International Conference on Learning Representations Β· 1682 papers
Neural Exploratory Landscape Analysis for Meta-Black-Box-Optimization
Zeyuan Ma (South China University of Technology), Yue-Jiao Gong (South China University of Technology)
CodeOptimizationMeta LearningTransformerNeural Radiance Field
π― What it does: This paper studies a learnable exploration landscape analyzer, NeurELA, which dynamically extracts low-level optimization states using an end-to-end two-stage attention network, replacing traditional handcrafted features to enhance the performance of Meta-black-box optimization.
π― What it does: This study proposes a unified framework for Neural Interactive Proofs, designs various new protocols (nip, mnip, zk-nip, zk-mnip), provides theoretical equivalence proofs, and conducts experimental evaluations on two major tasks: graph isomorphism and code verification.
π― What it does: A Graph-Image Multimodal Fusion (GIMF) framework is proposed to enhance the performance of neural multi-objective combinatorial optimization by constructing instance images and jointly learning with graph structures.
Neural networks on Symmetric Spaces of Noncompact Type
Xuan Son Nguyen (CY Cergy Paris University), Aymeric Histace (CY Cergy Paris University)
CodeClassificationImageTime Series
π― What it does: A unified framework is proposed for constructing the distance from a point to a hyperplane in non-compact symmetric spaces (including hyperplane spaces and SPD manifolds), and based on this, a fully connected layer and attention mechanism are designed to build a novel neural network.
π― What it does: Proposes to express the Transformer architecture as a non-autonomous neural ODE (DiffEqTransformer), generating time-dependent weights for attention and feedforward layers, and using an ODE solver to achieve adaptive layer counts;
π― What it does: A Neuralized Markov Random Field (Neuralized MRF) model is proposed to simultaneously capture the Markov evolution of individual movements and the effects of group interactions, thereby achieving stochastic human trajectory prediction with interaction awareness.
NeurFlow: Interpreting Neural Networks through Neuron Groups and Functional Interactions
Tue Minh Cao, My T. Thai (University of Florida)
CodeExplainability and InterpretabilityConvolutional Neural NetworkLarge Language ModelImage
π― What it does: An automated framework called NeurFlow is proposed, which uses neuron groups instead of individual neurons to explain the internal mechanisms of CNNs and constructs inter-layer interaction circuits.
NeuroLM: A Universal Multi-task Foundation Model for Bridging the Gap between Language and EEG Signals
Weibang Jiang, Dongsheng Li (Microsoft Research Asia)
CodeRecognitionAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderMultimodalityTime SeriesBiomedical DataElectrocardiogram
π― What it does: This study proposes NeuroLM, a multi-task foundational model that treats EEG signals as a 'foreign language', utilizing large language models (LLM) to achieve a unified EEG processing framework capable of completing six different EEG tasks at once.
Neuron Platonic Intrinsic Representation From Dynamics Using Contrastive Learning
Wei Wu (Peking University), Jinzhuo Wang (Peking University)
CodeRepresentation LearningSpiking Neural NetworkContrastive LearningTime SeriesBiomedical Data
π― What it does: Proposes the NeurPIR framework, which learns time-invariant intrinsic representations of neurons from dynamic data of neuronal populations through contrastive learning.
NExT-Mol: 3D Diffusion Meets 1D Language Modeling for 3D Molecule Generation
Zhiyuan Liu (National University of Singapore), Tat-Seng Chua (National University of Singapore)
CodeGenerationData SynthesisDrug DiscoveryTransformerLarge Language ModelDiffusion modelGraph
π― What it does: This paper proposes a two-step 3D molecular generation framework NExT-Mol based on a large 1D SELFIES language model MoLlama and a 3D diffusion model DMT, utilizing pre-trained 1D representations to enhance 3D predictions through cross-modal projection, and improving generation diversity through random SELFIES data augmentation.
No Preference Left Behind: Group Distributional Preference Optimization
Binwei Yao (Stanford University), Junjie Hu (University of Wisconsin Madison)
CodeRecommendation SystemOptimizationLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper studies a new framework called GDPO, which allows large language models to generate diverse responses according to group preference distributions.
