These 1874 NeurIPS 2024 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 NeurIPS 2024 paper, free trial on arXivSub.
(FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning
Seungjoo Lee (Korea Advanced Institute of Science and Technology), Sung-Ju Lee (Korea Advanced Institute of Science and Technology)
CodeFederated LearningSupervised Fine-TuningImage
π― What it does: In the scenario of federated semi-supervised learning where labels are only on the server, a new federated training framework (FLΒ²) is proposed, which significantly reduces confirmation bias and improves model performance through adaptive thresholds, sharpness consistency regularization, and learning state-aware aggregation.
$\beta$-DPO: Direct Preference Optimization with Dynamic $\beta$
Junkang Wu (University of Science and Technology of China), Xiangnan He (University of Science and Technology of China)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
π― What it does: Dynamic Ξ² calibration and data filtering for DPO are conducted, proposing the Ξ²-DPO framework to enhance the alignment of LLM with human feedback.
π― What it does: A simple Ξ΅-softmax layer is designed in deep learning to approximate one-hot vectors, thereby alleviating label noise and enhancing model robustness.
π― What it does: A conditional diffusion model named ID3 and its sampling algorithm have been designed and implemented to automatically generate diverse, identity-preserving synthetic face data for training facial recognition models.
$\textit{Bifr\"ost}$: 3D-Aware Image Compositing with Language Instructions
Lingxiao Li (Chinese University of Hong Kong), Xiangyu Yue (Chinese University of Hong Kong)
CodeGenerationData SynthesisDepth EstimationTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelImageVideoTextMultimodality
π― What it does: This paper proposes the BifrΓΆst framework, which achieves 3D perception image synthesis based on language instructions, precisely inserting reference objects into the background image while maintaining lighting, occlusion, and other three-dimensional spatial relationships.
Donato Crisostomi (Sapienza University of Rome), Emanuele RodolΓ (Sapienza University of Rome)
CodeOptimizationConvolutional Neural NetworkImage
π― What it does: A data-independent weight matching and merging method is proposed, achieving multi-model periodic consistency merging through global optimization of neuron arrangement.
π― What it does: This paper proposes a low-bit post-training quantization method for image super-resolution models based on the Transformer architectureβ2DQuant.
π― What it does: This paper proposes a parameter-efficient fine-tuning method based on two-dimensional rotation, called RoAd, and evaluates its performance on tasks such as GLUE, commonsense reasoning, and arithmetic reasoning, demonstrating its advantages in batch processing and composability.
π― What it does: This paper addresses the low rendering efficiency of 3D Gaussian splatting on edge devices by proposing a Fragment Pruning method based on adaptive truncation thresholds for each Gaussian;
π― What it does: Proposes 4-bit Shampoo, which compresses the state of second-order optimizers to 4 bits for memory efficiency while maintaining training performance close to the 32-bit version.
4M-21: An Any-to-Any Vision Model for Tens of Tasks and Modalities
Roman Bachmann (Swiss Federal Institute of Technology Lausanne), Amir Zamir (Apple)
CodeGenerationRetrievalTransformerVision Language ModelTextMultimodality
π― What it does: Trained a multimodal model 4M-21 that can take arbitrary inputs and produce arbitrary outputs, supporting interactions and generation across 21 different modalities.
π― What it does: This paper proposes a point selection method based on Bayesian inference called BADS, which utilizes SGLD to simultaneously learn model parameters and sample weights, thereby addressing the issues of slow convergence and high memory consumption in traditional two-layer optimization.
A Canonicalization Perspective on Invariant and Equivariant Learning
George Ma (Peking University), Yisen Wang (Peking University)
CodeGraph Neural NetworkGraph
π― What it does: Research on framework-based averaging methods, proposing a unified and optimized approach to symbol/basis invariance learning from the perspective of canonicalization.
