π― What it does: This paper proposes MetaPhysiCa, a physical information machine learning framework that achieves robustness in predicting out-of-distribution (OOD) dynamical systems through meta-learning and causal structure discovery.
MetaTool Benchmark for Large Language Models: Deciding Whether to Use Tools and Which to Use
Yue Huang (Lehigh University), Lichao Sun (Lehigh University)
CodeTransformerLarge Language ModelAgentic AITextBenchmark
π― What it does: Proposes the METATOOL benchmark to evaluate the capabilities of large language models in tool awareness and tool selection; simultaneously constructs the TOOLE dataset, which includes 21,127 diverse user queries (single tool and multiple tools), and designs four sub-tasks to examine different dimensions of tool selection.
MG-TSD: Multi-Granularity Time Series Diffusion Models with Guided Learning Process
Xinyao Fan (University of British Columbia), Jiang Bian (Microsoft Research)
CodeRecurrent Neural NetworkDiffusion modelTime Series
π― What it does: This paper proposes a Multi-Granularity Time Series Diffusion Model (MG-TSD), which guides the diffusion model learning by using time series of different granularities as targets in the intermediate steps of the diffusion process, thereby enhancing the stability and accuracy of probabilistic time series forecasting.
π― What it does: This study investigates the object coupling problem in dense self-supervised learning and proposes a solution using Region Collaborative Cutout (RCC) and a decoupling branch.
CodeKnowledge DistillationTransformerLarge Language ModelReinforcement LearningText
π― What it does: This paper proposes a knowledge distillation method based on reverse KL divergence, called MINILLM, to effectively transfer the generative distribution of large language models to smaller models.
CodeClassificationRecognitionTransformerLarge Language ModelVideoTextMultimodalityBenchmarkAudio
π― What it does: A large-scale multimodal dialogue intent recognition and out-of-domain detection dataset MIntRec2.0 has been constructed, along with a unified evaluation framework.
CodeOptimizationAdversarial AttackGraph Neural NetworkMixture of ExpertsAuto EncoderGraphBenchmark
π― What it does: To address the issue of 'destructive degradation' in graph neural networks under high-intensity node injection attacks, the DRAGON framework is proposed, which enhances the model's robustness and scalability against attacks through a preprocessed denoising autoencoder and a hierarchical differential privacy mixture of experts mechanism.
Mitigating Hallucination in Large Multi-Modal Models via Robust Instruction Tuning
Fuxiao Liu (University of Maryland), Lijuan Wang (Microsoft Corporation)
CodeObject DetectionAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
π― What it does: A large-scale multimodal instruction tuning dataset (LRV-Instruction) and evaluation method (GAVIE) are proposed to reduce the hallucination phenomenon of large-scale multimodal models (LMM) under image + text instructions. Fine-tuning MiniGPT4 and mPLUG-Owl on this dataset significantly improves the model's authenticity and instruction-following ability.
π― What it does: In response to the dimensionality curse faced by Randomized Smoothing for high-dimensional inputs when providing β2 certified robustness, this paper proposes the Dual Randomized Smoothing (DRS) mechanism, which first divides the input image into two low-dimensional sub-images, then smooths them separately and aggregates the results to obtain the certified robustness of the original input.
π― What it does: A hybrid table data synthesis framework named TABSYN is proposed, which maps table data to the VAE latent space and trains a score-based diffusion model in the latent space to achieve high-quality synthesis.
π― What it does: This paper proposes MixSATGEN, a method for generating instances with adjustable hardness through graph matching and substructure replacement on the literal-clause graph of SAT.
CodeTransformerLarge Language ModelMixture of ExpertsImageTextMultimodality
π― What it does: This paper studies and proposes a new method for combining multiple LoRA (Low-Rank Adaptation) modelsβMOLE (Mixture of LoRA Experts). It achieves hierarchical weighted combinations by using learnable gating functions at each layer, allowing for a more efficient and dynamic integration of multiple pre-trained LoRAs.
Hanqing Zeng (Meta AI), Jiebo Luo (University of Rochester)
CodeClassificationOptimizationExplainability and InterpretabilityGraph Neural NetworkMixture of ExpertsGraph
π― What it does: A mixed weak-strong expert (Mowst) model is proposed, combining a lightweight MLP with a powerful GNN to adaptively separate node features and neighborhood structures, and achieve expert collaboration through confidence gating.
