ICLR 2024 Papers — Page 12
International Conference on Learning Representations · 2260 papers
Learning Decentralized Partially Observable Mean Field Control for Artificial Collective Behavior
Kai Cui (Technische Universitaet Darmstadt), Heinz Koeppl (Technische Universitaet Darmstadt)
Robotic IntelligenceReinforcement Learning
🎯 What it does: A scalable decentralized, partially observable multi-agent reinforcement learning framework is proposed, engineering collective behavior based on mean field control (MFC);
Learning Delays in Spiking Neural Networks using Dilated Convolutions with Learnable Spacings
Ilyass Hammouamri (Centre de Recherche Cerveau et Cognition), Timothée Masquelier (Centre de Recherche Cerveau et Cognition)
Spiking Neural NetworkTime SeriesAudio
🎯 What it does: Utilize learnable dilated convolution (DCLS) to learn delays in deep feedforward spiking neural networks, obtaining a unique discrete delay for each synapse at the end of training;
Learning dynamic representations of the functional connectome in neurobiological networks
Luciano Dyballa (Yale University), Steven W. Zucker (Yale University)
Time SeriesBiomedical Data
🎯 What it does: An unsupervised method is proposed to construct a functional connectivity map of C. elegans by calculating the differential affinity of neuronal activity over time and extracting dynamic communities using Non-negative Tensor Factorization (NTF).
Learning Energy Decompositions for Partial Inference in GFlowNets
Hyosoon Jang (POSTECH), Sungsoo Ahn (KAIST)
GenerationOptimizationReinforcement LearningFlow-based Model
🎯 What it does: This paper proposes a method for partial inference using a generative flow network through learned energy decomposition (LED-GFN), which assigns learnable potential energy as a local credit signal for each state transition, thereby enhancing sample diversity and quality.
Learning Energy-Based Models by Cooperative Diffusion Recovery Likelihood
Yaxuan Zhu (University of California Los Angeles), Ruiqi Gao (Google DeepMind)
RestorationGenerationAnomaly DetectionDiffusion modelImage
🎯 What it does: A Cooperative Diffusion Recovery Likelihood (CDRL) framework is designed, which jointly trains an Energy-Based Model (EBM) and an MCMC initializer, allowing each noise layer in the diffusion process to generate high-quality samples with only a few sampling steps, and further supports downstream tasks such as conditional generation, compositional generation, image inpainting, and anomaly detection.
Learning Flexible Body Collision Dynamics with Hierarchical Contact Mesh Transformer
Youn-Yeol Yu (Yonsei University), Noseong Park (KAIST)
Graph Neural NetworkTransformerMeshGraph
🎯 What it does: Proposes the Hierarchical Contact Mesh Transformer (HCMT) for learning the collision dynamics of flexible objects;
Learning from Aggregate responses: Instance Level versus Bag Level Loss Functions
Adel Javanmard (Google Research), Gang Fu (Google Research)
Federated LearningSafty and PrivacyTabular
🎯 What it does: This paper studies the layer loss and instance loss under the framework of aggregated response learning, proposing to view the instance layer as a regularized layer loss, and designs an interpolation estimator along with an ε-label differential privacy mechanism for aggregated responses.
Learning from Label Proportions: Bootstrapping Supervised Learners via Belief Propagation
Shreyas Havaldar (Google Research India), Aravindan Raghuveer (Google India)
ClassificationOptimizationRepresentation LearningContrastive LearningImageTabular
🎯 What it does: In the problem of Learning from Label Proportions (LLP), pseudo-labels are obtained using Bayesian propagation, and these pseudo-labels are then used to train and iteratively optimize embeddings to enhance instance-level predictions.
Learning From Simplicial Data Based on Random Walks and 1D Convolutions
Florian Frantzen (RWTH Aachen University), Michael T Schaub
Convolutional Neural NetworkGraph
🎯 What it does: This paper proposes a new simplex network structure called SCRaWl, based on random walks and 1D convolution, to handle high-order data on simplex complexes.
Learning from Sparse Offline Datasets via Conservative Density Estimation
Zhepeng Cen (Carnegie Mellon University), Ding Zhao (Columbia University)
Reinforcement LearningBenchmark
🎯 What it does: A novel offline reinforcement learning algorithm named Conservative Density Estimation (CDE) is proposed, which utilizes a conservative estimate of the state-action occupancy distribution to reduce offline data sparsity and out-of-distribution (OOD) extrapolation errors.