A Cat Is A Cat (Not A Dog!): Unraveling Information Mix-ups in Text-to-Image Encoders through Causal Analysis and Embedding Optimization
Chieh-Yun Chen (National Yang Ming Chiao Tung University), Hong-Han Shuai (National Yang Ming Chiao Tung University)
CodeObject DetectionGenerationOptimizationTransformerVision Language ModelDiffusion modelImageText
π― What it does: This paper studies the impact of the causal attention mechanism in text encoders on text-to-image diffusion models, proposing a training-independent text embedding balance optimization method (TEBOpt) to eliminate information bias and loss, and presents new automatic evaluation metrics.
π― What it does: This study investigates the role of the CLS token in Vision Transformer in cross-domain few-shot learning and proposes enhancing model generalization performance by decoupling domain information.
CodeExplainability and InterpretabilityTransformerVision Language ModelImageTextMultimodality
π― What it does: This paper proposes an explanation framework CoX-LMM based on vocabulary concept dictionary learning, aimed at interpreting the internal representations of large multimodal models (LMM) and conducting multimodal concept mining.
π― What it does: This paper proposes a Consistency-Aware Spot-Guided Transformer (CAST) that achieves semi-dense, geometrically consistent coarse matching and designs a lightweight sparse-to-dense refinement module for efficient and accurate point cloud registration.
A Critical Evaluation of AI Feedback for Aligning Large Language Models
Archit Sharma (Stanford University), Thomas Kollar (Toyota Research Institute)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: This paper explores the necessity of the RL stage in aligning large language models by comparing two alignment methods: Supervised Fine-Tuning (SFT) and Reinforcement Learning with AI Feedback (LAIF).
A Fast Convoluted Story: Scaling Probabilistic Inference for Integer Arithmetics
Lennert De Smet (KU Leuven), Pedro Zuidberg Dos Martires (Γrebro University)
CodeComputational EfficiencyImageBenchmark
π― What it does: A differentiable probabilistic integer linear arithmetic framework PLIAt based on tensorization and Fast Fourier Transform (FFT) is proposed for efficient integer probabilistic inference and learning.
A Flexible, Equivariant Framework for Subgraph GNNs via Graph Products and Graph Coarsening
Guy Bar-Shalom (Technion - Israel Institute of Technology), Haggai Maron (NVIDIA Research)
CodeDrug DiscoveryGraph Neural NetworkGraphBiomedical Data
π― What it does: This paper proposes a Subgraph GNN framework with a variable subgraph set (CS-GNN), which first refines the original graph to obtain a set of supernodes, then constructs a product graph through the Cartesian product with the original graph, and performs message passing on this graph, supporting subgraph bags of arbitrary size; it also introduces symmetry-based equivariant linear layers and various node labeling strategies.
A Foundation Model for Zero-shot Logical Query Reasoning
Mikhail Galkin (Intel AI Lab), Zhaocheng Zhu (Mila - Quebec AI Institute)
CodeGraph Neural NetworkGraph
π― What it does: This paper proposes ULTRAQUERY, the first zero-shot knowledge graph complex logic query answering foundational model, capable of reasoning over any new knowledge graph.
A Globally Optimal Portfolio for m-Sparse Sharpe Ratio Maximization
Yizun Lin (Jinan University), Cheng Li (Jinan University)
CodeOptimizationTabularTime SeriesFinance Related
π― What it does: This paper proposes an m-sparse (at most m active assets) Sharpe ratio maximization model and provides its equivalent quadratic programming form. Subsequently, a Proximal Gradient Algorithm (PGA) based on semi-algebraic properties is developed to solve this non-convex problem, theoretically achieving a global optimal solution under certain conditions.
π― What it does: A gradient accumulation method called CONTACCUM with a dual memory pool is proposed for stabilizing the training of dense retrievers in low-resource environments.
A hierarchical decomposition for explaining ML performance discrepancies
Harvineet Singh (University of California, San Francisco), Jean Feng (University of California, San Francisco)
CodeDomain AdaptationExplainability and InterpretabilityBiomedical DataElectronic Health Records
π― What it does: This paper proposes a hierarchical, non-parametric framework (HDPD) to explain the fundamental reasons for performance differences of machine learning models across different domains.
