark
π― What it does: This paper proposes and implements GraphArena, a benchmark tool for graph computation problems, covering real-world graphs, ten multidimensional tasks, and a rigorous path-level evaluation process.
π― What it does: Proposes the GraphBridge framework, which constructs a two-stage pre-training + fine-tuning process, utilizing learnable input-output bridging and side networks (GSST, GMST) to achieve GNN transfer learning for any task and any domain;
GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
Tao Feng (University of Illinois at Urbana Champaign), Jiaxuan You (University of Illinois at Urbana Champaign)
CodeGraph Neural NetworkLarge Language ModelPrompt EngineeringTextGraph
π― What it does: Developed the GraphEval framework, which first uses a small LLM to decompose research ideas into viewpoint nodes, then constructs a viewpoint graph through BERT similarity or LLM relation extraction, and subsequently predicts paper review results using label propagation (LP) or graph neural networks (GNN), while incorporating novelty/plagiarism detection mechanisms.
GraphRouter: A Graph-based Router for LLM Selections
Tao Feng (University of Illinois Urbana Champaign), Jiaxuan You (University of Illinois Urbana Champaign)
CodeRecommendation SystemGraph Neural NetworkLarge Language ModelTextGraph
π― What it does: Designed and implemented a heterogeneous graph-based LLM router called GraphRouter, which recommends the optimal language model for user queries in multi-task scenarios.
GReaTer: Gradients Over Reasoning Makes Smaller Language Models Strong Prompt Optimizers
Sarkar Snigdha Sarathi Das (Pennsylvania State University), Rui Zhang (Salesforce Research)
CodeOptimizationTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
π― What it does: A method named GREATER is proposed, which uses small language models to optimize prompts through gradient optimization, without relying on large proprietary LLMs for feedback.
π― What it does: The MARBLE framework is proposed, which enhances the accuracy and generalization ability of PDE solutions based on Implicit Neural Representations (INR) through GridMix and spatial domain augmentation.
Grounding Multimodal Large Language Model in GUI World
Weixian Lei (National University of Singapore), Mike Zheng Shou (National University of Singapore)
CodeTransformerLarge Language ModelVision Language ModelImageTextMultimodality
π― What it does: This paper proposes an end-to-end GUI location positioning framework, which includes an automated data collection engine, a lightweight GUI Grounding model (AGG), and combines it with a multimodal large language model (MLLM) to build a visual agent capable of executing complex GUI tasks across various platforms.
Jiaqi Guan (University of Illinois), Jianzhu Ma (Tsinghua University)
CodeDrug DiscoveryDiffusion modelBiomedical Data
π― What it does: The GROUPBIND framework is proposed, which utilizes multiple ligands of the same protein pocket for molecular docking to enhance the accuracy of single-ligand docking.
Group-robust Sample Reweighting for Subpopulation Shifts via Influence Functions
Rui Qiao (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
CodeDomain AdaptationOptimizationText
π― What it does: This paper proposes a group label-based sample reweighting method called GSR, which utilizes a small number of group labels as the target set to iteratively optimize the sample weights of unlabeled data through influence functions, thereby enhancing the model's robustness to changes in subgroup distributions.
π― What it does: A framework utilizing two-dimensional Gaussian pulses for panoramic Gaussian splitting (GS-LiDAR) is proposed to generate realistic and controllable LiDAR point clouds, supporting perspective synthesis of dynamic scenes.
π― What it does: This paper proposes a model-free attack method based on geometric decision boundaries, GSBA K, which can generate imperceptible adversarial examples in a top-K manner for single-label multi-class and multi-label learning tasks.
GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models
Seyed Iman Mirzadeh, Mehrdad Farajtabar (Apple)
CodeLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: Two new mathematical reasoning benchmarks, GSM-Symbolic and GSM-NoOp, are proposed to systematically evaluate the reasoning performance of various LLMs under different instances, numerical variations, and additional irrelevant information.
Minbeom Kim (Seoul National University), Marc Dymetman (Independent Researcher)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
π― What it does: This paper proposes the GUARD framework, which combines autoregressive approximation during training with rejection sampling during inference to achieve constrained generation (guaranteed generation) for large language models.
π― What it does: A data-free resolution generation method is proposed, which combines Stable Diffusion with Score Identity Distillation (SiD) and Classifier-Free Guidance (CFG) to obtain a single-step text-to-image generator.