π― What it does: A graph kernel based on Maximum Mean Discrepancy (MMD) (MMD-GK) and its deep learnable version (Deep MMD-GK) are proposed. After obtaining node representations through message passing, MMD comparison is directly performed on graph distributions to obtain inter-graph similarity.
CodeRecognitionGenerationData-Centric LearningTransformerLarge Language ModelVision Language ModelImageTextMultimodality
π― What it does: A novel visual-language model MMICL is proposed, which can efficiently handle interleaved multimodal prompts containing multiple images, achieving multimodal context learning (ICL).
Model Tells You What to Discard: Adaptive KV Cache Compression for LLMs
Suyu Ge (University of Illinois Urbana-Champaign), Jianfeng Gao (Microsoft)
CodeGenerationCompressionComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Designed and implemented an adaptive KV cache compression method called FastGen, which significantly reduces GPU memory usage during LLM inference while maintaining generation quality.
π― What it does: A complex vector drawing generation model based on 'stroke cloud' has been proposed, which can view drawings as a collection of an arbitrary number of Bezier curves. It generates new high-complexity vector drawings through a Set Transformer for encoding, MLP-based conditional diffusion decoding, and a latent diffusion generator.
π― What it does: This paper proposes the Spectral Transformation framework, exploring spectral transformations beyond whitening to avoid dimensional collapse in self-supervised learning, and introduces the IterNorm with trace loss (INTL) method.
Modulated Phase Diffusor: Content-Oriented Feature Synthesis for Detecting Unknown Objects
Aming WU, Cheng Deng (Xidian University)
CodeObject DetectionDiffusion modelImage
π― What it does: This paper proposes a phase-based diffusion generative model called Modulated Phase Diffusion (MPD), which enhances unsupervised out-of-distribution (OOD) object detection performance by performing Gaussian averaging on the phase of ID features during forward diffusion and designing two reverse processes to synthesize OOD features and augmented features.
π― What it does: A framework for generating and optimizing MOFs based on a coarse-grained diffusion model, named MOFDiff, is proposed, which can generate and optimize MOF structures without relying on predefined topologies.
π― What it does: In this paper, the authors constructed a large-scale image-entity dataset (I2E) and trained a new visual foundation model MOFI using three pre-training methods: supervised, contrastive, and multi-task. The model demonstrated strong performance on multiple image retrieval and classification benchmarks.
π― What it does: A fully convolutional network called MogaNet is proposed, based on multi-order game theory interactions, integrating multi-order gated aggregation and channel redistribution modules to achieve more efficient visual representation.
CodeDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBiomedical Data
π― What it does: Designed and released the Mol-Instructions dataset, covering three main categories of instructions: molecules, proteins, and biological texts, with approximately 2,043,587 entries, and subsequently used this data for instruction tuning of LLMs;
π― What it does: A sampling method for Bayesian linear inverse problems, MCGdiff, is proposed, utilizing a diffusion generative model prior to reconstruct unknown quantities.
Most discriminative stimuli for functional cell type clustering
Max F Burg, Alexander S Ecker
CodeOptimizationBiomedical Data
π― What it does: A method combining deep prediction models with EM-style clustering is proposed, utilizing the most discriminative stimuli (MDS) to simultaneously optimize stimuli and cell type clustering.
Motif: Intrinsic Motivation from Artificial Intelligence Feedback
Martin Klissarov (Mila), Mikael Henaff (FAIR at Meta)
CodeTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: Proposes the Motif method, which utilizes large language models (LLM) to evaluate preferences for environmental event descriptions (captions), generating intrinsic rewards for training agents in reinforcement learning;
MOTOR: A Time-to-Event Foundation Model For Structured Medical Records
Ethan Steinberg (Stanford University), Nigam Shah
CodeTransformerTime SeriesSequentialBiomedical DataElectronic Health Records
π― What it does: This study proposes and pre-trains a time-to-event (TTE) foundational model named MOTOR, utilizing self-supervised TTE objective learning on large-scale clinical event sequences and transferring it to various medical prediction tasks.
MT-Ranker: Reference-free machine translation evaluation by inter-system ranking
Ibraheem Muhammad Moosa (Pennsylvania State University), Wenpeng Yin (Pennsylvania State University)
CodeTransformerSupervised Fine-TuningText
π― What it does: A reference-free MT evaluation method called MT-Ranker is proposed, which directly predicts which of the two translations is better by comparing translations of the same source sentence, rather than providing an absolute quality score.