A Label is Worth A Thousand Images in Dataset Distillation
Tian Qin (Harvard University), David Alvarez-Melis (Harvard University)
CodeClassificationKnowledge DistillationImage
π― What it does: Through extensive ablation experiments and a simple soft label baseline, it is demonstrated that soft labels are the core factor for the success of data distillation methods, and that randomly sampling real images paired with soft labels from pre-trained experts can approach or even surpass existing state-of-the-art synthetic image distillation methods under significant data compression.
A Local Method for Satisfying Interventional Fairness with Partially Known Causal Graphs
Haoxuan Li (Peking University), Kun Zhang (Carnegie Mellon University)
CodeOptimizationGraph Neural NetworkTabular
π― What it does: A local method is proposed to enumerate the possible parent sets of sensitive attributes, estimate propensity scores, and achieve causal intervention fairness through min-max joint optimization, under the condition of only observing data and having partial knowledge of the causal graph.
π― What it does: A motion-aware spatiotemporal graph model is proposed for video salient object ranking, and video redirection is achieved based on the ranking results.
π― What it does: This paper proposes a technique that dynamically adjusts the Softmax gradient decay rate using a PID controller, thereby optimizing both model accuracy and calibration performance during training.
π― What it does: A penalty reconstruction based on Lagrangian duality is proposed, and a full-step algorithm BLOCC is designed to solve the bilevel optimization problem with coupling constraints.
A Prompt-Based Knowledge Graph Foundation Model for Universal In-Context Reasoning
Yuanning Cui (Nanjing University), Wei Hu (Nanjing University)
CodeGraph Neural NetworkPrompt EngineeringGraph
π― What it does: A KG-based model KG-ICL is proposed, which implements context learning without parameter updates to complete cross-KG reasoning tasks.
A robust inlier identification algorithm for point cloud registration via $\mathbf{\ell_0}$-minimization
Yinuo Jiang (Huazhong University of Science and Technology), Ye Yuan (Huazhong University of Science and Technology)
CodeAutonomous DrivingOptimizationPoint Cloud
π― What it does: A robust inlier identification algorithm is proposed that transforms the point cloud registration problem into an β0-minimization problem of alignment error for each local set.
π― What it does: This paper proposes a scalable generative model that utilizes control theory's teacher forcing (GTF) and Wiener deconvolution techniques to reconstruct dynamic systems from convolutional observational data such as BOLD fMRI, and generates interpretable generative models.
π― What it does: This paper studies a general in-context example-based image segmentation framework called SINE, which can simultaneously output masks at three granularities: object, instance, and semantic, addressing the task ambiguity problem in traditional in-context segmentation.
π― What it does: A unified depth completion (UniDC) problem is proposed, constructing a lightweight framework based on deep foundational models. It utilizes the relative depth features from a monocular camera, hyperplane geometry, and multi-scale feature fusion to achieve rapid conversion from sparse depth to dense depth, and completes depth refinement through pixel-level affinity graphs, supporting few-shot learning with very few labeled data.
π― What it does: This paper proposes a single-step Sharpness-Aware Minimization (S2-SAM) with no additional computational cost, which is applied as a plugin to various sparse training methods, significantly enhancing the generalization performance and robustness of sparse networks.
A Surprisingly Simple Approach to Generalized Few-Shot Semantic Segmentation
Tomoya Sakai (IBM Research), Tadanobu Inoue (IBM)
CodeSegmentationSupervised Fine-TuningImage
π― What it does: A method for base class mining (BCM) based on simple rules is proposed, utilizing standard supervised learning to identify new categories in generalized few-shot semantic segmentation (GFSS) while maintaining the segmentation performance of most base classes.
π― What it does: A federated multi-task learning framework FedSAK based on tensor trace norm is proposed, which can simultaneously handle the heterogeneity of data, models, and tasks within the same framework.
