🎯 What it does: This paper studies a zero-shot talking head generation model GAIA, which decouples motion and appearance using VAE and utilizes a diffusion model to predict motion based on speech, thereby generating natural and diverse videos.
Gene Regulatory Network Inference in the Presence of Dropouts: a Causal View
Haoyue Dai (Carnegie Mellon University), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)
CodeBiomedical Data
🎯 What it does: To address the bias in gene regulatory network (GRN) inference caused by technical zeros (dropout) in single-cell RNA-seq, a 'Causal Dropout Model' based on causal graphs is proposed. This model deletes samples with all conditional variables set to zero during conditional independence (CI) testing (test-wise deletion), thereby restoring the same CI relationships as in the absence of dropout, ultimately achieving unbiased GRN structure learning.
🎯 What it does: A general graph random feature (g-GRF) algorithm based on random walks is proposed for unbiased estimation of arbitrary weighted adjacency matrix functions, thereby efficiently approximating the graph kernel matrix.
🎯 What it does: This study investigates the generalization ability of diffusion models with sufficiently large training sets and reveals the denoising mechanism induced by the internal Geometric Adaptive Harmonic Basis (GAHB).
Junlong Li (Shanghai Jiao Tong University), Pengfei Liu (Shanghai Jiao Tong University)
CodeGenerationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A generative evaluator AUTO-J with 13 billion parameters is proposed to assess the alignment performance of LLMs in multiple scenarios.
🎯 What it does: This paper proposes a general speech generation pre-training model named SpeechFlow and demonstrates its transfer performance in tasks such as denoising, speech separation, and zero-shot TTS.
GenSim: Generating Robotic Simulation Tasks via Large Language Models
Lirui Wang (Massachusetts Institute of Technology), Xiaolong Wang (University of California San Diego)
CodeRobotic IntelligenceTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Utilizing large language models (GPT-4, Code-Llama, etc.) to automatically generate rich robotic simulation tasks and expert demonstrations, and training multi-task visual-language control strategies based on these tasks to enhance the robot's generalization ability in new tasks.
🎯 What it does: This paper proposes a geographic location encoding method that combines Spherical Harmonics (SH) with Sinusoidal Representation Networks (SIREN) to efficiently learn representations of geographic coordinates on a global scale.
Get more for less: Principled Data Selection for Warming Up Fine-Tuning in LLMs
Feiyang Kang (Virginia Tech), Ruoxi Jia (Virginia Tech)
CodeDomain AdaptationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A pre-fine-tuning scheme is proposed, utilizing a large amount of unlabeled open data to select samples through the GOT-D method for lightweight preheating of pre-trained language models, followed by final fine-tuning on a small amount of labeled data, thereby improving task performance and reducing costs.
🎯 What it does: To address the issue of text-to-image diffusion models struggling to suppress negative targets, a method is proposed to remove unwanted content through soft-weighted regularization and optimization of text embeddings during inference.
GIO: Gradient Information Optimization for Training Dataset Selection
Dante Everaert (Amazon), Christopher Potts (Stanford University)
CodeOptimizationData-Centric LearningImageText
🎯 What it does: Proposes the Gradient Information Optimization (GIO) method, which selects a subset of training data by minimizing KL divergence in an unlabeled setting;
🎯 What it does: This paper proposes GNNCert, a verifiable robustness defense for graph classification tasks that ensures the prediction label remains unchanged when the structure and node features are perturbed a limited number of times.
GOAt: Explaining Graph Neural Networks via Graph Output Attribution
Shengyao Lu (University of Alberta), Di Niu (University of Alberta)
CodeExplainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: The Graph Output Attribution (GOAt) method is proposed for local interpretation of pre-trained Graph Neural Networks (GNNs), decomposing model outputs into scalar products and attributing them based on input node/edge features.
🎯 What it does: Proposes a definition of functionally equivalent features, explains the complexity of network features using category theory, and based on this, introduces the Iterative Feature Merging (IFM) algorithm to evaluate and compress networks;
Oscar Sainz (University of the Basque Country), Eneko Agirre (University of the Basque Country)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Introducing GoLLIE, a large language model (LLM) designed for information extraction tasks, specifically for zero-shot learning through refined annotation guidelines;
🎯 What it does: The GPAvatar framework is proposed, capable of reconstructing an animatable 3D head avatar from one or more images in a single forward inference, achieving precise control over expressions and poses.