Multi-Scale Representations by Varying Window Attention for Semantic Segmentation
Haotian Yan (Beijing University of Posts and Telecommunications), Chuang Zhang (Beijing University of Posts and Telecommunications)
CodeSegmentationTransformerImage
π― What it does: A variable window attention (VWA) and a multi-scale decoder (VWFormer) based on it are proposed for efficiently learning multi-scale feature representations in semantic segmentation tasks.
Multi-Task Reinforcement Learning with Mixture of Orthogonal Experts
Ahmed Hendawy (TU Darmstadt), Carlo D'Eramo
CodeReinforcement LearningMixture of Experts
π― What it does: Designed and implemented the MOORE (Mixture of Orthogonal Experts) framework, which enforces orthogonality among expert representations on the Stiefel manifold using the Gram-Schmidt process, thereby providing diverse and interpolable shared representations for multi-task reinforcement learning, ultimately learning a unified policy.
π― What it does: This study explores a partially observable representation learning framework for multi-view nonlinear mixtures, providing a unified identifiability theory.
π― What it does: In a semi-supervised scenario with only single-modal labeled data and unlabeled multi-modal paired data, a method for estimating multi-modal interactions (redundant, unique, and collaborative) is proposed, utilizing these interaction quantities to predict multi-modal model performance, guide data collection, and model selection.
Multimodal Patient Representation Learning with Missing Modalities and Labels
Zhenbang Wu (University of Illinois Urbana-Champaign), Faraz Faghri (National Institutes of Health)
CodeRepresentation LearningGraph Neural NetworkContrastive LearningMultimodalityBiomedical DataAlzheimer's DiseaseElectronic Health Records
π― What it does: The MUSE method is proposed, which utilizes a patient-modal bipartite graph combined with mutually consistent contrastive learning to handle missing modalities and labels in multimodal clinical representation learning.
Yang He (Agency for Science Technology and Research), Ivor Tsang
CodeData SynthesisCompressionImage
π― What it does: A multi-size dataset compression method (MDC) is proposed, which can obtain a multi-size synthetic dataset with a single compression.
π― What it does: A completely zero-shot industrial anomaly classification and segmentation method called MuSc is proposed, which detects anomalies by mutual scoring between unlabeled test images.
MuSR: Testing the Limits of Chain-of-thought with Multistep Soft Reasoning
Zayne Rea Sprague (University of Texas), Greg Durrett (University of Texas)
CodeTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
π― What it does: The MuSR dataset is proposed, constructing multi-step soft reasoning tasks through natural language narratives, covering three domains: murder reasoning, object placement, and team allocation.
MUSTARD: Mastering Uniform Synthesis of Theorem and Proof Data
Yinya Huang (City University of Hong Kong), Xiaodan Liang (Shenzhen Campus of Sun Yat-sen University)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: By interacting with large language models and the Lean prover, automatically generate step-by-step mathematical problems, natural language answers, and formal proofs, while filtering data verified by the prover.
MVSFormer++: Revealing the Devil in Transformer's Details for Multi-View Stereo
Chenjie Cao (Fudan University), Yanwei Fu (Fudan University)
CodeDepth EstimationTransformerImage
π― What it does: The multi-view stereo reconstruction method MVSFormer++ based on Transformer introduces cross-view attention and position encoding optimization in two major modules: feature extraction and cost volume regularization.
NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers
Kai Shen (Zhejiang University), Jiang Bian (Zhejiang University)
CodeGenerationData SynthesisDiffusion modelAudio
π― What it does: Developed NaturalSpeech 2, which combines continuous vector audio codecs and latent diffusion models to achieve natural, zero-shot TTS and singing synthesis;
Navigating Dataset Documentations in AI: A Large-Scale Analysis of Dataset Cards on HuggingFace
Xinyu Yang (Cornell University), James Zou (Stanford University)
CodeText
π― What it does: This study examined the structure, completeness, and content of 7,433 dataset documents on Hugging Face, combining it with download statistics, usage details, and human evaluations;
Neel Jain (University of Maryland), Tom Goldstein (University of Maryland)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes an enhancement method (NEFTune) that injects random noise into embedding vectors during instruction fine-tuning, and verifies that it can significantly improve the performance of large language models in dialogue quality.