π― What it does: A framework for Noise Partial Label Learning (NPLL) is proposed, which reduces the noise rate and shortens the candidate label set length through sample separation and candidate label set reconstruction, thereby enhancing the generalization performance of the classifier.
Noisy Test-Time Adaptation in Vision-Language Models
Chentao Cao (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
CodeDomain AdaptationAnomaly DetectionTransformerVision Language ModelImageMultimodalityBenchmark
π― What it does: A zero-shot noise testing adaptation (ZS-NTTA) framework is proposed, and the AdaND method is designed for this task, utilizing a pre-trained vision-language model (CLIP) to freeze the classifier, a single-layer adaptive noise detector with dynamic thresholds, and injecting Gaussian noise into clean data streams to avoid misjudgments.
Non-Adversarial Inverse Reinforcement Learning via Successor Feature Matching
Arnav Kumar Jain (Mila Quebec AI Institute), Sanjiban Choudhury (Mila Quebec AI Institute)
CodeReinforcement LearningSequential
π― What it does: Inverse reinforcement learning is achieved by directly matching the expert's Successor Features, without the need for adversarial reward learning, and can use only the expert's state sequences for imitation.
Not All Heads Matter: A Head-Level KV Cache Compression Method with Integrated Retrieval and Reasoning
Yu Fu (University of California), Wen Xiao (Microsoft)
CodeRetrievalCompressionTransformerLarge Language ModelText
π― What it does: This paper proposes a KV cache compression method for attention heads, HeadKV-R2, which significantly reduces KV cache usage while maintaining the long text reasoning capabilities of LLMs.
Not All Language Model Features Are One-Dimensionally Linear
Joshua Engels (Massachusetts Institute of Technology), Max Tegmark (Massachusetts Institute of Technology)
CodeExplainability and InterpretabilityTransformerLarge Language ModelAuto EncoderText
π― What it does: This paper first provides a strict definition of multidimensional features and uses Sparse Autoencoders (SAE) to automatically retrieve interpretable multidimensional features from the hidden layers of GPT-2, Mistral-7B, and Llama-3-8B. Through clustering and visualization, they discovered features with a circular distribution (such as the seven days of the week and twelve months). Subsequently, the authors designed circular subspace intervention experiments based on activation patches, demonstrating that these circular features play a causal role in modular arithmetic tasks (such as 'What day is it seven days after Monday?') and compared different models and layers.
Not All LLM-Generated Data Are Equal: Rethinking Data Weighting in Text Classification
Hsun-Yu Kuo (Swiss Federal Institute of Technology in Lausanne), Pu-Jen Cheng (National Taiwan University)
CodeClassificationKnowledge DistillationData-Centric LearningTransformerLarge Language ModelTextFinance Related
π― What it does: Two weighted loss functions (IMP-Loss and DIMP-Loss) are proposed, which dynamically or statically weight the synthetic data generated by LLM through a quality checker and a diversity checker to align it with the real data distribution, thereby improving the performance of text classification models.
NoVo: Norm Voting off Hallucinations with Attention Heads in Large Language Models
Zheng Yi Ho, Dacheng Tao (Nanyang Technological University)
CodeTransformerLarge Language ModelText
π― What it does: The Norm Voting (NoVo) method is proposed, which utilizes the L2 norm of attention heads for truth voting in zero-shot multiple-choice questions, significantly reducing hallucinations in large language models and improving factual accuracy.
CodeGenerationData SynthesisExplainability and InterpretabilityGenerative Adversarial NetworkTabular
π― What it does: This paper proposes an energy-based generative model called NRGBoost, which transforms gradient boosting trees into a generative model. It uses second-order approximate likelihood maximization during training and supports conditional inference for arbitrary variables and handling of missing values.
Number Cookbook: Number Understanding of Language Models and How to Improve It
Haotong Yang (Peking University), Muhan Zhang (Peking University)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
π― What it does: This paper constructs the NUPA Test benchmark, systematically evaluating LLMs on 41 basic numerical understanding and processing tasks under four types of numerical representations: integers, floating-point numbers, fractions, and scientific notation, and conducts zero-shot testing on various large models. It also explores the enhancement effects of tokenizers, positional encodings, numerical format modifications during the pre-training phase, post-training fine-tuning, and the Chain-of-Thought method on NUPA.