A theoretical design of concept sets: improving the predictability of concept bottleneck models
Max Ruiz Luyten (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
CodeClassificationDomain AdaptationTransformerLarge Language ModelContrastive LearningImageTextMultimodality
π― What it does: In the paper, the authors propose and validate a theoretical framework and empirical effects of concept sets in the Concept Bottleneck Model (CBM), studying the expressiveness of concept sets and the model awareness bias, and providing corresponding theoretical derivations and experimental validations.
A Theoretical Understanding of Self-Correction through In-context Alignment
Yifei Wang (Massachusetts Institute of Technology), Yisen Wang (Peking University)
CodeTransformerLarge Language ModelText
π― What it does: This study investigates the self-correction mechanism of LLMs and provides a theoretical analysis from the perspective of contextual alignment.
A Topology-aware Graph Coarsening Framework for Continual Graph Learning
Xiaoxue Han (Stevens Institute of Technology), Yue Ning (Stevens Institute of Technology)
CodeOptimizationRepresentation LearningGraph Neural NetworkGraphTime Series
π― What it does: A topology-aware graph coarsening framework, TA CO, is proposed to alleviate the problem of catastrophic forgetting in continual graph learning (CGL) by storing compressed graphs of previous tasks.
Xuejie Liu (Peking University), Yitao Liang (Peking University)
CodeTransformerReinforcement LearningSequential
π― What it does: The Trifle algorithm is proposed, utilizing a tractable probabilistic model (TPM) to make the inference process of offline RL tractable, significantly improving the performance of offline RL in action sampling and reward estimation.
A Unified Confidence Sequence for Generalized Linear Models, with Applications to Bandits
Junghyun Lee, Kwang-Sung Jun (University of Arizona)
CodeOptimizationReinforcement LearningTabular
π― What it does: A unified likelihood ratio confidence sequence (CS) framework is proposed, applicable to all self-conjugate (GLM) models, and based on this, a universal UCB algorithm OFUGLB is designed.
A Unifying Normative Framework of Decision Confidence
Amelia Johnson, Koosha Khalvati (Allen Institute)
CodeOptimizationReinforcement LearningTabular
π― What it does: A unified normative framework is proposed to measure decision confidence using probabilistic models and map it to planning as inference (maximum entropy reinforcement learning).
A Unifying Post-Processing Framework for Multi-Objective Learn-to-Defer Problems
Mohammad-Amin Charusaie (Max Planck Institute for Intelligent Systems), Samira Samadi (Max Planck Institute for Intelligent Systems)
CodeAnomaly DetectionOptimizationTabular
π― What it does: A post-processing framework based on the d-dimensional generalized Neyman-Pearson rule is proposed to simultaneously satisfy accuracy and various constraints (fairness, expert intervention budget, anomaly detection, etc.) in the multi-objective learn-to-defer (L2D) problem.
A2PO: Towards Effective Offline Reinforcement Learning from an Advantage-aware Perspective
Yunpeng Qing (Zhejiang University), Mingli Song (Zhejiang University)
CodeReinforcement LearningAuto EncoderTabular
π― What it does: A new offline reinforcement learning method A2PO is proposed, which addresses constraint conflicts in mixed-quality datasets through advantage-aware policy optimization.
π― What it does: Define and solve the Maximum A Posteriori (MAP) and Marginal MAP (MMAP) inference tasks in Logic Credible Networks (LCNs), proposing two types of solution approaches: exact search and approximate message passing.
Abrupt Learning in Transformers: A Case Study on Matrix Completion
Pulkit Gopalani (University of Michigan), Wei Hu (University of Michigan)
CodeTransformerLarge Language ModelTabular
π― What it does: This study explores the training dynamics of the Transformer model in the low-rank matrix completion task, finding that the training loss experiences a plateau in the early stages of training, followed by a sudden drop to near-optimal values.
Absorb & Escape: Overcoming Single Model Limitations in Generating Heterogeneous Genomic Sequences
Zehui Li (Imperial College London), Yiren Zhao (Imperial College London)
CodeGenerationData SynthesisOptimizationDiffusion modelBiomedical Data
π― What it does: This paper proposes an Absorb & Escape (A&E) framework and a fast implementation called Fast A&E, which utilizes a combination of pre-trained autoregressive models and diffusion models to generate high-quality genomic sequences.