Negative Label Guided OOD Detection with Pretrained Vision-Language Models
Xue Jiang (Southern University of Science and Technology), Bo Han (Hong Kong Baptist University)
CodeDomain AdaptationAnomaly DetectionTransformerVision Language ModelContrastive LearningImageMultimodality
π― What it does: A zero-shot OOD detection framework called NegLabel is proposed, which is based on a pre-trained vision-language model and utilizes negative labels mined from a large corpus to distinguish between ID and OOD samples.
Nemesis: Normalizing the Soft-prompt Vectors of Vision-Language Models
Shuai Fu (Southern University of Science and Technology), Yu Zhang (Southern University of Science and Technology)
CodeClassificationDomain AdaptationTransformerPrompt EngineeringVision Language ModelImageMultimodality
π― What it does: This paper conducts systematic experiments on the norm of learnable soft prompt vectors in visual-language models, discovering the 'Low-Norm Effect.' Based on this, a soft prompt normalization method named Nemesis is proposed, which can dynamically adjust the norm during the soft prompt tuning process to enhance model performance.
π― What it does: The NETINFOF framework is proposed to measure and utilize the available information in graph structures and node features (NUI) without training a model, thereby achieving efficient inference in link prediction and node classification tasks.
Justin Dumouchelle (University of Toronto), Elias Boutros Khalil
CodeOptimizationTabular
π― What it does: This paper proposes Neur2RO, a two-stage robust optimization solving framework that embeds neural network predictions of the second-stage value function into Column-Constraint Generation (CCG).
π― What it does: Proposes the Neural Atoms method, which maps original atoms to a few learnable neural atoms, performs information exchange, and then projects back to atoms to enhance the GNN's ability to capture long-range interactions.
π― What it does: Developed the QuKerNet framework, which automates the design of quantum feature maps tailored for specific tasks to enhance the performance of quantum kernels on NISQ devices.
π― What it does: The Neural Common Neighbor (NCN) model is proposed, and based on this, a soft completion technique (NCNC) is introduced to address the issue of graph incompleteness, thereby achieving more accurate link prediction.
π― What it does: This paper proposes a framework that transforms neural fields from regression to classification - the Neural Field Classifier (NFC), which enhances the rendering and reconstruction performance of neural fields through target encoding and classification loss.
Neural Spectral Methods: Self-supervised learning in the spectral domain
Yiheng Du (University of California), Aditi S. Krishnapriyan (University of California)
CodeTabularPhysics Related
π― What it does: In the spectral domain, a neural operator is used to learn the mapping of parameterized PDE solutions, achieving self-supervised training solely by minimizing the residual, without the need for internal solution data.
Neural structure learning with stochastic differential equations
Benjie Wang (University of California Los Angeles), Wenbo Gong (Microsoft Research)
CodeTime SeriesStochastic Differential Equation
π― What it does: A new structure learning method called SCOTCH is proposed, which combines neural stochastic differential equations (SDE) and variational inference to infer the posterior distribution of possible structures.
Neuroformer: Multimodal and Multitask Generative Pretraining for Brain Data
Antonis Antoniades (University of California), Spencer Smith
CodeGenerationData SynthesisTransformerContrastive LearningMultimodalityBiomedical Data
π― What it does: Construct and train a multimodal multi-task generative pre-trained Transformer (Neuroformer) to autoregressively generate neural activity and predict behavior from cellular-level neural discharge, visual stimuli, and behavioral data.
π― What it does: Proposes the neuron activation state and neuron activation coverage (NAC), and utilizes NAC for OOD detection (NAC-UE) and OOD generalization evaluation (NAC-ME).
π― What it does: The NeurRev framework is proposed to revive neurons by identifying and pruning large negative weights that lead to neuron 'exhaustion', thereby enhancing the learning effectiveness of dynamic sparse training.
CodeTransformerSupervised Fine-TuningMultimodalityTime SeriesSequentialBiomedical Data
π― What it does: This paper evaluates the true capabilities of Transformers and State-Space Models (SSMs) in modeling long-range dependencies by comparing the effects of training with random initialization on long sequence tasks versus self-supervised pretraining (Self-Pretraining, SPT) on task data.
π― What it does: A deep network named NfgTransformer is proposed to learn the representation of regularized, interpretable, and equivariant regularized games (NFG), which can be used for various game theory tasks such as Nash equilibrium solving, maximum deviation profit estimation, and ranking.