NutriBench: A Dataset for Evaluating Large Language Models in Nutrition Estimation from Meal Descriptions
Mehak Preet Dhaliwal (University of California), Yao Qin (University of California)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Developed the NUTRIBENCH benchmark set, which includes 11,857 manually verified natural language meal descriptions, and evaluated the performance of 12 LLMs on carbohydrate estimation tasks.
π― What it does: Using a pre-trained large video diffusion model, a zero-shot novel view synthesis method is proposed, which can generate images from arbitrary viewpoints from a single view, a sequence of views, or a monocular video without additional training.
OATS: Outlier-Aware Pruning Through Sparse and Low Rank Decomposition
Stephen Zhang (University of Toronto), Vardan Papyan (University of Toronto)
CodeCompressionAnomaly DetectionComputational EfficiencyTransformerLarge Language ModelImageText
π― What it does: A model compression method that does not require retraining is proposedβOATS, which decomposes the transformer weight matrix into the sum of a sparse matrix and a low-rank matrix, and utilizes the second-order moment of input embeddings for weight scaling, preserving key features of the model;
CodeClassificationRecognitionRetrievalTransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
π― What it does: A comprehensive benchmark for the retrieval and interpretation of ancient script stone tablets, named OBI-Bench, has been proposed. This benchmark evaluates the performance of 23 large-scale multimodal models (LMM) on five major tasks: recognition, stitching, classification, retrieval, and interpretation.
π― What it does: A mixed reasoning framework OCCAM is proposed, which dynamically allocates different capacities of classifiers for different queries, thereby maximizing overall accuracy while satisfying user budget constraints.
Occlusion-aware Non-Rigid Point Cloud Registration via Unsupervised Neural Deformation Correntropy
Mingyang Zhao (Hong Kong Institute of Science and Innovation), Dong-ming Yan
CodeOptimizationPoint CloudBiomedical Data
π― What it does: This paper proposes an unsupervised occlusion-adaptive non-rigid point cloud registration method called OAR, which achieves physically reasonable registration of occluded areas using the maximum mutual information criterion and local linear reconstruction.
OccProphet: Pushing the Efficiency Frontier of Camera-Only 4D Occupancy Forecasting with an Observer-Forecaster-Refiner Framework
Junliang Chen (Hong Kong Polytechnic University), Lap-Pui Chau (Hong Kong Polytechnic University)
CodeDepth EstimationAutonomous DrivingComputational EfficiencyConvolutional Neural NetworkSimultaneous Localization and MappingImage
π― What it does: A camera-based 4D occupancy prediction framework named OccProphet is proposed, which efficiently predicts future 3D occupancy states using an Observer-Forecaster-Refiner three-step pipeline.
Offline Model-Based Optimization by Learning to Rank
Rong-Xi Tan (Nanjing University), Chao Qian (Nanjing University)
CodeOptimizationReinforcement LearningTabular
π― What it does: A learning-to-rank based offline model optimization framework RaM is proposed, replacing the traditional MSE regression surrogate and directly optimizing the relative ranking of designs.
π― What it does: A smooth Q-function out-of-distribution (OOD) generalization method for offline reinforcement learning (RL) using Convex Hull Neighborhood (CHN) is proposed, which enhances the Q-value estimation in the OOD region through the Smooth Bellman Operator.
Niklas Muennighoff (Allen Institute for AI), Hannaneh Hajishirzi (Allen Institute for AI)
CodeTransformerLarge Language ModelMixture of ExpertsText
π― What it does: Proposed and open-sourced the sparse Mixture-of-Experts language model OLMOE-1B-7B and its instruction version for open-source research.
π― What it does: A reverse rendering plugin module OMG based on 3D Gaussian Splatting is proposed, which couples material with opacity using the Beer-Lambert law.
OmniCorpus: A Unified Multimodal Corpus of 10 Billion-Level Images Interleaved with Text
Qingyun Li (Shanghai AI Laboratory), Jifeng Dai (Tsinghua University)
CodeData SynthesisRetrievalTransformerLarge Language ModelImageTextMultimodality
π― What it does: Constructed and released the OmniCorpus, an open multimodal document dataset with a scale of 10B (8.6B images, 1.696T text tokens, 2.2B documents), providing a unified streaming format and an efficient data engine.