Learning Grounded Action Abstractions from Language
Lionel Wong (Massachusetts Institute of Technology), Jacob Andreas (Massachusetts Institute of Technology)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes the Ada framework, which utilizes language models to automatically generate and validate high-level planning abstractions (in PDDL format) and low-level controllers, gradually building a reusable action library to achieve language-driven long-term planning.
Learning Hierarchical Image Segmentation For Recognition and By Recognition
Tsung-Wei Ke (University of California), Stella X. Yu (University of Michigan)
RecognitionSegmentationTransformerContrastive LearningImage
🎯 What it does: The CAST model is proposed, which achieves structured understanding of images through internal hierarchical segmentation in image recognition tasks, completing both segmentation and recognition without the need for pixel-level annotations.
Learning Hierarchical Polynomials with Three-Layer Neural Networks
Zihao Wang (Peking University), Jason D. Lee (Princeton University)
Tabular
🎯 What it does: The paper proposes and proves a hierarchical polynomial learning method: by performing layered gradient descent on a three-layer neural network under the standard normal distribution, the learning objective function h = g · p (where g is a univariate polynomial and p is a multivariate polynomial) achieves a sample complexity of O(d^k).
Learning Hierarchical World Models with Adaptive Temporal Abstractions from Discrete Latent Dynamics
Christian Gumbsch (University of Tubingen), Martin V. Butz (University of Tubingen)
Robotic IntelligenceReinforcement LearningWorld ModelSequential
🎯 What it does: A self-supervised two-layer hierarchical world model THICK is designed, utilizing sparsely updatable context encoding to learn interpretable temporal abstractions, and applying it to model-based reinforcement learning and model predictive control.
Learning Implicit Representation for Reconstructing Articulated Objects
Hao Zhang (University of Illinois), Narendra Ahuja (University of Illinois)
Pose EstimationOptimizationRepresentation LearningOptical FlowVideo
🎯 What it does: This paper proposes an unsupervised learning framework for implicit skeleton and elastic parameters 3D reconstruction from monocular videos, modeling moving objects using skeletons, skin weights, stiffness coefficients, and time-varying transformations.
Learning in reverse causal strategic environments with ramifications on two sided markets
Seamus Somerstep (University of Michigan), Yaacov Ritov
Tabular
🎯 What it does: Proposes a counterfactual strategy learning framework to study employers' predictions and learning of workers' counterfactual strategies (i.e., workers directly manipulating labels leading to changes in features) in the labor market, and applies it to the Coate-Loury model to compare the effects of optimal (performative) and stable (reactive) hiring strategies.
Learning Interactive Real-World Simulators
Sherry Yang (University of California Berkeley), Pieter Abbeel (University of California Berkeley)
GenerationRobotic IntelligenceTransformerReinforcement LearningVision Language ModelDiffusion modelImageVideoTextMultimodality
🎯 What it does: A general interactive real-world simulator called UniSim is proposed, which can accept multimodal actions (language, robot control, camera movement, etc.) and generate corresponding video sequences; through this simulator, high-level visual language strategies, low-level reinforcement learning control strategies, and video caption models are trained to achieve zero-shot deployment in real environments.
Learning interpretable control inputs and dynamics underlying animal locomotion
Thomas Soares Mullen (Champalimaud Foundation), Adrien Jouary (Champalimaud Foundation)
Explainability and InterpretabilityRobotic IntelligenceRecurrent Neural NetworkTime SeriesSequential
🎯 What it does: Using the iLQR-VAE framework to learn interpretable sparse control inputs and latent dynamics, modeling the complete movement spectrum of zebrafish larvae and Caenorhabditis elegans, and obtaining an interpretable linear dynamic model through balanced model simplification.
Learning invariant representations of time-homogeneous stochastic dynamical systems
Vladimir R Kostic, Massimiliano Pontil (Istituto Italiano di Tecnologia)
OptimizationRepresentation LearningTime SeriesSequential
🎯 What it does: This paper proposes DPNets (Deep Projection Networks), which obtain a low-dimensional representation of dynamic systems that is reversible and free from metric distortion by optimizing the projection operator, thereby directly capturing the main invariant subspace of the transfer operator or generator.
Learning Large DAGs is Harder than you Think: Many Losses are Minimal for the Wrong DAG
Jonas Seng (Technical University of Darmstadt), Kristian Kersting (Technical University of Darmstadt)
Graph Neural NetworkGraphBiomedical Data
🎯 What it does: This study investigates the impact of measurement scale on structure learning algorithms (particularly those based on least squares error and log-likelihood loss) and demonstrates that in high-dimensional chain, fork, and collision structures, the error is scale-dependent, leading to incorrect DAGs being misidentified as optimal.