CodeOptimizationRepresentation LearningGraph Neural NetworkGraphPhysics Related
π― What it does: This paper proposes a method for constructing high-dimensional network embeddings in quantum Hilbert space, utilizing the 'product state' generated by tensor products to achieve exponential embedding dimensions, and achieving O(p) training complexity through an inner product objective; it also introduces a parameter-sharing compressed variant called node2ket+; develops a Riemannian-Adagrad optimizer with normalization and positive inner product constraints; and implements the LIBN2K C++ library.
π― What it does: A non-optimized Noise Map Guidance (NMG) inverse method is proposed, which directly retains spatial context using noise maps, thereby enhancing the editing quality of real images.
π― What it does: A NoiseDiffusion method is proposed, which performs noise correction on natural images in the diffusion model before interpolation, eliminating artifacts while preserving the original image information.
NOLA: Compressing LoRA using Linear Combination of Random Basis
Soroush Abbasi Koohpayegani (University of California), Hamed Pirsiavash (University of California)
CodeCompressionOptimizationTransformerLarge Language ModelSupervised Fine-TuningImageText
π― What it does: NOLA achieves efficient compression and fine-tuning of LLM weights by reparameterizing the low-rank matrices in LoRA as a linear combination of random bases.
António Farinhas (Instituto de Telecomunicaçáes), Andre Martins
CodeClassificationOptimizationTextTime Series
π― What it does: A non-exchangeable conformal risk control method is proposed, which can provide an upper bound on the expected value of any monotonic loss function in the case of non-exchangeable data, and degenerates to traditional CRC when the data is exchangeable.
Yifei Wang (Peking University), Yisen Wang (Peking University)
CodeRetrievalExplainability and InterpretabilityRepresentation LearningContrastive LearningImage
π― What it does: Non-negative Contrastive Learning (NCL) is proposed, which enhances interpretability and sparsity by enforcing non-negativity of features in contrastive learning.
Nougat: Neural Optical Understanding for Academic Documents
Lukas Blecher (Meta AI), Robert Stojnic (Meta AI)
CodeRecognitionGenerationTransformerLarge Language ModelText
π― What it does: The Nougat model and its dataset generation pipeline are proposed, capable of directly converting the PDF pages of academic papers into a lightweight markup language (similar to LaTeX), without relying on traditional OCR.
Novel Quadratic Constraints for Extending LipSDP beyond Slope-Restricted Activations
Patricia Pauli (Institute for Systems Theory and Automatic Control, University of Stuttgart), Bin Hu (University of Illinois Urbana-Champaign)
CodeImage
π― What it does: This study investigates how to extend LipSDP to estimate the Lipschitz constants of neural networks using MaxMin, GroupSort, and Householder activation functions.
π― What it does: Proposes the NuwaDynamics framework, which discovers causal patches in self-supervised reconstruction tasks and enhances the model's robustness to different distributions through intervention patches.
π― What it does: This paper proposes a method that combines causal representation learning with object-centric learning, utilizing a Slot Attention-based object segmentation network and a weakly supervised sparse perturbation learning framework to achieve identifiable attribute decoupling in multi-object scenes.
π― What it does: This paper proposes the Slot Mixture Module (SMM), a slot attention mechanism based on Gaussian Mixture Models (GMM) to improve object-centric representation learning.
OctoPack: Instruction Tuning Code Large Language Models
Niklas Muennighoff, Shayne Longpre
CodeAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: This paper constructs a 4TB COMMITPACK dataset by utilizing the natural structure of Git commit messages and fine-tunes LLMs for code instructions. It proposes a multilingual evaluation benchmark, HUMANEVALPACK, covering code repair, explanation, and synthesis, and generates and releases two commercially usable code LLMs, OCTOCODER and OCTOGEEX, which perform best on this benchmark.
ODE Discovery for Longitudinal Heterogeneous Treatment Effects Inference
Krzysztof Kacprzyk (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
CodeOptimizationExplainability and InterpretabilityDrug DiscoveryTime SeriesBiomedical DataOrdinary Differential Equation
π― What it does: This paper proposes a method for transforming ordinary differential equation (ODE) discovery into a framework for inferring long-term heterogeneous treatment effects, and implements the INSITE method based on this framework, which can predict treatment effects in an interpretable manner and is robust to irregular sampling without using neural networks.