π― What it does: The OmniSep model is proposed, achieving audio separation based on arbitrary single-modal (text, image, audio) or multi-modal combination queries, and supports negative queries and open vocabulary queries.
π― What it does: Developed the Omni Γ R benchmark to evaluate the performance of multimodal language models in cross-modal reasoning, providing synthetic and real data subsets.
On a Connection Between Imitation Learning and RLHF
Teng Xiao (Pennsylvania State University), Vasant G Honavar (Pennsylvania State University)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
π― What it does: This paper re-examines RLHF from the perspective of imitation learning, proposing the DIL framework, which directly optimizes the reverse KL imitation learning objective and estimates the density ratio through Bregman divergence, thereby achieving efficient preference alignment.
CodeLarge Language ModelReinforcement LearningText
π― What it does: This paper studies the trade-off relationship between information quantity (measured in bits) and cumulative loss (return) in Bayesian interactive decision-making problems, proposing a general information-theoretic method to obtain lower and upper bounds, and validating this theory in practical tasks.
On Calibration of LLM-based Guard Models for Reliable Content Moderation
Hongfu Liu (National University of Singapore), Ye Wang (National University of Singapore)
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This study investigates the confidence calibration and reliability of guardian models based on large language models (LLM) in content moderation, systematically evaluating the ECE and F1 of 9 open-source guardian models across 12 public benchmarks, and discusses post-hoc calibration methods.
On Conformal Isometry of Grid Cells: Learning Distance-Preserving Position Embedding
Dehong Xu (University of California, Los Angeles), Ying Nian Wu (University of California, Los Angeles)
CodeRecurrent Neural NetworkSequential
π― What it does: This paper proposes and verifies that the reason for the formation of hexagonal lattice patterns in grid cell response maps is that their neural spatial mapping satisfies the assumption of local distance-preserving conformal isometry.
π― What it does: This paper proposes a discriminative probability modeling framework for continuous domains, combining multiple importance sampling (MIS) to address the partition function integral challenge, and based on this, designs a new non-parametric approximation method and corresponding contrastive loss;
On Disentangled Training for Nonlinear Transform in Learned Image Compression
Han Li (Shanghai Jiao Tong University), Hongkai Xiong (Shanghai Jiao Tong University)
CodeCompressionImage
π― What it does: An auxiliary linear transformation, AuxT, is proposed to help learn image compression models achieve energy compression through feature decorrelation and non-uniform energy modulation during the training phase, significantly accelerating training.
On Generalization Across Environments In Multi-Objective Reinforcement Learning
Jayden Teoh (Singapore Management University), Peter Vamplew (Federation University Australia)
CodeOptimizationReinforcement LearningBenchmark
π― What it does: This paper proposes a framework for environment generalization in Multi-Objective Reinforcement Learning (MORL) and establishes a new MORL generalization benchmark and evaluation metrics, subsequently evaluating various state-of-the-art (SOTA) algorithms on this benchmark.
On Linear Representations and Pretraining Data Frequency in Language Models
Jack Merullo (Brown University), Yanai Elazar (Allen Institute for AI)
CodeTransformerLarge Language ModelText
π― What it does: Investigate the correlation between the linear representation of factual relationships in language models and word frequency in pre-training corpora, and construct a regression model to predict word frequency based on linear representation.
π― What it does: This paper proposes a robust reinforcement learning method for the problem of partial observability caused by observation disturbances, defining the Adversarial Counterfactual Error (ACoE) and introducing a scalable C-ACoE objective. It then maximizes rewards while minimizing C-ACoE within the PPO/DQN framework to enhance adversarial robustness.
On Quantizing Neural Representation for Variable-Rate Video Coding
Junqi Shi (Nanjing University), Zhan Ma (Nanjing University)
CodeCompressionVideo
π― What it does: Developed NeuroQuant, a post-training quantization method for variable bitrate implementation of non-general implicit neural representation video coding;
Bernd Frauenknecht (RWTH Aachen University), Sebastian Trimpe (RWTH Aachen University)
CodeReinforcement LearningTabular
π― What it does: This paper proposes the Infoprop mechanism, which reduces model error accumulation in model-based reinforcement learning by distinguishing between naive and aleatoric uncertainty, achieving longer and more reliable simulation rollouts.