Learning Mean Field Games on Sparse Graphs: A Hybrid Graphex Approach
Christian Fabian (Technische Universitat Darmstadt), Heinz Koeppl (Technische Universitat Darmstadt)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: A large-scale multi-agent equilibrium learning framework for sparse graph networks is proposed and implemented - the Gravitational Graphical Equilibrium Game (GXMFG), along with a scalable learning algorithm.
Learning model uncertainty as variance-minimizing instance weights
Nishant Jain (Google Research India), Pradeep Shenoy (Google Research India)
Domain AdaptationOptimizationMeta LearningConvolutional Neural NetworkReinforcement LearningImageBiomedical Data
🎯 What it does: Learn an instance conditional weight function (U-SCORE) that associates training instance weights with prediction uncertainty through bi-level optimization, allowing for the assessment of model uncertainty during both training and testing.
Learning Multi-Agent Communication from Graph Modeling Perspective
Shengchao Hu (Shanghai Jiao Tong University), Dacheng Tao (Nanyang Technological University)
OptimizationGraph Neural NetworkTransformerReinforcement LearningAgentic AIGraph
🎯 What it does: Proposes CommFormer, which learns variable sparse communication graphs for multi-agent collaboration.
Learning Multi-Agent Communication with Contrastive Learning
Yat Long Lo (Dyson), Michael Noukhovitch (Mila)
Recurrent Neural NetworkReinforcement LearningContrastive LearningSequential
🎯 What it does: The study focuses on decentralized multi-agent reinforcement learning, learning a common protocol through contrastive learning of communication information to enhance cooperation effectiveness.
Learning Multi-Faceted Prototypical User Interests
Nhu-Thuat Tran (Singapore Management University), Hady W. Lauw (Singapore Management University)
Recommendation SystemAuto EncoderTabular
🎯 What it does: The FACETVAE model is proposed, which utilizes the VAE framework to construct low-level interests based on multi-faceted (multi-dimensional) prototype learning and synthesizes high-level user interests through a binding mechanism, thereby achieving multi-faceted and interpretable modeling of user preferences.
Learning Nash Equilibria in Rank-1 Games
Nikolas Patris (University of California), Ioannis Panageas (University of California)
OptimizationReinforcement Learning
🎯 What it does: A decentralized learning algorithm is proposed, utilizing parameterized zero-sum games and optimized structures to find approximate Nash equilibria in rank-1 bimatrix games.
Learning No-Regret Sparse Generalized Linear Models with Varying Observation(s)
Diyang Li (Cornell University), Bin Gu (Mohamed bin Zayed University of Artificial Intelligence)
OptimizationTabularFinance RelatedOrdinary Differential Equation
🎯 What it does: A novel online learning framework and algorithm SAGO is proposed for sparse generalized linear models (GLM), which can achieve no-regret updates when the observed data is dynamically added or removed, and automatically adjusts the regularization parameters in each iteration.
Learning Optimal Contracts: How to Exploit Small Action Spaces
Francesco Bacchiocchi (Politecnico di Milano), Nicola Gatti (Politecnico di Milano)
Optimization
🎯 What it does: This paper studies the principal-agent problem of hidden actions and proposes a result-based payment contract that learns the optimal or approximately optimal contract through observing outcomes in multi-round interactions.
Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks
Yanqiao Zhu (University of California Los Angeles), Wei Wang
Drug DiscoveryGraph Neural NetworkGraphBenchmark
🎯 What it does: A MARCEL benchmark is proposed to evaluate the learning effectiveness on a set of molecular conjugates.
Learning Performance-Improving Code Edits
Alexander G Shypula, Amir Yazdanbakhsh (Google DeepMind)
OptimizationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Utilizing large pre-trained code language models (LLMs) for high-level program performance optimization, primarily by generating code rewrites that significantly enhance execution speed;
Learning Personalized Causally Invariant Representations for Heterogeneous Federated Clients
Xueyang Tang (Hong Kong Polytechnic University), Jingcai Guo (Hong Kong Polytechnic University)
Federated LearningImage
🎯 What it does: In the heterogeneous federated learning scenario, the FedSDR method is proposed, which first collaboratively discovers all shortcut features and removes them, and then generates the best personalized causal invariant predictors for each client.