GPT-4 Is Too Smart To Be Safe: Stealthy Chat with LLMs via Cipher
Youliang Yuan (Chinese University of Hong Kong), Zhaopeng Tu (Tencent AI Lab)
CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The study verifies whether existing safety alignment mechanisms can be bypassed under non-natural language input by using various ciphers in dialogue with large language models (LLMs), and proposes a systematic evaluation framework called CipherChat.
🎯 What it does: Using neural networks (NNP) to learn and predict the potential energy of molecular conformations, and employing its gradients for energy minimization, improves traditional physics-based conformational optimization.
🎯 What it does: A gradient descent-based hard-axis alignment decision tree ensemble method called GRANDE is proposed, and end-to-end training is achieved using dense representation and straight-through operations;
🎯 What it does: Unified encoding of neural network parameters and architectures into a graph structure (neural graph), and using graph neural networks/Transformers for invariant/homogeneous learning, thereby achieving generalization for networks of different architectures;
🎯 What it does: This paper proposes an adaptive graph pooling framework based on graph parsing (Graph Parsing Network, GPN), which can learn personalized pooling trees for each graph and achieve efficient node aggregation.
Graph Transformers on EHRs: Better Representation Improves Downstream Performance
Raphael Poulain (University of Delaware), Rahmatollah Beheshti (University of Delaware)
CodeClassificationRepresentation LearningDrug DiscoveryGraph Neural NetworkTransformerTabularTime SeriesSequentialBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes a hybrid model named GT-BEHRT, which utilizes a graph Transformer to generate embeddings for each visit, and then captures the temporal relationships in the patient's visit sequence through a BERT encoder, resulting in more robust patient representations for various prediction tasks.
GraphCare: Enhancing Healthcare Predictions with Personalized Knowledge Graphs
Pengcheng Jiang (University of Illinois Urbana-Champaign), Jimeng Sun (University of Illinois Urbana-Champaign)
CodeGraph Neural NetworkLarge Language ModelTabularBiomedical DataElectronic Health Records
🎯 What it does: Construct a personalized knowledge graph and use a dual attention-enhanced graph neural network for multiple medical predictions for patients.
🎯 What it does: A graphical multi-output Gaussian process (GMOGP) framework is proposed, which dynamically learns the conditional dependencies between outputs through probabilistic graphical models and attention mechanisms, supports non-Gaussian priors, and achieves interpretable output associations.
GraphPulse: Topological representations for temporal graph property prediction
Kiarash Shamsi (University of Manitoba), Cuneyt Gurcan Akcora (University of Central Florida)
CodeRecurrent Neural NetworkGraph Neural NetworkGraphTime Series
🎯 What it does: A framework named GraphPulse is proposed for predicting future attributes of time-evolving graphs, combining Mapper to generate topological summaries and using RNN for sequence modeling.
Grounding Multimodal Large Language Models to the World
Zhiliang Peng (University of Chinese Academy of Sciences), Furu Wei (Microsoft Research)
CodeGenerationRetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: A multimodal large language model KOSMOS-2 with visual localization and text reference capabilities has been designed and trained. It can directly output the corresponding image area coordinates when generating text and supports users in pointing to objects using bounding boxes or referring expressions.
GTA: A Geometry-Aware Attention Mechanism for Multi-View Transformers
Takeru Miyato (University of Tübingen), Andreas Geiger (University of Tübingen)
CodeTransformerImage
🎯 What it does: Geometric Perception Attention (GTA) is introduced in the Transformer, applying relative geometric transformations to queries, keys, and values to achieve natural alignment in multi-view scenes.
🎯 What it does: An end-to-end model GTMGC based on Graph Transformer is proposed, which directly predicts the 3D ground state conformation from the 2D topological structure of molecules.
Guess & Sketch: Language Model Guided Transpilation
Celine Lee (Cornell University), Alexander M Rush
CodeAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: A framework for assembly code translation called GUESS & SKETCH, which combines neural networks and symbolic reasoning, is proposed to automatically convert assembly programs between ARMv8 and RISC-V.
🎯 What it does: Through the Masked Autoencoder (MAE) framework of self-supervised learning, we reconstruct 12-lead electrocardiograms (ECGs) using spatio-temporal patching to learn general ECG representations and fine-tune them for downstream disease screening tasks.