π― What it does: The Grendel system is proposed, achieving distributed training of 3D Gaussian Splatting on multiple GPUs, supporting large batch views, dynamic load balancing, and sparse All-to-All communication.
Jin Peng Zhou (Cornell University), Kilian Q Weinberger
CodeLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes two adaptive evaluation methods based on multi-armed bandits to quickly identify the optimal LLM or prompt under a limited budget.
π― What it does: This paper proposes a testing-time data poisoning (RTTDP) framework that is more aligned with practical deployment scenarios. It designs a complete protocol for gray-box attacks, without the need to access other normal samples, online attack sequences, and attack budget constraints. Based on this, it develops adaptive 'in-distribution' poisoning methods, feature consistency regularization, and two attack targets for testing-time adaptation (TTA) methods (high-entropy attack and low-entropy attack).
π― What it does: This study quantifies the vulnerability of test-time adaptive (TTA) methods to adversarial attacks in the absence of labels and proposes a novel attack algorithm that does not rely on true labels (Feature Collapse Attack, FCA).
On the Crucial Role of Initialization for Matrix Factorization
Bingcong Li (ETH Zurich), Niao He (ETH Zurich)
CodeGenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelImageText
π― What it does: This paper proposes an initialization method based on Nystrom sampling and applies it to the fine-tuning of ScaledGD and LoRA, significantly improving the convergence speed and final performance of low-rank matrix decomposition and LLM/diffusion models.
π― What it does: A new framework based on HΓΆlder stability expectations is proposed to evaluate the separation quality of multi-set and graph neural networks, and SortMPNN, a sorting-based aggregation MPNN, is designed.
π― What it does: Proposes the FedGLCL framework, which employs language-driven contrastive learning in federated learning, aligning global text embeddings with local image features, replacing traditional label-driven training to address the performance degradation caused by non-IID data.
On the Optimization Landscape of Low Rank Adaptation Methods for Large Language Models
Xu-Hui Liu (Nanjing University), Yang Yu (Nanjing University)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper analyzes the optimization landscape of low-rank adaptation methods through theoretical analysis and experimental evaluation, and proposes a new algorithm called GaRare.
On the Performance Analysis of Momentum Method: A Frequency Domain Perspective
Xianliang Li (Shenzhen Institutes of Advanced Technology Chinese Academy of Sciences), Sheng Xu (Shenzhen Institutes of Advanced Technology Chinese Academy of Sciences)
CodeOptimizationReinforcement LearningImageText
π― What it does: By viewing the momentum method as a time-varying filter, a frequency domain analysis framework is proposed to explore the impact of different momentum coefficients on high and low frequency gradients, and based on this, an optimizer FSGDM is designed to dynamically adjust the momentum filtering characteristics, verifying its performance improvement in multiple tasks.
On the Role of Attention Heads in Large Language Model Safety
Zhenhong Zhou (Tongyi Lab), Yongbin Li (Tongyi Lab)
CodeSafty and PrivacyTransformerLarge Language ModelText
π― What it does: This paper quantifies and attributes the attention heads responsible for safety in large language models by defining the Attention Head Importance Score (Ships) and the Safety Attention Head Attribution Algorithm (Sahara);
On-the-fly Preference Alignment via Principle-Guided Decoding
Mingye Zhu (University of Science and Technology of China), Zhendong Mao (University of Science and Technology of China)
CodeRecommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
π― What it does: A framework named OPAD is proposed for preference alignment during inference without fine-tuning. It dynamically adjusts token probabilities through a principle-guided reward mechanism, enabling the model to adhere to user preferences and principles during the inference phase.
π― What it does: The PAWL algorithm has been designed and implemented for efficiently computing the partial Wasserstein distance in one-dimensional space, and a slicing strategy for Partial OT has been proposed, which can obtain the solution for all transport quantities at once.