Accelerating Greedy Coordinate Gradient and General Prompt Optimization via Probe Sampling
Yiran Zhao (National University of Singapore), Michael Shieh (National University of Singapore)
CodeOptimizationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes the Probe Sampling algorithm, which uses a small draft model to filter candidate prompts in GCG search, significantly accelerating the adversarial prompt optimization for LLMs.
π― What it does: A systematic analysis of the internal structure of Non-Maximum Suppression (NMS) from the perspective of graph theory is presented, proposing two efficient implementations (QSI-NMS, eQSI-NMS, and BOE-NMS), and constructing the NMS-Bench benchmark for quick evaluation and comparison of different NMS algorithms.
Achieving Domain-Independent Certified Robustness via Knowledge Continuity
Alan Sun (Carnegie Mellon University), Soroush Vosoughi (Dartmouth College)
CodeClassificationAdversarial AttackText
π― What it does: A new definition of robustness called Knowledge Continuity is proposed, along with its theoretical proof and practical applications.
ActAnywhere: Subject-Aware Video Background Generation
Boxiao Pan (Stanford University), Jimei Yang (Runway)
CodeGenerationData SynthesisDiffusion modelVideo
π― What it does: This paper proposes a method for automatically generating video backgrounds that match the motion of foreground subjects, using a single frame background image to create a complete video with realistic interactions with the subject.
π― What it does: This paper proposes a unified diffusion model (ActFusion) that simultaneously performs temporal action segmentation and long-term action prediction in videos, utilizing learnable masking tokens to achieve segmentation of visible parts and prediction of invisible parts.
π― What it does: This study investigates continuous-time distributed reinforcement learning, where action values collapse as decision frequency increases. It proposes the concept of distributed superiority and designs an algorithm called DSUP based on this concept to address performance instability under high-frequency decision-making.
π― What it does: When training deep networks on edge devices, this paper compresses storage by using low-rank tensor decomposition (SVD and HOSVD) on activation maps, significantly reducing the memory requirements for backpropagation.
π― What it does: An active learning method for derivative-based global sensitivity analysis (DGSM) is proposed, which directly targets the quantification metrics of gradient, absolute gradient, and squared gradient for sample-efficient collection.
Active Learning with LLMs for Partially Observed and Cost-Aware Scenarios
NicolΓ‘s Astorga (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
CodeOptimizationData-Centric LearningTransformerLarge Language ModelTabular
π― What it does: A proactive learning framework named POCA is proposed, specifically designed to address the issues of partial observation and cost-constrained data collection.
Active preference learning for ordering items in- and out-of-sample
Herman BergstrΓΆm (Chalmers University of Technology and University of Gothenburg), Fredrik D. Johansson (Chalmers University of Technology and University of Gothenburg)
CodeRecommendation SystemOptimizationImageTabular
π― What it does: Learn complete rankings on item sets with contextual attributes through Active Preference Learning, and provide an upper bound on ranking error along with an implemented active sampling strategy.
Active Sequential Posterior Estimation for Sample-Efficient Simulation-Based Inference
Sam Griesemer (University of Southern California), Yan Liu (University of Southern California)
CodeFlow-based ModelTabular
π― What it does: The Active Sequential Neural Posterior Estimation (ASNPE) method is proposed for efficient inference of posterior distributions in expensive simulation models.
π― What it does: This paper presents ActSort, an active learning-based cell sorting algorithm that can quickly perform quality control of cell candidates in large calcium imaging datasets.
AdaNeg: Adaptive Negative Proxy Guided OOD Detection with Vision-Language Models
Yabin Zhang (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
CodeAnomaly DetectionTransformerVision Language ModelContrastive LearningImageMultimodality
π― What it does: The AdaNeg method is proposed, which dynamically generates adaptive negative proxies through a feature memory bank during testing to better align with the OOD distribution.