Harnessing Explanations: LLM-to-LM Interpreter for Enhanced Text-Attributed Graph Representation Learning
Xiaoxin He (National University of Singapore), Bryan Hooi (National University of Singapore)
CodeClassificationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextGraph
🎯 What it does: This paper proposes a TAPE framework that enriches the node features of Text Attribute Graphs (TAG) using explanation texts generated by large language models (LLM) and inputs them into Graph Neural Networks (GNN) for node classification.
🎯 What it does: This paper proposes an adaptive image de-raining method called CoIC, which is based on joint rain/detail perception representation. It can train CNN/Transformer models on mixed multi-source datasets and achieve dynamic adaptation to different rain intensities/background details through instance-level modulation.
Headless Language Models: Learning without Predicting with Contrastive Weight Tying
Nathan Godey (Inria), Benoît Sagot (Inria)
CodeTransformerLarge Language ModelContrastive LearningText
🎯 What it does: A pre-training method without a language model projection head is proposed—Headless Language Modeling, which directly learns word embedding representations using Contrastive Weight Tying.
Hebbian Learning based Orthogonal Projection for Continual Learning of Spiking Neural Networks
Mingqing Xiao (Peking University), Zhouchen Lin (Peking University)
CodeSpiking Neural NetworkImage
🎯 What it does: A Hebbian learning-based orthogonal projection method (HLOP) is proposed, which achieves online extraction of the neuronal activity subspace through lateral recurrent connections, thereby protecting old task knowledge during continual learning and avoiding catastrophic forgetting.
Heterogeneous Personalized Federated Learning by Local-Global Updates Mixing via Convergence Rate
Meirui Jiang (Chinese University of Hong Kong), Qi Dou (Chinese University of Hong Kong)
CodeFederated LearningBiomedical Data
🎯 What it does: The LG-Mix method is proposed, which dynamically mixes local and global updates through the NTK convergence rate to achieve personalized federated learning under heterogeneous features.
Hiding in Plain Sight: Disguising Data Stealing Attacks in Federated Learning
Kostadin Garov (INSAIT), Martin Vechev (ETH Zurich)
CodeFederated LearningSafty and PrivacyAdversarial AttackImage
🎯 What it does: This paper proposes an attack framework called SEER that utilizes malicious servers to steal data from large batches of gradients through a secret decoder in federated learning.
Hierarchical Context Merging: Better Long Context Understanding for Pre-trained LLMs
Woomin Song (KAIST), Jinwoo Shin (KAIST)
CodeRetrievalComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A training-independent hierarchical context merging (HOMER) scheme is proposed, which expands the context window of LLMs while maintaining computational efficiency by merging layer by layer and performing token pruning on each block before merging.
🎯 What it does: A hierarchical graph generation network HiGen is proposed, which can first generate community subgraphs in parallel and then predict inter-community edges, thus achieving a coarse-to-fine hierarchical graph generation.
🎯 What it does: The HoloNet framework is proposed, realizing a trainable spectral convolutional network on directed graphs, breaking free from the limitations of traditional graph Fourier transforms.
Jiahai Feng (University of California Berkeley), Jacob Steinhardt (University of California Berkeley)
CodeTransformerLarge Language ModelText
🎯 What it does: The study investigates how language models bind entities to attributes in context and identifies a general internal mechanism called binding ID.
How Does Unlabeled Data Provably Help Out-of-Distribution Detection?
Xuefeng Du (University of Wisconsin-Madison), Yixuan Li (University of Wisconsin-Madison)
CodeAnomaly DetectionImage
🎯 What it does: Proposes the SAL framework, which first filters candidate anomaly samples from unlabeled field data, and then trains a binary classification OOD detector using these samples along with labeled ID samples.
🎯 What it does: This paper proposes a method to automatically convert linear inequality constraints into a differentiable constraint layer and embed it into various deep generative models (GAN, CTGAN, TableGAN, TVAE, GOGGLE) to generate synthetic tabular data that satisfies the constraints.
How to Catch an AI Liar: Lie Detection in Black-Box LLMs by Asking Unrelated Questions
Lorenzo Pacchiardi (University of Oxford), Jan M. Brauner (University of Oxford)
CodeClassificationTransformerLarge Language ModelText
🎯 What it does: A lie detection method is proposed that only uses black-box access, based on asking the model a set of yes/no questions unrelated to lies, and then using logistic regression to discriminate the response patterns.