One Model Transfer to All: On Robust Jailbreak Prompts Generation against LLMs
Linbao Li (Harbin Institute of Technology), YU LI
CodeGenerationAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: The ArrAttack framework is proposed, which utilizes rewriting attacks and a general robustness evaluation model to automatically generate jailbreak prompts that can bypass various defense strategies, and quickly produces high-quality, semantically consistent robust prompts through fine-tuning of the generative model.
π― What it does: A single network, single-stage training Shortcut Models is proposed, capable of generating high-quality images at any number of steps (including single-step).
π― What it does: This paper proposes a training-free consistency text-to-image generation method called One-Prompt-One-Story. By concatenating identity prompts with multiple frame descriptions into a single long prompt, and applying singular value reweighting and identity-preserving cross-attention on the prompt embeddings during generation, it can maintain the identity consistency of the same subject and text alignment across different scenes.
Online Reinforcement Learning in Non-Stationary Context-Driven Environments
Pouya Hamadanian (Massachusetts Institute of Technology), Mohammad Alizadeh (Massachusetts Institute of Technology)
CodeOptimizationReinforcement LearningTabularTime Series
π― What it does: An online reinforcement learning algorithm LCPO is proposed, which suppresses catastrophic forgetting through local constraint optimization using observed non-stationary contexts.
Open-Set Graph Anomaly Detection via Normal Structure Regularisation
Qizhou Wang (University of Melbourne), Christopher Leckie (University of Melbourne)
CodeAnomaly DetectionGraph Neural NetworkGraph
π― What it does: A new open set graph anomaly detection method called NSReg is proposed, which enhances the model's generalization to unseen anomalies through normal structure regularization.
OpenHands: An Open Platform for AI Software Developers as Generalist Agents
Xingyao Wang (University of Illinois Urbana-Champaign), Graham Neubig (Carnegie Mellon University)
CodeAI Code AssistantTransformerLarge Language ModelAgentic AITextBenchmark
π― What it does: An open-source platform called OpenHands has been built, supporting AI agents to interact with the environment through software interfaces such as code, command line, and browser, and providing a framework for multi-agent collaboration and unified evaluation.
OpenMathInstruct-2: Accelerating AI for Math with Massive Open-Source Instruction Data
Shubham Toshniwal (NVIDIA), Igor Gitman (NVIDIA)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
π― What it does: Created the OpenMathInstruct-2 public mathematical reasoning dataset (14M problem-solution pairs) and fine-tuned it on the Llama3.1 base model, resulting in the high-performance OpenMath2-Llama3.1-8B/70B models.
OPTAMI: Global Superlinear Convergence of High-order Methods
Dmitry Kamzolov (Mohamed bin Zayed University of Artificial Intelligence), Martin TakΓ‘Δ (Mohamed bin Zayed University of Artificial Intelligence)
CodeOptimizationTabular
π― What it does: The OPTAMI high-order optimization library is proposed, with a new NATA (Nesterov Accelerated Tensor Method with A_t Adaptation) designed, along with a global superlinear convergence theory based on high-order methods, which has been validated in practice for effectiveness.
OptiBench Meets ReSocratic: Measure and Improve LLMs for Optimization Modeling
Zhicheng Yang (Hong Kong University of Science and Technology), Jing Tang (Hong Kong University of Science and Technology)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningTextTabularBenchmark
π― What it does: An optimization modeling benchmark for LLMs, OPTIBENCH, has been constructed, and a reverse Socratic data synthesis method, ReSocratic, has been proposed to generate a large number of optimization problems.
π― What it does: A novel network pruning method called Optimal Brain Apoptosis (OBA) is proposed, which directly computes the Hessian vector product to achieve parameter importance assessment for efficient pruning.
Optimal Flow Transport and its Entropic Regularization: a GPU-friendly Matrix Iterative Algorithm for Flow Balance Satisfaction
Liangliang Shi (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
CodeOptimizationGraph
π― What it does: A framework for solving Optimal Flow Transport (OFT) on general graphs is proposed, and a GPU-friendly Sinkhorn iterative algorithm (OFT-Sinkhorn) is obtained through entropy regularization, along with an EOFT-Sinkhorn with capacity constraints, to solve the minimum cost flow problem.