Learning Planning Abstractions from Language
Weiyu Liu (Stanford University), Jiajun Wu (Stanford University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextPoint Cloud
🎯 What it does: Utilize language-annotated demonstration learning to abstract states and actions in sequential decision-making, and implement hierarchical control from high-level planning to low-level execution.
Learning Polynomial Problems with $SL(2, \mathbb{R})$-Equivariance
Hannah Lawrence (Massachusetts Institute of Technology), Mitchell Tong Harris (Massachusetts Institute of Technology)
OptimizationTabular
🎯 What it does: Applying machine learning to polynomial extremum and positivity determination, constructing and training neural networks for fast solving.
Learning Robust Generalizable Radiance Field with Visibility and Feature Augmented Point Representation
WANG Jiaxu, Renjing Xu (Hong Kong University of Science and Technology)
Data SynthesisDepth EstimationNeural Radiance FieldPoint Cloud
🎯 What it does: A generalizable NeRF framework based on point clouds, GPF, is proposed, achieving high-quality view synthesis through visibility-guided feature extraction, robust logarithmic sampling, and feature-enhanced learnable kernels.
Learning Semantic Proxies from Visual Prompts for Parameter-Efficient Fine-Tuning in Deep Metric Learning
Li Ren (University of Central Florida), Kien A. Hua
RetrievalTransformerPrompt EngineeringImage
🎯 What it does: The study achieves parameter-efficient deep metric learning fine-tuning on a pre-trained ViT through Visual Prompt Tuning and semantic proxies.
Learning semilinear neural operators: A unified recursive framework for prediction and data assimilation.
Ashutosh Singh (Northeastern University), Tales Imbiriba (Northeastern University)
Recurrent Neural NetworkTime SeriesPhysics Related
🎯 What it does: A recursive neural operator framework NODA is proposed, capable of long-term prediction of semi-linear PDEs in the absence of complete measurements and performing data assimilation in the presence of noisy measurements.
Learning Stackable and Skippable LEGO Bricks for Efficient, Reconfigurable, and Variable-Resolution Diffusion Modeling
Huangjie Zheng (University of Texas at Austin), Mingyuan Zhou (University of Texas at Austin)
GenerationSuper ResolutionComputational EfficiencyTransformerDiffusion modelImage
🎯 What it does: Proposed stackable and skippable LEGO blocks for constructing a diffusion model backbone that can be flexibly reconfigured during both training and sampling phases.
Learning the greatest common divisor: explaining transformer predictions
Francois Charton
Explainability and InterpretabilityTransformerSupervised Fine-TuningTabular
🎯 What it does: Train a small Transformer to learn the greatest common divisor (GCD) of two positive integers, revealing its internal interpretable algorithm through input-output analysis.
Learning Thresholds with Latent Values and Censored Feedback
Jiahao Zhang (Peking University), Xiaotie Deng (Peking University)
🎯 What it does: This paper studies the problem of active learning thresholds in latent space, exploring the relationship between the unknown reward function, the threshold, and the latent value, and proposes a general active learning abstraction.
Learning to Act from Actionless Videos through Dense Correspondences
Po-Chen Ko (National Taiwan University), Joshua B. Tenenbaum (Massachusetts Institute of Technology)
Robotic IntelligenceDiffusion modelOptical FlowVideo
🎯 What it does: Construct a video-based robotic strategy that can learn and perform diverse tasks solely from action-free labeled RGB videos.
Learning to Act without Actions
Dominik Schmidt (Weco AI), Minqi Jiang (Meta AI)
Robotic IntelligenceReinforcement LearningVideo
🎯 What it does: The LAPO method is proposed, which infers the potential action space and trains the corresponding policy using only video data without action labels through inverse and forward dynamics models.
Learning to Compose: Improving Object Centric Learning by Injecting Compositionality
Whie Jung (Korea Advanced Institute of Science and Technology), Seunghoon Hong (Korea Advanced Institute of Science and Technology)
SegmentationGenerationData-Centric LearningTransformerDiffusion modelAuto EncoderImage
🎯 What it does: This paper proposes the incorporation of a combinatorial regularization objective into the object-centric learning framework, allowing slot representations to generate valid images through mixing, thereby explicitly enhancing combinability.
Learning to design protein-protein interactions with enhanced generalization
Anton Bushuiev (Czech Technical University), Josef Sivic (Czech Technical University)
Drug DiscoveryProtein Structure PredictionTransformerBiomedical Data
🎯 What it does: A large-scale non-redundant protein-protein interaction (PPI) dataset, PPIRef, was constructed, and the SE(3) equivariant Transformer model, PPIFORMER, was pre-trained using this dataset. It was fine-tuned on labeled data such as SKEMPI to predict the effect of mutations on binding affinity's ΔΔG.