🎯 What it does: We constructed and released the largest 3D CT dataset, AbdomenAtlas 1.1 (9,262 CT scans, 25 anatomical structures, and 7 tumor pseudo-labels), and trained and publicly released a series of supervised pre-trained models called SuPreM using this dataset. We then evaluated their transfer learning performance on various 3D medical image segmentation and classification tasks.
Tom Hosking (University of Edinburgh), Max Bartolo (Cohere)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: By conducting human evaluations on the text generated by large language models (LLMs), this study explores whether a single preference score can cover all important types of errors and analyzes potential biases and confounding factors in the annotations.
🎯 What it does: This paper proposes three combination methods based on pre-trained motion diffusion models: sequential combination (DoubleTake) for generating actions of arbitrary length; parallel combination (ComMDM) for achieving two-person interactive actions; and model combination (DiffusionBlending) for multi-dimensional fine-grained control.
Hybrid Directional Graph Neural Network for Molecules
Junyi An (Nanjing University), Furao Shen (Nanjing University)
CodeDrug DiscoveryGraph Neural NetworkGraph
🎯 What it does: This paper proposes the Hybrid Directional Graph Neural Network (HDGNN), which enhances molecular property prediction performance by integrating strict equivariant operations with learnable modules to balance equivariance and expressive power.
🎯 What it does: Proposes the Hybrid Distill framework, which utilizes both CL/supervised teacher and MIM teacher to guide the student model's learning;
Hybrid LLM: Cost-Efficient and Quality-Aware Query Routing
Dujian Ding (University of British Columbia), Ahmed Hassan Awadallah
CodeTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: A quality-aware hybrid large language model inference framework is proposed, which allocates 'easy' queries to small models and 'difficult' queries to large models through a router, in order to reduce costs while maintaining high quality;
🎯 What it does: A generative self-supervised learning task based on hyperedge filling is designed, and the corresponding pre-training framework HYPEBOY is implemented to enhance the performance of hypergraph neural networks in node classification and hyperedge prediction tasks.
Haoyue Bai (University of Wisconsin Madison), Yixuan Li (University of Wisconsin Madison)
CodeDomain AdaptationRepresentation LearningImage
🎯 What it does: Proposes the HYPO framework, which learns domain-invariant representations on the unit hypersphere to reduce intra-class domain variation and maximize inter-class separation.
IDEAL: Influence-Driven Selective Annotations Empower In-Context Learners in Large Language Models
Shaokun Zhang (Pennsylvania State University), Tongliang Liu (University of Sydney)
CodeClassificationOptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes an influence-driven selective labeling method called IDEAL, which is used to select a small subset of examples for manual labeling from a large pool of unlabeled data, in order to reduce labeling costs in in-context learning (ICL) and improve model performance.
Sehyun Kwon (Seoul National University), Kangwook Lee (University of Wisconsin-Madison)
CodeClassificationRecognitionTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageText
🎯 What it does: A text-conditioned image clustering method called IC TC is proposed, allowing users to directly control clustering results through natural language descriptions.
🎯 What it does: This paper proposes an image clustering pipeline CPP that utilizes features from the large-scale pre-trained model CLIP, combined with the principle of Maximum Coding Rate Reduction (MCR).
🎯 What it does: This paper proposes the use of Probabilistic Circuits to guide the sampling process of diffusion models, achieving high-quality image inpainting.
Impact of Computation in Integral Reinforcement Learning for Continuous-Time Control
Wenhan Cao (Tsinghua University), Wei Pan (University of Manchester)
CodeOptimizationReinforcement LearningTime Series
🎯 What it does: This study investigates the impact of using different numerical integration methods (trapezoidal rule and Bayesian quadrature) on control performance during the policy evaluation phase of Integral Reinforcement Learning (IntRL) under continuous-time control, and provides convergence rate analysis and simulation validation.
ImplicitSLIM and How it Improves Embedding-based Collaborative Filtering
Ilya Shenbin (Steklov Mathematical Institute of Russian Academy of Sciences), Sergey Nikolenko (Steklov Mathematical Institute of Russian Academy of Sciences)
CodeRecommendation SystemTabular
🎯 What it does: This paper proposes ImplicitSLIM, an unsupervised learning method that extracts embeddings from SLIM-like models for collaborative filtering.