Optimistic Games for Combinatorial Bayesian Optimization with Application to Protein Design
Melis Ilayda Bal (Max Planck Institute for Intelligent Systems), Andreas Krause (Google DeepMind)
CodeOptimizationDrug DiscoveryProtein Structure PredictionBiomedical Data
π― What it does: This paper proposes a game-theory-based combinatorial Bayesian optimization method called GAMEOPT, which treats discrete variables as players in a cooperative game, using the Upper Confidence Bound (UCB) as a reward function to compute the game equilibrium point for iterative selection of evaluation samples.
Optimized Multi-Token Joint Decoding With Auxiliary Model for LLM Inference
Zongyue Qin (University of California), Yizhou Sun (California Institute of Technology)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: This paper proposes Multi-Word Joint Decoding (MTJD) and its efficient approximate version, Multi-Word Assisted Decoding (MTAD), which achieves the one-time generation of multiple words by sampling on a small model and validating on a large model, while ensuring output quality.
Optimizing Posterior Samples for Bayesian Optimization via Rootfinding
Taiwo Adebiyi, Ruda Zhang (University of Houston)
CodeOptimization
π― What it does: This paper proposes a global optimization strategy named TS-roots for gradient multi-start optimization of posterior sample paths in high-dimensional Bayesian optimization.
ORSO: Accelerating Reward Design via Online Reward Selection and Policy Optimization
Chen Bo Calvin Zhang (Massachusetts Institute of Technology), Pulkit Agrawal
CodeOptimizationRobotic IntelligenceLarge Language ModelReinforcement LearningSequential
π― What it does: This study proposes a framework for online reward selection and strategy optimization (ORSO) that automatically selects the optimal reward function from a set of candidate reward functions to accelerate reward design in reinforcement learning.
π― What it does: A linear oscillation state space model (LinOSS) based on the dynamics of a forced harmonic oscillator is proposed, and it is demonstrated that it can maintain stability, interpretability, and universal approximation capability in long sequence learning.
OSTQuant: Refining Large Language Model Quantization with Orthogonal and Scaling Transformations for Better Distribution Fitting
Xing Hu (Houmo AI), Sifan Zhou
CodeTransformerLarge Language ModelText
π― What it does: This paper proposes OSTQuant, which utilizes learnable orthogonal and scaling transformations for post-training quantization of LLM weights and activations.
π― What it does: A Lagrangian multiplier framework based on Total Variation (TV) called OOD-TV-IRM is proposed, which constructs an adversarial learning process of primal-dual optimization and semi-Nash equilibrium, aimed at enhancing the model's out-of-distribution (OOD) generalization ability.
Outlier Synthesis via Hamiltonian Monte Carlo for Out-of-Distribution Detection
Hengzhuang Li (Huazhong University of Science and Technology), Teng Zhang (Huazhong University of Science and Technology)
CodeAnomaly DetectionContrastive LearningImage
π― What it does: The HamOS framework is proposed, which utilizes Hamiltonian Monte Carlo to generate diverse and representative virtual anomaly samples in the unit sphere feature space, and uses them for training to enhance OOD detection performance.
Overcoming Lower-Level Constraints in Bilevel Optimization: A Novel Approach with Regularized Gap Functions
Wei Yao (National Center for Applied Mathematics Shenzhen), Jin Zhang (Southern University of Science and Technology)
CodeOptimizationTabularBiomedical Data
π― What it does: This paper proposes a single-loop, Hessian-free solver BiC-GAFFA for solving lower-level constraint-coupled bilevel optimization problems, and extends it to lower-level min-max problems.
π― What it does: This study addresses the problem of Zero-Shot Cooperation (ZSC) in the Overcooked environment, proposing a state-enhanced training method and designing a new version, OvercookedV2, as a more challenging ZSC benchmark.
π― What it does: An end-to-end open vocabulary multi-object tracking framework (OVTR) is proposed, capable of achieving continuous tracking and classification without the need for candidate boxes and post-processing.
π― What it does: A scalable spiking network based on a probabilistic state space model (PβSpikeSSM) is proposed, which utilizes a SpikeSampler layer to randomly generate spikes and achieves multi-layer parallel communication through SpikeMixer and ClampFuse.
π― What it does: A new parameter-efficient fine-tuning method called PaCA is proposed, which fine-tunes by randomly selecting a portion of connections in the pre-trained weights, avoiding the sequential processing of adapter layers.