AdaNovo: Towards Robust \emph{De Novo} Peptide Sequencing in Proteomics against Data Biases
Jun Xia (Westlake University), Stan Z. Li (Westlake University)
CodeRecognitionData-Centric LearningTransformerBiomedical Data
π― What it does: This paper proposes the AdaNovo framework, which utilizes Conditional Mutual Information (CMI) to re-weight training loss in order to improve the performance of de-biasing in de novo peptide sequencing, particularly for the identification of PTMs.
π― What it does: For the radar semantic segmentation task, Adaptive Peak-aware Convolution (AdaPKC) is proposed with two implementations (metric-based AdaPKC ΞΎ and learning-based AdaPKC ΞΈ). Additionally, a threshold online switching fine-tuning strategy (FiTOS) is introduced to further enhance performance.
Adaptable Logical Control for Large Language Models
Honghua Zhang (University of California Los Angeles), Nanyun Peng (University of California Los Angeles)
CodeGenerationKnowledge DistillationTransformerLarge Language ModelText
π― What it does: The Ctrl-G framework is proposed, using hidden Markov models (HMM) as an interpretable approximation of large language models, and employing deterministic finite automata (DFA) for logical constraint control, thereby achieving inferable and satisfiable generation.
π― What it does: A network architecture with predictable and adjustable depth is designed, dividing the residual blocks into necessary and skipable sub-paths, and using self-distillation to refine features in the latter, allowing multiple depth sub-networks to be obtained from a single training.
π― What it does: This paper proposes and implements GVFExplorer, which can efficiently evaluate multiple Generalized Value Functions (GVFs) in parallel by adaptively learning a single behavior policy, and achieves minimum variance updates on offline data using TD variance estimation.
Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare
Hanwei Zhu (City University of Hong Kong), Shiqi Wang (City University of Hong Kong)
CodeOptimizationTransformerLarge Language ModelVision Language ModelImageMultimodality
π― What it does: A large model-based NR-IQA model called Compare2Score is proposed, which learns image quality comparisons and converts them into continuous scores.
CodeDomain AdaptationReinforcement LearningBiomedical Data
π― What it does: An adaptive label sampling framework is proposed for efficiently evaluating the performance of machine learning models under severe distribution shifts.
Adaptive Layer Sparsity for Large Language Models via Activation Correlation Assessment
Wei Li (University of Birmingham), Shengjie Sun (AISpeech Co., Ltd.)
CodeCompressionOptimizationTransformerLarge Language ModelText
π― What it does: Proposes the Adaptive Layer Sparsity (ALS) method, which compresses large language models by adaptively allocating layer-wise sparsity rates while maintaining or improving inference performance.
Adaptive Proximal Gradient Method for Convex Optimization
Yura Malitsky (University of Vienna), Konstantin Mishchenko (Samsung AI Center)
CodeOptimization
π― What it does: An adaptive gradient descent and adaptive proximal gradient descent algorithm (AdGD/AdProxGD) based on local curvature information is proposed, which does not require a global Lipschitz constant and can automatically adjust the step size, theoretically converging to the optimal solution of convex problems.
π― What it does: This paper proposes Ada-STORM, an adaptive variance reduction algorithm that achieves optimal convergence rates in scenarios such as non-convex optimization, component optimization, and finite summation.
Addressing Asynchronicity in Clinical Multimodal Fusion via Individualized Chest X-ray Generation
Wenfang Yao (Hong Kong Polytechnic University), Jing Qin (Hong Kong Baptist University)
CodeGenerationData SynthesisAnomaly DetectionTransformerDiffusion modelAuto EncoderContrastive LearningMultimodalityBiomedical DataElectronic Health Records
π― What it does: This paper proposes a method called DDL-CXR based on the Latent Diffusion Model (LDM), which dynamically generates personalized chest X-ray (CXR) latent representations synchronized with the prediction time to address the asynchrony issue of multimodal clinical data (EHR and CXR) and integrates the generated latent CXR with historical data for clinical predictions.