Improved algorithm and bounds for successive projection
Jiashun Jin (Carnegie Mellon University), Jingming Wang (Harvard University)
CodeOptimizationSupervised Fine-TuningPoint Cloud
🎯 What it does: An improved vertex finding algorithm, pp-SPA, is proposed to address the performance degradation of traditional SPA under strong noise or outliers.
🎯 What it does: An improved probabilistic image-text matching model PCME++ is proposed, which captures the many-to-many relationships and uncertainties brought by label sparsity through probabilistic embedding.
🎯 What it does: A unified perspective on path space is proposed, utilizing time-reversed controlled diffusion processes to achieve sampling from prior distributions to target distributions, and extending it to arbitrary path space divergences; within this framework, a new log-variance loss is introduced to address issues such as mode collapse and high variance associated with traditional KL loss.
🎯 What it does: A general framework for using holographic equipotential propagation (hEP) under the condition of no weight symmetry is proposed, and the gradient estimation bias is reduced through Jacobian homeostasis.
🎯 What it does: A protein fitness landscape smoothing method based on graph signal processing is proposed, and the GGS algorithm is constructed in conjunction with Gibbs sampling for protein optimization under limited noise data.
🎯 What it does: This paper proposes a lightweight regularization based on optimal transport (sliced Wasserstein) to improve the convergence and rendering quality of dynamic NeRF, achieving this without the need for additional networks or preprocessing, making it easy to integrate into any existing architecture.
🎯 What it does: This paper proposes an image deformation method called IMPUS based on diffusion models, which achieves smooth, direct, and realistic transitions between two images through linear interpolation in the optimized CLIP text embedding space and latent space, combined with probabilistic flow ODE.
In-context Autoencoder for Context Compression in a Large Language Model
Tao Ge (Microsoft Corporation), Furu Wei (Microsoft Corporation)
CodeGenerationCompressionTransformerLarge Language ModelPrompt EngineeringAuto EncoderText
🎯 What it does: Designed and implemented the In-Context Autoencoder (ICAE), which uses a lightweight LoRA encoder to compress long contexts into a small number of memory slots, allowing the original LLM to be used as a decoder to complete generation tasks directly with these slots.
🎯 What it does: This paper experimentally explores the Bayesian perspective on in-context learning (ICL) in large-scale language models, extending to a hierarchical ICL (HMICL) setting with multi-task mixing. It verifies whether Transformers can approximate Bayesian predictors across various families of linear and nonlinear functions (such as linear regression, sparse regression, neural networks, decision trees, Fourier series, quadratic monomials, GMMs, etc.) and studies their generalization and simplification bias on unseen tasks.
🎯 What it does: This paper proposes a plugin framework named CAPABOOST, which implements parallel multi-branching of weights by using random binary masks for sparse masking of shared weights in the PEFT module, thereby enhancing the effective rank and capacity of the model without increasing parameters or FLOP.
🎯 What it does: The Incremental Randomized Smoothing (IRS) method is proposed, which can quickly re-verify the robustness of an already certified DNN after approximate modifications such as quantization or pruning.
Inducing High Energy-Latency of Large Vision-Language Models with Verbose Images
Kuofeng Gao (Tsinghua University), Wei Liu (Tencent Data Platform)
CodeGenerationComputational EfficiencyAdversarial AttackTransformerVision Language ModelImageText
🎯 What it does: A method for energy consumption and latency attacks on large-scale visual-language models (VLM) has been designed and implemented—verbose images, which induce the model to generate extremely long text sequences, significantly increasing energy consumption and latency during inference.
🎯 What it does: This study investigates backdoor attacks on semantic segmentation models and proposes the Influencer Backdoor Attack (IBA), which induces the model to misclassify target categories by injecting small triggers on non-target pixels.
🎯 What it does: Proposes the principle of information retention and designs a three-stage supervised learning framework InfoR‑LSF to simultaneously learn main features and supplementary features, aiming to retain as much relevant information as possible.
Inherently Interpretable Time Series Classification via Multiple Instance Learning
Joseph Early (University of Southampton), Niall Twomey (Amazon Prime Video)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkTime Series
🎯 What it does: A framework called MILLET is proposed, which integrates Multiple Instance Learning (MIL) into Time Series Classification (TSC), allowing the model itself to provide interpretable point predictions, thereby enhancing interpretability.