Learning to Embed Time Series Patches Independently
Seunghan Lee (Yonsei University), Kibok Lee (Yonsei University)
ClassificationRepresentation LearningContrastive LearningTime Series
🎯 What it does: A self-supervised pre-training method for time series called PITS is proposed, which achieves time series representation learning using a patch reconstruction task and an MLP encoder.
Learning to Jointly Understand Visual and Tactile Signals
Yichen Li (Massachusetts Institute of Technology), Wojciech Matusik (NVIDIA)
RetrievalAuto EncoderVideoMultimodality
🎯 What it does: A cross-modal dataset of hand force pressure and visual video was constructed, and a shared latent manifold was learned, enabling cross-modal prediction and retrieval from vision to touch and from touch to vision, revealing the implicit structures in different modalities.
Learning to Make Adherence-aware Advice
Guanting Chen (University of North Carolina at Chapel Hill), Hanzhao Wang (Imperial College London)
Recommendation SystemAutonomous DrivingReinforcement LearningSequential
🎯 What it does: A decision-making model considering human compliance and deferral suggestions is proposed, along with a learning algorithm designed to approximate the optimal suggestion strategy and timing.
Learning to Reject Meets Long-tail Learning
Harikrishna Narasimhan (Google Research), Sanjiv Kumar (Google Research)
ClassificationImage
🎯 What it does: A learning rejection (L2R) method is proposed for long-tail classification scenarios, studying the optimal rejection strategy under non-standard evaluation metrics such as balanced error and worst group error.
Learning to Reject with a Fixed Predictor: Application to Decontextualization
Christopher Mohri (Stanford University), Yutao Zhong (Courant Institute)
GenerationOptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: A rejection learning framework under a fixed predictor is proposed, and a new convex differentiable surrogate loss function is designed to achieve high precision and high coverage for rejection options in natural language processing generation tasks (taking decontextualization as an example).
Learning to Relax: Setting Solver Parameters Across a Sequence of Linear System Instances
Mikhail Khodak (Carnegie Mellon University), Ameet Talwalkar (Carnegie Mellon University)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper studies how to automatically adjust the relaxation parameter ω of the SOR (Successive Over-Relaxation) solver through online learning when a series of linear equations need to be solved sequentially, thereby approaching the optimal number of iterations without additional matrix operations.
Learning to Solve Bilevel Programs with Binary Tender
Bo Zhou (University of Michigan), Siqian Shen (University of Michigan)
OptimizationGraph Neural NetworkSupervised Fine-Tuning
🎯 What it does: A learning framework based on neural networks is proposed to solve bi-level programming (BP) with binary linked variables, by approximating the lower-level value function and embedding it into a single-level integer programming for solution.
Learning to solve Class-Constrained Bin Packing Problems via Encoder-Decoder Model
Hanni Cheng (Hikvision Research Institute), Shiliang Pu (Hikvision Research Institute)
OptimizationGraph Neural NetworkReinforcement LearningTabular
🎯 What it does: A learning framework based on Encoder-Decoder is proposed to solve the Class-Constrained Bin Packing Problem (CCBPP) and its application in manufacturing order consolidation (OCP).
Learning with a Mole: Transferable latent spatial representations for navigation without reconstruction
Guillaume Bono (Naver Labs Europe), Christian Wolf (Naver Labs Europe)
Robotic IntelligenceRecurrent Neural NetworkReinforcement LearningAgentic AISequential
🎯 What it does: Learn a transferable latent space representation (Navigability) by allowing a blind auxiliary agent to navigate using this representation in short sub-tasks, thereby constructing an actionable map-like representation for 3D navigation tasks without explicitly reconstructing the scene.
Learning with Language-Guided State Abstractions
Andi Peng (Massachusetts Institute of Technology), Julie Shah
Robotic IntelligenceReinforcement Learning from Human FeedbackConvolutional Neural NetworkLarge Language ModelReinforcement LearningImageMultimodality
🎯 What it does: This paper proposes a framework (LGA) that automatically constructs task-specific state abstractions using natural language and large language models, and performs imitation learning on this abstraction to obtain sample-efficient and robust control strategies.