Gumbel Counterfactual Generation From Language Models
Shauli Ravfogel (New York University), Ryan Cotterell (ETH Zurich)
CodeGenerationLarge Language ModelText
π― What it does: This paper proposes to reconstruct language models as structural equation models using the Gumbel-max technique, thereby achieving string-based causal counterfactual generation.
Ziheng Chen (University of Trento), Nicu Sebe (University of Trento)
CodeClassificationRecognitionVideoGraph
π― What it does: A general GyroBN batch normalization framework is proposed, suitable for pseudo-reductive Gyrogroups, and implemented on Grassmannian and hyperbolic spaces;
CodeAnomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: The HADEMIF framework is proposed, utilizing two small networks (an interpretable D3T and MLP) to detect hallucinations in the output space and internal hidden states of LLMs, and calibrating logits through network outputs, achieving a unified process for hallucination detection and calibration.
HALL-E: Hierarchical Neural Codec Language Model for Minute-Long Zero-Shot Text-to-Speech Synthesis
Yuto Nishimura (University of Tokyo), Nakamasa Inoue (University of Tokyo)
CodeGenerationData SynthesisKnowledge DistillationTransformerLarge Language ModelAudio
π― What it does: Two technologies, MReQ and HALL-E, are proposed to achieve minute-level zero-shot text-to-speech (TTS) synthesis, breaking through the frame rate bottleneck of traditional LLM-TTS in long audio generation.
Halton Scheduler for Masked Generative Image Transformer
Victor Besnier (Valeo.ai), Matthieu Cord (Valeo.ai)
CodeGenerationTransformerImage
π― What it does: A new Halton scheduler is proposed for the token decoding order in the Masked Generative Image Transformers (MaskGIT) generation process, improving the sampling strategy;
Handling Delay in Real-Time Reinforcement Learning
Ivan Anokhin (Mila), Samira Ebrahimi Kahou (CIFAR AI Chair)
CodeReinforcement LearningSequential
π― What it does: This study investigates the delay problem in real-time reinforcement learning and proposes a network architecture that introduces temporal skip connections and historical observation enhancement within a parallel layer computing framework, validated through theoretical and experimental evidence of its advantages.
HARDMath: A Benchmark Dataset for Challenging Problems in Applied Mathematics
Jingxuan Fan (Harvard University), Michael Brenner
CodeLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: An automatically generated dataset HARDMATH containing 1060 advanced applied mathematics problems was created, and a subset HARDMATH-MINI (366 problems) along with 40 context-based semantic problems were constructed through manual verification.
HART: Efficient Visual Generation with Hybrid Autoregressive Transformer
Haotian Tang (Massachusetts Institute of Technology), Song Han (Massachusetts Institute of Technology)
CodeGenerationTransformerDiffusion modelImageText
π― What it does: The HART model is proposed, which combines autoregressive Transformers with lightweight residual diffusion, capable of directly generating high-quality images of 1024Γ1024 from text prompts.
Has the Deep Neural Network learned the Stochastic Process? An Evaluation Viewpoint
Harshit Kumar (Georgia Institute of Technology), Saibal Mukhopadhyay (Georgia Institute of Technology)
CodeConvolutional Neural NetworkRecurrent Neural NetworkAuto EncoderTime SeriesSequentialFinance Related
π― What it does: A new evaluation framework is proposed for assessing deep neural networks (DNN) in predicting the evolution of random complex systems, which includes the statistical ground truth (Statistic-GT) and a fidelity metric to stochastic processes (Fidelity to Stochastic Process, F2SP). It is demonstrated that the expected calibration error (ECE) is the only metric that can test F2SP based solely on a single observation.
HeadMap: Locating and Enhancing Knowledge Circuits in LLMs
Xuehao Wang (Southern University of Science and Technology), Yu Zhang (Tencent Technology Co., Ltd)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A layer-conditioned localization algorithm is proposed to identify knowledge circuits composed of key attention heads in large language models, and based on this, a parameter-efficient fine-tuning method called HeadMap is designed.
HELMET: How to Evaluate Long-context Models Effectively and Thoroughly
Howard Yen (Princeton University), Danqi Chen (Princeton University)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
π― What it does: The HELMET benchmark is proposed for the systematic evaluation of the multidimensional performance of Long Context Language Models (LCLMs).