π― What it does: This paper proposes a Transformer-based ODEFormer model that can directly infer the symbolic expressions of multi-dimensional ordinary differential equations from a single noisy and irregularly sampled observation trajectory.
ODICE: Revealing the Mystery of Distribution Correction Estimation via Orthogonal-gradient Update
Liyuan Mao (Shanghai Jiao Tong University), Xianyuan Zhan (Institute for AI Industry Research)
CodeReinforcement LearningTabularBenchmark
π― What it does: The DICE (Distributionally Corrected Estimation) method has been researched and improved, introducing Orthogonal-Gradient Update in offline reinforcement learning and offline imitation learning.
Zifan Wu (Sun Yat-sen University), Dong Wang (Meituan)
CodeOptimizationSafty and PrivacyReinforcement LearningTabularBenchmark
π― What it does: This paper proposes an offline-prioritized master-slave safety reinforcement learning method called CAL, which integrates conservative policy optimization and local policy convexification.
Offline Data Enhanced On-Policy Policy Gradient with Provable Guarantees
Yifei Zhou (University of California), Wen Sun (Cornell University)
CodeReinforcement LearningImage
π― What it does: A new hybrid reinforcement learning algorithm is proposed, which integrates online policy updates based on natural policy gradients with offline data fitting for policy evaluation, thereby achieving the joint utilization of online and offline data.
OmniQuant: Omnidirectionally Calibrated Quantization for Large Language Models
Wenqi Shao (OpenGVLab), Ping Luo (The University of Hong Kong)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: A post-training quantization method named OmniQuant is proposed, which can improve the accuracy of large language models while maintaining low computation and memory usage.
π― What it does: This paper proposes Subset Adversarial Training (SAT), which generates adversarial samples only on a subset A of the training set, while the remaining subset B is not attacked, exploring the robustness transfer effects of this approach across different categories, samples, and downstream tasks.
π― What it does: Utilizing the time posterior distribution of the diffusion model to detect anomalies, the Diffusion Time Estimation (DTE) method is derived by simplifying DDPM;
On Double Descent in Reinforcement Learning with LSTD and Random Features
David Brellmann (ENSTA Paris), Goran Frehse (ENSTA Paris)
CodeReinforcement LearningSequential
π― What it does: Analyzes the theoretical performance of regularized LSTD (TD learning) based on random features in the dual asymptotic limit where the ratio of parameter dimensions to the number of visited states (i.e., model complexity) approaches infinity, revealing the double descent phenomenon;
Yangming Li (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
CodeGenerationData SynthesisDiffusion modelImage
π― What it does: This study investigates the error propagation problem in diffusion models (DM), proposing a theoretical framework and designing regularization methods to reduce cumulative errors and improve generation quality.
On Representation Complexity of Model-based and Model-free Reinforcement Learning
Hanlin Zhu (University of California Berkeley), Stuart Russell (University of California Berkeley)
CodeReinforcement LearningSequential
π― What it does: This study investigates the representation complexity of different functions (transition, reward, Q-function) in model-based and model-free reinforcement learning, proving that in certain classes of MDPs, transitions and rewards can be implemented using constant-depth polynomial-size circuits, while the optimal Q-function requires exponential-size circuits. Experiments in the MuJoCo environment further validate that the Q-function is more difficult to approximate with small networks compared to transitions and rewards.
π― What it does: The ROAD framework is proposed, combining distributionally robust optimization and adversarial learning to achieve local fairness and high accuracy on unknown subpopulations.
π― What it does: A configurable multimodal question-answering benchmark gCOG was constructed, and the performance of various benchmark neural networks was evaluated on three types of OOD generalization (interference, system combination, productive combination).
On the Humanity of Conversational AI: Evaluating the Psychological Portrayal of LLMs
Jen-tse Huang (Chinese University of Hong Kong), Michael Lyu
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: Constructed the PsychoBench framework to evaluate five LLMs (text-davinci-003, ChatGPT, GPT-4, LLaMA-2-7B, LLaMA-2-13B) on 13 clinical psychological scales (personality, relationships, motivation, emotions).
CodeKnowledge DistillationAdversarial AttackTransformerLarge Language ModelText
π― What it does: This paper studies the learnability of watermarking in language models, proposing that through knowledge distillation, models can naturally generate watermarked text based solely on weights, thus validating the feasibility of learning watermarks.