Learning with Mixture of Prototypes for Out-of-Distribution Detection
Haodong Lu (University of New South Wales), Kristen Moore (CSIRO)
Anomaly DetectionConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes a distance-based OOD detection method called PALM, which models each category as multiple von Mises-Fisher prototypes to learn a more compact and separable embedding space, achieving precise detection of out-of-distribution samples from the training set.
Leave-one-out Distinguishability in Machine Learning
Jiayuan Ye (National University of Singapore), Reza Shokri (Imperial College London)
Safty and PrivacyComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A leave-one-out discrepancy (LOOD) framework based on Gaussian processes is proposed and validated to accurately assess the changes in output distribution of machine learning models when a single data point is added or removed from the training set, thereby measuring memorization, information leakage, and the influence of data points.
Leftover Lunch: Advantage-based Offline Reinforcement Learning for Language Models
Ashutosh Baheti (Georgia Institute of Technology), Mark Riedl (Georgia Institute of Technology)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Designed and implemented an offline policy gradient method A-LOL, which utilizes a single action assumption for RLHF on language models, thereby avoiding online sampling and high computational costs.
LEGO-Prover: Neural Theorem Proving with Growing Libraries
Haiming Wang (Sun Yat-sen University), Xiaodan Liang (Sun Yat-sen University)
Large Language ModelText
🎯 What it does: A neural theorem proving framework named LEGO-Prover is proposed, which utilizes a growable skill library (verified lemmas/theorems) to construct proofs in a modular, block-like manner, retrieving existing skills while also generating and evolving new skills during the proving process.
LEMON: Lossless model expansion
Yite Wang (University of Illinois Urbana-Champaign), Hongxia Yang (ByteDance)
TransformerImageText
🎯 What it does: The LEMON method is proposed, which utilizes the pre-trained weights of small models for lossless model expansion and quickly converges with an optimized learning rate schedule.
Lemur: Harmonizing Natural Language and Code for Language Agents
Yiheng Xu (University of Hong Kong), Tao Yu (University of Hong Kong)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Two open-source large language models, Lemur and Lemur-Chat, have been developed with the goal of balancing capabilities in natural language and programming languages to support the construction of language agents.
Lemur: Integrating Large Language Models in Automated Program Verification
Haoze Wu (Stanford University), Nina Narodytska (VMware)
OptimizationAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: This paper proposes a framework named LEMUR, which combines large language models (LLMs) with automatic verifiers for automated program verification; it constructs an executable and provably correct program verification process through a formal reasoning calculator and proves its completeness.
Less is More: Fewer Interpretable Region via Submodular Subset Selection
Ruoyu Chen (Institute of Information Engineering, Chinese Academy of Sciences), Xiaochun Cao (Sun Yat-sen University)
Explainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: By re-framing the image interpretability problem as a submodular set selection problem, a submodular objective function based on four scores: confidence, effectiveness, consistency, and synergy is designed. A greedy algorithm is used to select a limited number of local sub-regions, resulting in more refined and accurate interpretable areas.
Less is More: One-shot Subgraph Reasoning on Large-scale Knowledge Graphs
Zhanke Zhou (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
Hyperparameter SearchGraph Neural NetworkGraph
🎯 What it does: A one-shot subgraph link prediction framework is proposed, which first uses Personalized PageRank to extract query-relevant subgraphs, and then employs a structured GNN model for prediction on these subgraphs.
Less or More From Teacher: Exploiting Trilateral Geometry For Knowledge Distillation
Chengming Hu (McGill University), Xue Liu (McGill University)
Anomaly DetectionKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: A method called TGeo-KD is proposed for adaptively learning the sample-level knowledge fusion ratio during the knowledge distillation process.
Let 2D Diffusion Model Know 3D-Consistency for Robust Text-to-3D Generation
Junyoung Seo (Korea University), Seungryong Kim (Korea University)
GenerationData SynthesisDiffusion modelScore-based ModelPoint Cloud
🎯 What it does: By combining a pre-trained two-dimensional diffusion model with sparse depth projection, a 3DFuse framework is constructed to enhance the geometric consistency and detail quality of text-to-3D generation based on score distillation.
Let Models Speak Ciphers: Multiagent Debate through Embeddings
Chau Pham (Boston University), Hongxia Yang (ByteDance Inc.)
GenerationOptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The CIPHER protocol is proposed, using embeddings output by the Transformer instead of tokens for LLM multi-agent debates, eliminating information loss caused by token sampling.