Herald: A Natural Language Annotated Lean 4 Dataset
Guoxiong Gao (Peking University), Bin Dong (Peking University)
CodeGenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
π― What it does: A natural language to Lean 4 formal statement translation pipeline based on hierarchical retrieval enhancement is proposed, and the HeralD dataset is generated.
π― What it does: The HG-Adapter framework is proposed, which improves the generalization performance of pre-trained heterogeneous graph neural networks in downstream tasks through a dual-structure-aware adapter, contrastive loss with label propagation, and two types of self-supervised losses.
HiBug2: Efficient and Interpretable Error Slice Discovery for Comprehensive Model Debugging
Muxi Chen (Chinese University of Hong Kong), Qiang Xu (Chinese University of Hong Kong)
CodeClassificationObject DetectionPose EstimationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTransformerLarge Language ModelPrompt EngineeringImage
π― What it does: HiBug2 is proposed, an automated error slice discovery and model repair framework designed to enhance the robustness of visual models in real-world scenarios.
π― What it does: A two-stage robust watermarking method (WIND) based on the initial noise of the diffusion model is proposed, which can achieve watermark embedding and detection without affecting image quality.
π― What it does: A hierarchical uncertainty estimation framework is proposed, which can propagate the local uncertainty predicted by deep learning to the global transformation model and downstream tasks;
High-Dimensional Bayesian Optimisation with Gaussian Process Prior Variational Autoencoders
Siddharth Ramchandran (Aalto University), Harri LΓ€hdesmΓ€ki
CodeOptimizationDrug DiscoveryAuto EncoderTabular
π― What it does: Utilize GP prior VAE to learn a structured latent space and perform Bayesian Optimization in that space to efficiently search for the objective function in high-dimensional black-box optimization.
HiRA: Parameter-Efficient Hadamard High-Rank Adaptation for Large Language Models
Qiushi Huang (Southern University of Science and Technology), Yu Zhang (University of Surrey)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: The HiRA method is proposed, which combines pre-trained weights with low-rank matrices using the Hadamard product to achieve high-rank adaptation of parameters while maintaining the parameter and computational advantages of PEFT.
π― What it does: This paper presents HiSplat, a scalable 3D Gaussian rendering framework for sparse two-view scenarios, capable of simultaneously reconstructing large-scale structures and fine details through a multi-scale hierarchical structure.
HMoRA: Making LLMs More Effective with Hierarchical Mixture of LoRA Experts
Mengqi Liao (Beijing Jiaotong University), Huaiyu Wan (Beijing Jiaotong University)
CodeTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
π― What it does: A new LLM fine-tuning method called HMoRA is proposed, which combines LoRA experts with hierarchical mixed routing to achieve efficient multi-task learning.
Holistic Reasoning with Long-Context LMs: A Benchmark for Database Operations on Massive Textual Data
Seiji Maekawa (Megagon Labs), Nikita Bhutani (Megagon Labs)
CodeTransformerLarge Language ModelTextTabularBenchmarkChain-of-Thought
π― What it does: A benchmark called HoloBench is proposed for system evaluation of long-context language models' global reasoning capabilities on large-scale text data.
π― What it does: Designed and implemented HoloGNN, a method capable of pre-training task-agnostic node representations, which can quickly adapt to different order tasks (node-level, edge-level, and higher-order) on new tasks through a lightweight reduction map.
Hot-pluggable Federated Learning: Bridging General and Personalized FL via Dynamic Selection
Lei Shen (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
CodeFederated LearningSafty and PrivacyConvolutional Neural NetworkImage
π― What it does: A hot-plug federated learning framework (HPFL) is proposed, which utilizes a shared feature extractor and client-trained pluggable plugins to dynamically select the most suitable plugin during inference to enhance global generalization performance.
π― What it does: A hotspot-driven autoregressive generative model, PepHAR, is proposed for designing peptide bundles that meet geometric and interaction requirements under a given target protein.
How Does Critical Batch Size Scale in Pre-training?
Hanlin Zhang (Harvard University), Sham M. Kakade
CodeHyperparameter SearchTransformerLarge Language ModelText
π― What it does: This study investigates and quantifies the scaling laws of the critical batch size (CBS) in large-scale pre-training, conducting systematic experiments on autoregressive Transformers with parameters ranging from 85M to 1.2B, and derives empirical and theoretical models for CBS.