🎯 What it does: By inserting a pluggable HyperNet module into the NeRF framework, dynamic generation of scene-specific network weights is achieved using reference scene features, thereby enabling the generalization of NeRF.
InstructCV: Instruction-Tuned Text-to-Image Diffusion Models as Vision Generalists
Yulu Gan (Peking University), Ahmed Alaa
CodeObject DetectionSegmentationGenerationDepth EstimationTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelImageMultimodality
🎯 What it does: By mapping text instructions to image generation tasks, a unified visual task framework based on natural language was achieved using the stable diffusion model;
🎯 What it does: A reset-free reinforcement learning algorithm named RISC is proposed, which can intelligently switch between the forward task controller and the rollback controller without frequently resetting the environment, thereby collecting experiences more efficiently and improving learning speed.
🎯 What it does: A method called InCo Aggregation is proposed, which utilizes internal cross-layer gradients in federated learning to enhance the similarity of deep client models, significantly improving model performance in heterogeneous device environments.
InternVid: A Large-scale Video-Text Dataset for Multimodal Understanding and Generation
Yi Wang (OpenGVLab, Shanghai AI Laboratory), Yu Qiao (OpenGVLab, Shanghai AI Laboratory)
CodeRecognitionGenerationRetrievalTransformerLarge Language ModelVision Language ModelDiffusion modelContrastive LearningVideoTextMultimodality
🎯 What it does: A large-scale multimodal dataset, InternVid, was constructed, which has a high correspondence between video and text. The ViCLIP vision-language model was trained on this dataset and further applied to tasks such as action recognition, video retrieval, text-to-video generation, and video dialogue.
Matthieu Blanke (Inria Paris), Marc Lelarge (Inria Paris)
CodeExplainability and InterpretabilityComputational EfficiencyRobotic IntelligenceMeta LearningTime SeriesSequentialPhysics Related
🎯 What it does: This paper proposes an interpretable meta-learning framework suitable for learning physical systems in multiple environments (CAMEL), which achieves rapid adaptation and physical parameter identification through task-related low-dimensional linear context parameters.
Interpreting Robustness Proofs of Deep Neural Networks
Debangshu Banerjee (University of Illinois Urbana-Champaign), Gagandeep Singh (VMware Research)
CodeExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: This paper proposes the ProFIt method for analyzing the robustness proofs of deep neural networks, helping humans understand the content of the proofs.
🎯 What it does: This paper studies how to perform data attribution on diffusion models (DDPM and Stable Diffusion), proposing an improved version of the TRAK method called D-TRAK, and conducts large-scale experiments on various datasets.
🎯 What it does: The LaDID framework is proposed, which learns the latent dynamics of high-dimensional observations through variational autoencoders and Transformers, supporting continuous time prediction and cross-system generalization.
INViTE: INterpret and Control Vision-Language Models with Text Explanations
Haozhe Chen (Columbia University), Chengzhi Mao
CodeExplainability and InterpretabilityAdversarial AttackTransformerVision Language ModelImage
🎯 What it does: This paper proposes the INViTE framework, which disables self-attention in the Transformer and retains only local operations, mapping each potential token to the CLS layer, and then provides natural language explanations using CLIP's text retrieval.
🎯 What it does: This paper proposes an image resampling method based on implicit continuous representation (IRAD), which utilizes SampleNet to automatically predict pixel-level offsets to counter adversarial attacks during the inference phase.
iTransformer: Inverted Transformers Are Effective for Time Series Forecasting
Yong Liu (Tsinghua University), Mingsheng Long (Tsinghua University)
CodeTransformerTime Series
🎯 What it does: A method called 'iTransformer' is proposed, which captures multivariate correlations using attention in the Transformer structure by reversing the dimensions of time series data (treating the entire sequence of each variable as a token) and learning sequence representations with a feedforward network, while keeping the original Transformer components unchanged.
JoMA: Demystifying Multilayer Transformers via Joint Dynamics of MLP and Attention
Yuandong Tian (Meta), Simon Shaolei Du
CodeTransformerLarge Language ModelText
🎯 What it does: The JoMA framework is proposed, which combines the self-attention of Transformers with the training dynamics of MLP layers, deriving the implicit dynamics of self-attention and explaining the mechanism by which multi-layer Transformers learn hierarchical structures.