CodeAnomaly DetectionRepresentation LearningTransformerTime Series
π― What it does: This paper proposes the MiTSformer framework for unified modeling of mixed temporal data (including continuous and discrete variables) to achieve general representation learning for multiple tasks.
π― What it does: This paper proposes a new adaptive gradient optimizer called ADOPT, which theoretically achieves a convergence rate of O(1/βT) without requiring a specific choice of Ξ²2 and without relying on the bounded gradient noise assumption. It outperforms Adam and its variants across various tasks.
π― What it does: This paper proposes a non-parametric diffusion process-based adversarial attack framework, AdvAD, and its extreme version, AdvAD-X, to generate imperceptible adversarial samples.
π― What it does: An adaptive domain scheduling method for Open Set Domain Generalization (OSDG) is proposedβEvidential Bi-Level Hardest Domain Scheduler (EBiL-HaDS), which enhances the model's generalization and recognition capabilities for unknown domains and unknown categories by dynamically selecting the hardest domains.
Advancing Spiking Neural Networks for Sequential Modeling with Central Pattern Generators
Changze Lv (Fudan University), Dongsheng Li (Microsoft Research Asia)
CodeClassificationRecognitionOptimizationSpiking Neural NetworkTransformerImageTextTime Series
π― What it does: A location information encoding method based on Central Pattern Generators (CPG) is proposed in spiking neural networks (SNN), and its effectiveness is validated in tasks such as time series prediction, text classification, and image classification.
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: By constructing a stepwise preference dataset based on failure exploration and applying Direct Preference Optimization (DPO) after Supervised Fine-Tuning (SFT), the multi-step reasoning ability of tool-enhanced large language models has been improved.
Advancing Training Efficiency of Deep Spiking Neural Networks through Rate-based Backpropagation
Chengting Yu (Zhejiang University), Aili Wang (Zhejiang University)
CodeSpiking Neural NetworkImage
π― What it does: A rate coding-based backpropagation method is proposed to simplify the time dependency of SNNs and reduce the computational and memory costs of BPTT.
Niloufar Zakariaei (University of British Columbia), Moshe Eliasof (University of Cambridge)
CodeRestorationGenerationOptimizationConvolutional Neural NetworkVideoTime SeriesPhysics Related
π― What it does: This paper proposes an architecture that integrates the advection-diffusion-reaction (ADR) process into convolutional neural networks for efficient prediction of spatio-temporal sequences.
Adversarial Moment-Matching Distillation of Large Language Models
Chen Jia (SI-TECH Information Technology)
CodeKnowledge DistillationTransformerLarge Language ModelReinforcement LearningText
π― What it does: This paper proposes a knowledge distillation method based on adversarial action value moment matching, utilizing a reinforcement learning framework to view distillation as imitation learning. The goal is to minimize the difference in action value moments between the teacher and the student, combining both on-policy and off-policy perspectives.
π― What it does: A discrete-time iterative Markov fitting (D-IMF) method is proposed to efficiently solve the SchrΓΆdinger bridge problem, achieving unsupervised domain transformation from distribution p0 to p1.
Shuofei Qiao (Zhejiang University), Huajun Chen (National University of Singapore)
CodeTransformerLarge Language ModelAgentic AIContrastive LearningText
π― What it does: A parameterizable World Knowledge Model (WKM) is proposed, which synthesizes task knowledge and state knowledge from expert trajectories and experiential exploration trajectories to assist large language models in interactive planning.
AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases
Zhaorun Chen (University of Chicago), Bo Li (University of Chicago)
CodeAutonomous DrivingOptimizationAdversarial AttackTransformerLarge Language ModelAgentic AITextSequentialBiomedical DataElectronic Health RecordsRetrieval-Augmented Generation
π― What it does: This paper proposes a backdoor attack method called AGENTPOISON for large language model (LLM) agents using retrieval-augmented generation (RAG). By injecting a small number of malicious examples into the agent's long-term memory or knowledge base, the attack can be triggered when specific keywords are retrieved, inducing the agent to perform malicious actions.