Kai Ye (University of Hong Kong), Chenxiong Qian (University of Hong Kong)
CodeImage
π― What it does: This paper proposes two metrics, Sharpness-Aware Learnability (SAL) and Unlearnable Distance (UD), by analyzing the model optimization process and loss landscape, to evaluate and compare the unlearnability of different Unlearnable Examples (UE) methods under single-task, multi-task, and various model architectures.
π― What it does: This study investigates how to utilize the Euclidean distance model to compress complex networks to very low dimensions without losing information, and provides efficient dimension search and verification methods.
How Much is Unseen Depends Chiefly on Information About the Seen
Seongmin Lee (Max Planck Institute for Security and Privacy), Marcel Boehme
CodeTextTabular
π― What it does: This paper derives the exact expected expression for missing mass under a multinomial distribution setting and proposes a search-based minimum mean square error (MSE) estimator design framework utilizing the dependency between frequencies, ultimately obtaining a distribution-specific estimator that outperforms Good-Turing through a genetic algorithm.
How to Probe: Simple Yet Effective Techniques for Improving Post-hoc Explanations
Siddhartha Gairola (Max Planck Institute for Informatics), Bernt Schiele (Institute of Science and Technology Austria)
CodeClassificationObject DetectionExplainability and InterpretabilityImage
π― What it does: This paper studies the impact of training details of the classification head (probe) of pre-trained models on the quality of post-hoc importance attribution methods (XAI). It finds that using binary cross-entropy (BCE) loss instead of cross-entropy (CE) significantly improves the class specificity and localization accuracy of the explanations. It also explores the use of a multi-layer interpretable MLP probe to further enhance classification performance and explanation localization.
π― What it does: To improve the quality of novel view synthesis in degraded scenarios, HQGS is proposed, which guides 3D Gaussian splatting through edge-semantic fusion and incorporates structural cosine similarity loss to compensate for detail loss and global structural inconsistency in low-quality images.
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: This paper proposes a hierarchical key-value sharing framework (HShare) that reduces the computational overhead of dynamic sparse selection and improves the decoding speed of LLMs by sharing key indices in the KV cache at the layer, head, and query levels.
Human-inspired Episodic Memory for Infinite Context LLMs
Zafeirios Fountas (Huawei), Jun Wang (University College London)
CodeGenerationRetrievalTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
π― What it does: This paper proposes EM-LLM, an architecture that embeds human episodic memory and event cognition mechanisms into existing large language models (LLMs), enabling them to handle nearly infinite context lengths without fine-tuning.
π― What it does: This paper proposes a hybrid regularization framework (HRDIS) that improves the inverse problem-solving process based on diffusion models through consistency regularization (CR) and noise mixing techniques.
π― What it does: A full hyper-surface convolutional neural network (HCNN) is proposed and implemented for representation learning of DNA sequences, operating directly in negatively curved space, avoiding the explicit mapping to evolutionary trees used in traditional methods.
HyPoGen: Optimization-Biased Hypernetworks for Generalizable Policy Generation
Hanxiang Ren (Zhejiang University), Yanchao Yang (Hong Kong University)
CodeOptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningSequential
π― What it does: This paper proposes an optimized bias hypernetwork, HyPoGen, which directly generates generalizable policy network parameters from task specifications without the need for demonstration data from the target task.
Hypothetical Minds: Scaffolding Theory of Mind for Multi-Agent Tasks with Large Language Models
Logan Cross (Stanford University), Nick Haber (Stanford University)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
π― What it does: A multi-agent system called Hypothetical Minds based on large language models has been developed, utilizing a Theory of Mind module to generate, evaluate, and improve hypotheses about the strategies of other agents in natural language, thereby driving high-level planning and sub-goal generation.
π― What it does: This paper proposes an identifiable latent dynamical system framework that can simultaneously recover latent states, nonlinear state transition functions, and observation mappings, identifying the true system dynamics from high-dimensional perception sequences.
IDInit: A Universal and Stable Initialization Method for Neural Network Training
Yu Pan (Huawei Noah's Ark Lab), Zenglin Xu (Fudan University)
CodeConvolutional Neural NetworkImage
π― What it does: A novel identity initialization method called IDInit is proposed, which can maintain the complete identity mapping of the backbone and sub-backbone in residual networks.