CodeRetrievalSafty and PrivacyComputational EfficiencyGraph Neural Network
π― What it does: A scheme for performing approximate nearest neighbor search on a massive vector database without disclosing the query vector is provided;
Painting with Words: Elevating Detailed Image Captioning with Benchmark and Alignment Learning
Qinghao Ye (ByteDance Research), Haoqi Fan
CodeGenerationOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextBenchmark
π― What it does: This paper proposes the DCSCORE fine-grained image description evaluation metric, the DECAPBENCH detailed image description benchmark, and designs the FEEDQUILL fine-grained feedback collection and reinforcement learning method to improve the image description quality of VLMs and reduce hallucinations.
π― What it does: A new algorithm called Pairwise Elimination (PE) and its extension PE-CS are proposed and analyzed for the cost-subsidized multi-armed bandit (MAB-CS) problem, specifically addressing two constraint scenarios: known reference arms and optimal subsidy rewards.
PaLD: Detection of Text Partially Written by Large Language Models
Eric Lei (University of Pennsylvania), Chun-Fu Chen
CodeClassificationRecognitionOptimizationTransformerLarge Language ModelText
π― What it does: This paper proposes the Partial-LLM Detector (PaLD), which can estimate the proportion of LLM-generated content in mixed text and locate LLM paragraphs.
PALMBENCH: A COMPREHENSIVE BENCHMARK OF COMPRESSED LARGE LANGUAGE MODELS ON MOBILE PLATFORMS
Yilong Li (University of Wisconsin Madison), Suman Banerjee (University of Wisconsin Madison)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextMultimodalityBenchmark
π― What it does: This paper presents PalmBench, a lightweight automated LLM benchmark framework for mobile devices, designed to evaluate compressed models on different mobile platforms in terms of memory, power consumption, throughput, as well as accuracy, toxicity, and hallucination metrics.
Palu: KV-Cache Compression with Low-Rank Projection
Chi-Chih Chang (National Yang Ming Chiao Tung University), Kai-Chiang Wu (National Yang Ming Chiao Tung University)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelText
π― What it does: A post-training KV-Cache compression framework called Palu is proposed, which significantly reduces KV-Cache memory usage and accelerates inference while maintaining accuracy by performing low-rank decomposition on Key/Value projection weights and caching low-dimensional latent representations.
Parameter and Memory Efficient Pretraining via Low-rank Riemannian Optimization
Zhanfeng Mo (Nanyang Technological University), Sinno Jialin Pan (Chinese University of Hong Kong)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: The researchers propose LORO, a low-rank Riemannian optimizer capable of pre-training low-rank parameterized language models from scratch.
π― What it does: Transform the sequential sampling process of diffusion models into solving banded nonlinear equations to achieve hierarchical parallel sampling, proposing the ParaSolver framework.
Guang Zhao (Brookhaven National Laboratory), Xiaoning Qian (Texas A&M University)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: A reinforcement learning framework based on multi-objective dominance relations (ParetoPrompt) is proposed for the automatic generation of prompts on the Pareto front.
ParetoFlow: Guided Flows in Multi-Objective Optimization
Ye Yuan (McGill), Xue Liu (MILA - Quebec AI Institute)
CodeOptimizationFlow-based ModelTabularBenchmark
π― What it does: A flow matching-based offline multi-objective optimization framework called ParetoFlow is proposed, which utilizes a unified weight vector and neighborhood evolution to guide sampling in approximating the Pareto front.
ParFam -- (Neural Guided) Symbolic Regression via Continuous Global Optimization
Philipp Scholl (LMU Munich), Gitta Kutyniok (LMU Munich)
CodeOptimizationTransformerTabularBenchmarkPhysics Related
π― What it does: The ParFam method (and its pre-trained version DL-ParFam) is proposed, which transforms symbolic regression from discrete search to continuous optimization using parameterized rational function networks, and solves it through global optimization.
π― What it does: The Partial Gromov-Wasserstein (PGW) problem is proposed, its metric properties in metric spaces are proven, and two Frank-Wolfe solvers are provided; subsequently, its performance is validated in tasks such as shape matching, retrieval, and interpolation.