π― What it does: This paper studies the calibration limitations of temperature scaling when there is overlap in class distributions and proposes that using Mixup (including the extended d-Mixup) during the training phase can significantly improve the model's calibration performance.
On the Markov Property of Neural Algorithmic Reasoning: Analyses and Methods
Montgomery Bohde (Texas A&M University), Shuiwang Ji (Texas A&M University)
CodeGraph Neural NetworkGraphBenchmark
π― What it does: This paper proposes ForgetNet and its improved version G-ForgetNet, aimed at eliminating the dependence on historical embeddings in neural algorithm reasoning models, ensuring consistency with the Markov property of algorithm reasoning; by introducing a gating mechanism and regularization, G-ForgetNet achieves more stable gradients in the early training phase and gradually converges to the Markov mode;
π― What it does: By studying the memory effect of deep networks, a unified concept of 'over-memorization' is proposed, and a Distraction Over-Memorization (DOM) framework is designed to prevent the model from over-memorizing high-confidence training samples, thereby alleviating overfitting in three scenarios: natural training, robust training, and catastrophic overfitting.
π― What it does: This paper studies the role of discrete tokenization in self-supervised masked image modeling and proposes a clustering-based discrete tokenization method called ClusterMIM.
π― What it does: This paper proposes a method that transforms the semi-definite programming (SDP) estimation of the network Lipschitz constant into an eigenvalue optimization problem (EP-LipSDP) and implements a GPU-friendly first-order subgradient solver called LipDiff. It also introduces techniques such as Lanczos approximation, sparse matrix multiplication, and analytical initialization, enabling efficient computation of Lipschitz upper bounds on large networks (e.g., AlexNet on ImageNet).
On the Stability of Expressive Positional Encodings for Graphs
Yinan Huang (Georgia Institute of Technology), Pan Li (Georgia Institute of Technology)
CodeDrug DiscoveryGraph Neural NetworkGraph
π― What it does: A stable and expressive position information encoding method based on graph Laplacian eigenvectors, called SPE, is proposed to address the non-uniqueness and instability issues of traditional encodings.
One-hot Generalized Linear Model for Switching Brain State Discovery
Chengrui Li (Georgia Institute of Technology), Anqi Wu (Georgia Institute of Technology)
CodeTime Series
π― What it does: A one-hot encoded hidden Markov model generalized linear model (one-hot HMM-GLM) is proposed, which estimates the functional connectivity that varies over time across multiple states by decomposing the weight matrix into a discrete adjacency matrix and a positive intensity matrix, and sharing a Gumbel-Softmax prior for the adjacency matrix.
π― What it does: This paper proposes a robot for interactive instruction following that can continuously learn new behaviors and new environments after deployment, addressing the limitations of traditional prior data learning scenarios.
Online GNN Evaluation Under Test-time Graph Distribution Shifts
Xin Zheng (Monash University), Shirui Pan (Griffith University)
CodeGraph Neural NetworkGraph
π― What it does: In an online deployment environment, an online GNN evaluation framework is proposed to estimate the generalization error of a trained GNN on unlabeled, distribution-drift test graphs.
π― What it does: Proposes Online Spike Re-normalization (OSR) and Online Threshold Stabilizer (OTS), achieving memory efficiency and biological interpretability in SNN online training without using future information.
π― What it does: This paper proposes Temporally-Correlated Episodic RL (TCE), a reinforcement learning framework that combines temporal correlation with periodic exploration, balancing the trajectory smoothness of ERL with the data efficiency of SRL.
Open-ended VQA benchmarking of Vision-Language models by exploiting Classification datasets and their semantic hierarchy
Simon Ging (University of Freiburg), Thomas Brox (University of Freiburg)
CodeClassificationRecognitionTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
π― What it does: This paper proposes a method to convert traditional classification datasets into an open visual question answering (oVQA) benchmark, and follows up with questions at the semantic hierarchy level to eliminate answer ambiguity.
OpenChat: Advancing Open-source Language Models with Mixed-Quality Data
Guan Wang (Tsinghua University), Yang Liu (Tsinghua University)
CodeTransformerLarge Language ModelReinforcement LearningText
π― What it does: Proposes the OpenChat framework, using Conditioned-RLFT for RL-free training on mixed quality data to enhance the performance of open-source LLMs.