Aggregating Quantitative Relative Judgments: From Social Choice to Ranking Prediction
Yixuan Even Xu (Carnegie Mellon University), Vincent Conitzer (Carnegie Mellon University)
CodeTabular
π― What it does: Proposes and studies the Quantitative Relative Judgment Aggregation (QRJA) model, which serves as a bridge between social choice and ranking prediction;
AGILE: A Novel Reinforcement Learning Framework of LLM Agents
Peiyuan Feng (ByteDance Research), Hang Li (ByteDance Research)
CodeTransformerLarge Language ModelReinforcement LearningAgentic AIText
π― What it does: A unified reinforcement learning framework called AGILE is proposed, allowing large language model agents to complete complex dialogue tasks through memory, tool invocation, reflection, and actively seeking human advice.
AHA: Human-Assisted Out-of-Distribution Generalization and Detection
Haoyue Bai (University of Wisconsin Madison), Robert D Nowak
CodeDomain AdaptationAnomaly DetectionImage
π― What it does: Proposes the AHA framework, which enhances the model's OOD generalization and detection capabilities using a small amount of manual labeling in the maximum discernment region (where the densities of covariate OOD and semantic OOD are roughly equal).
π― What it does: This paper presents AirSketch, which utilizes a controllable diffusion model to recover clean, coherent, and user-intent-compliant hand-drawn sketches from extremely noisy trajectory images generated by hand tracking, without the need for markers or expensive hardware.
AlchemistCoder: Harmonizing and Eliciting Code Capability by Hindsight Tuning on Multi-source Data
Zifan Song (Tongji University), Cairong Zhao (Tongji University)
CodeGenerationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: Developed the AlchemistCoder series of code LLMs, significantly enhancing code generation and reasoning capabilities through multi-source data, AlchemistPrompt, and fine-tuning on code understanding tasks.
Ziqian Zhong (Massachusetts Institute of Technology), Jacob Andreas (Massachusetts Institute of Technology)
CodeGenerationRetrievalTransformerText
π― What it does: This paper studies whether a randomly initialized Transformer that only trains the embedding layer (i.e., freezing the internal parameters) can perform algorithmic tasks such as arithmetic, associative retrieval, and bracket matching, and explores its potential in memory and natural language generation.
π― What it does: This paper studies the feasibility of enhancing the visibility of minority artists in a Transformer-based music recommendation system through algorithmic collective action. It proposes two lightweight, authenticity-constrained song insertion strategies (InClust, DirLoF) and experimentally validates their effectiveness using large industry-level models.
ALI-Agent: Assessing LLMs' Alignment with Human Values via Agent-based Evaluation
Jingnan Zheng (National University of Singapore), Tat-Seng Chua (Singapore Management University)
CodeTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextChain-of-Thought
π― What it does: The ALI-Agent framework is designed to automatically generate, evaluate, and iteratively improve alignment tests for human values through LLM agents.
π― What it does: This paper proposes a neural operator MambaNO that integrates the Mamba state space model with convolutional integration for efficiently approximating the analytical operator of PDEs while maintaining alias-free properties, with a time complexity of O(N).
π― What it does: This paper proposes a two-stage method called Efficient Diffusion Alignment (EDA), which breaks down offline reinforcement learning tasks into behavior pre-training and policy alignment. By pre-training a diffusion behavior model and fine-tuning it with a Q-function, efficient continuous control is achieved.
Aligning Large Language Models with Representation Editing: A Control Perspective
Lingkai Kong (Georgia Tech), Chao Zhang (Georgia Tech)
CodeOptimizationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: This paper proposes a framework for aligning large language models through dynamic representation editing during the inference phase;
Aligning LLM Agents by Learning Latent Preference from User Edits
Ge Gao (Cornell University), Dipendra Misra (Microsoft Research)
CodeRecommendation SystemTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: This paper proposes an interactive learning framework called PRELUDE, based on user edits, to learn and infer users' implicit preference descriptions, thereby guiding LLM agents to generate text that better meets user needs.