Let's do the time-warp-attend: Learning topological invariants of dynamical systems
Noa Moriel (Hebrew University), Mor Nitzan (Hebrew University)
ClassificationConvolutional Neural NetworkBiomedical DataPhysics Related
🎯 What it does: Using data augmentation and self-attention convolutional networks, we learn the topological invariant features of supercritical Hopf bifurcation and construct the Time-Warp-Attend method for cross-system identification of point attractors and limit cycles, as well as for locating bifurcation boundaries.
Let's Verify Step by Step
Hunter Lightman (OpenAI), Karl Cobbe (OpenAI)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Train a reward model through process supervision (labeling each step of chain reasoning) to enhance the reliability of mathematical reasoning tasks;
Leveraging augmented-Lagrangian techniques for differentiating over infeasible quadratic programs in machine learning
Antoine Bambade (Inria), Justin Carpentier (Inria)
OptimizationTabular
🎯 What it does: A unified augmented Lagrangian technique and Extended Conservative Jacobian (ECJ) method are proposed to differentiate feasible and infeasible convex quadratic programming layers, and an open-source QPLayer library has been implemented.
Leveraging Generative Models for Unsupervised Alignment of Neural Time Series Data
Ayesha Vermani (Champalimaud Foundation), Josue Nassar (RyvivyR)
Data SynthesisOptimizationRecurrent Neural NetworkAuto EncoderTime SeriesSequential
🎯 What it does: An unsupervised time series data alignment method is proposed, allowing the reuse of a pre-trained seqVAE model without source data or paired samples.
Leveraging Hyperbolic Embeddings for Coarse-to-Fine Robot Design
Heng Dong (Tsinghua University), Chongjie Zhang (Washington University in St. Louis)
Robotic IntelligenceTransformerReinforcement LearningTabular
🎯 What it does: A coarse-to-fine multi-cell robot design method is proposed, utilizing hypercurvature embedding to unify the representation of robots with different granularities and searching for optimal structures through an improved cross-entropy method.
Leveraging Low-Rank and Sparse Recurrent Connectivity for Robust Closed-Loop Control
Neehal Tumma (Harvard University), Daniela Rus (Massachusetts Institute of Technology)
Recurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: Proposed low-rank and sparse variable cyclic connection parameterization, and validated its robustness in closed-loop control tasks.
Leveraging Optimization for Adaptive Attacks on Image Watermarks
Nils Lukas (University of Waterloo), Florian Kerschbaum (University of Waterloo)
GenerationOptimizationAdversarial AttackDiffusion modelImage
🎯 What it does: This paper proposes an adaptive learnable attack method for image watermarking, utilizing optimization techniques to achieve watermark removal while maintaining image quality.
Leveraging Uncertainty Estimates To Improve Classifier Performance
Gundeep Arora (Amazon), Rajeev Rastogi (Amazon)
ClassificationOptimizationTabular
🎯 What it does: The study incorporates model prediction uncertainty estimation into binary classification tasks to improve decision thresholds.
Leveraging Unpaired Data for Vision-Language Generative Models via Cycle Consistency
Tianhong Li (Massachusetts Institute of Technology), Dilip Krishnan (Google Research)
GenerationData SynthesisTransformerVision Language ModelImageText
🎯 What it does: Proposes the ITIT training framework, which utilizes cycle consistency to achieve audiovisual language generation on unpaired image and text data.
Lewis's Signaling Game as beta-VAE For Natural Word Lengths and Segments
Ryo Ueda (University of Tokyo), Tadahiro Taniguchi (Kyoto University)
GenerationRecurrent Neural NetworkReinforcement LearningText
🎯 What it does: View the Lewis signaling game as a β-VAE, redefining the objective function using ELBO, and studying whether the generated language meets the statistical laws of word length and segmentation in natural language from the perspective of generative models.
LiDAR-PTQ: Post-Training Quantization for Point Cloud 3D Object Detection
Sifan Zhou (Southeast University), Xiangxiang Chu (Meituan Inc)
Object DetectionAutonomous DrivingPoint Cloud
🎯 What it does: This paper proposes a post-training quantization method for 3D LiDAR object detection called LiDAR-PTQ.
LiDAR: Sensing Linear Probing Performance in Joint Embedding SSL Architectures
Vimal Thilak (Apple), Etai Littwin (Apple)
Representation LearningConvolutional Neural NetworkTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: A new unlabeled evaluation metric called LiDAR is proposed to measure the representation quality of joint embedding self-supervised learning models, replacing traditional methods that require downstream task evaluations.