Knowledge Card: Filling LLMs' Knowledge Gaps with Plug-in Specialized Language Models
Shangbin Feng (University of Washington), Yulia Tsvetkov (University of Washington)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: The KNOWLEDGE CARD framework is proposed, which utilizes small language models (knowledge cards) trained from different fields and sources, along with a three-layer knowledge selector, to dynamically inject external knowledge during inference, enhancing the factuality and timeliness of large models.
🎯 What it does: This paper proposes a temperature scaling method applied only to the teacher side for knowledge distillation—Transform Teacher Matching (TTM) and its sample adaptive version (WTTM), and proves its equivalence to traditional KD plus R'‑enyi entropy regularization, thereby enhancing the generalization ability of the student model.
Fanqi Wan (Sun Yat-sen University), Shuming Shi (Tencent AI Lab)
CodeKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: A knowledge fusion framework called FUSELLM is proposed, which utilizes the probability distributions of various structurally different LLMs for lightweight continuous training of the target LLM, thereby transferring the capabilities of each source model into a single model.
KoLA: Carefully Benchmarking World Knowledge of Large Language Models
Jifan Yu (Tsinghua University), Juanzi Li (Tsinghua University)
CodeTransformerLarge Language ModelTextBenchmark
🎯 What it does: A multi-level evaluation benchmark KoLA aimed at global knowledge has been constructed to systematically assess the four cognitive abilities of LLM: knowledge memory, understanding, application, and creation.
L2MAC: Large Language Model Automatic Computer for Extensive Code Generation
Samuel Holt (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
CodeGenerationAI Code AssistantTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Designed and implemented L2MAC, a storage program computing framework based on large language models, for generating long texts and code, instantiated as Code-L2MAC, capable of gradually building a complete large-scale codebase without being constrained by traditional context window limits.
🎯 What it does: This paper proposes a learning-driven preprocessing method (Learning to Presolve, L2P) that predicts optimal preprocessing parameters (priority, maximum iterations, temporal mask) for each mixed-integer programming instance and integrates them with the solver to improve solving efficiency.
Label-Agnostic Forgetting: A Supervision-Free Unlearning in Deep Models
Shaofei Shen (University of Queensland), Miao Xu (University of Queensland)
CodeSafty and PrivacyData-Centric LearningAuto EncoderContrastive LearningImage
🎯 What it does: A framework LAF is proposed to achieve unlearning of deep models without label supervision, which can maintain the knowledge of remaining data while deleting specified data and resisting privacy attacks.
Label-Focused Inductive Bias over Latent Object Features in Visual Classification
Ilmin Kang (Gwangju Institute of Science and Technology), Kangil Kim (Gwangju Institute of Science and Technology)
CodeClassificationTransformerImage
🎯 What it does: Proposes the Label-focused Latent-object Biasing (LLB) method, which quantizes the visual features from the intermediate layers of ViT into latent objects and disconnects visual dependencies to learn non-visual features based solely on labels, then fuses them with the original visual features to enhance the generalization ability of image classification models.
Label-free Node Classification on Graphs with Large Language Models (LLMs)
Zhikai Chen (Michigan State University), Jiliang Tang (Michigan State University)
CodeClassificationGraph Neural NetworkLarge Language ModelTextGraph
🎯 What it does: A complete process for unsupervised node classification is proposed, utilizing a large language model (LLM) to annotate a small number of nodes, and then using these annotations to train a graph neural network (GNN) to complete the node classification for the entire graph.
Byeonghu Na (Korea Advanced Institute of Science and Technology), Il-chul Moon
CodeGenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper proposes a target function for training conditional diffusion models with noisy labels—Transition-aware Weighted Denoising Score Matching (TDSM), which eliminates the conditional bias caused by noisy labels through a weighted score network.
Fabricio Arend Torres (University of Basel), Volker Roth (University of Basel)
CodeFlow-based ModelTime SeriesPhysics Related
🎯 What it does: A Lagrangian flow network (LFlows) based on conditional normalizing flows is proposed, which models density and velocity fields in continuous time-space through a differentiable invertible mapping, naturally satisfying the continuity equation.
🎯 What it does: An end-to-end LaneSegNet network is proposed, which for the first time uses lane segments to uniformly represent map features and achieve complete road structure perception in autonomous driving scenarios.