Lie Group Decompositions for Equivariant Neural Networks
Mircea Mironenco (University of Amsterdam), Patrick Forré (University of Amsterdam)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: A global parameterization and integrability framework for constructing compatible convolution operations on non-compact, non-Abelian Lie groups (such as GL⁺(n,ℝ) and SL(n,ℝ)) is established, and a equivariant convolution layer based on this framework is implemented.
Lifting Architectural Constraints of Injective Flows
Peter Sorrenson (Heidelberg University), Ullrich Koethe
GenerationData SynthesisOptimizationConvolutional Neural NetworkFlow-based ModelAuto EncoderImageTabular
🎯 What it does: This paper proposes Free-Form Injective Flow (FIF), which addresses the dimensionality bottleneck and reversibility limitations of traditional invertible flows in low-dimensional manifold learning by jointly learning the data manifold and the distribution on that manifold.
Light Schrödinger Bridge
Alexander Korotin (Artificial Intelligence Research Institute), Evgeny Burnaev (Artificial Intelligence Research Institute)
Data SynthesisOptimizationComputational EfficiencyBiomedical DataStochastic Differential Equation
🎯 What it does: A new lightweight Schrödinger bridge solver is proposed, aimed at addressing the issues of complexity and high computational resource consumption in existing solvers.
Light-MILPopt: Solving Large-scale Mixed Integer Linear Programs with Lightweight Optimizer and Small-scale Training Dataset
Huigen Ye (Tsinghua University), Hongyan Wang (Tsinghua University)
OptimizationGraph Neural NetworkTabular
🎯 What it does: This paper presents Light-MILPopt, a lightweight optimization framework for large-scale Mixed Integer Linear Programming (MILP), which includes four stages: problem decomposition, model-based initial solution prediction, variable and constraint dimensionality reduction, and data-driven optimization.
LightHGNN: Distilling Hypergraph Neural Networks into MLPs for 100x Faster Inference
Yifan Feng (Tsinghua University), Yue Gao (Tsinghua University)
Computational EfficiencyKnowledge DistillationGraph Neural NetworkGraph
🎯 What it does: This paper proposes two models, LightHGNN and LightHGNN+, which distill the knowledge of higher-order graph neural networks (HGNN) into MLPs, eliminating the dependence on hypergraph structures during the inference phase, thus achieving high-speed and low-memory inference.
Like Oil and Water: Group Robustness Methods and Poisoning Defenses May Be at Odds
Michael-Andrei Panaitescu-Liess (University of Maryland), Tudor Dumitras (University of Maryland)
ClassificationFederated LearningAdversarial AttackImage
🎯 What it does: This paper studies the fundamental conflict between group robustness methods and data poisoning attacks in machine learning training without group labels, demonstrating that these methods cannot distinguish between legitimate minority samples and poisoned samples, making the model more vulnerable to attacks.
Likelihood Training of Cascaded Diffusion Models via Hierarchical Volume-preserving Maps
Henry Li (Yale University), Yuval Kluger (Yale University)
GenerationData SynthesisCompressionDiffusion modelImage
🎯 What it does: A method is proposed that utilizes hierarchical volume-preserving mappings (such as Laplacian pyramids and wavelet transforms) to train cascade diffusion models, enabling precise likelihood training while maintaining high-resolution generation quality.
LILO: Learning Interpretable Libraries by Compressing and Documenting Code
Gabriel Grand (Massachusetts Institute of Technology), Jacob Andreas (Massachusetts Institute of Technology)
OptimizationExplainability and InterpretabilityAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: LILO enhances cross-task code reuse and readability by cyclically learning interpretable function libraries through the combination of large language model-driven program synthesis, STITCH's symbolic compression, and AutoDoc's automatic documentation.
Linear attention is (maybe) all you need (to understand Transformer optimization)
Kwangjun Ahn (Massachusetts Institute of Technology), Suvrit Sra (Massachusetts Institute of Technology)
OptimizationTransformerTabular
🎯 What it does: This study investigates training shallow linear Transformer models on random linear regression data and demonstrates that the model can reproduce phenomena observed in full Transformer training, such as Adam > SGD, noise heavy tails, landscape pathology, and directional smoothness.
Linear Log-Normal Attention with Unbiased Concentration
Yury Nahshan (Huawei Technologies), Emir Haleva (Huawei Technologies)
TransformerSupervised Fine-TuningText
🎯 What it does: A linearized self-attention mechanism called LLN Attention is proposed and implemented, which can simulate the distribution and concentration of the original Softmax attention, achieving linear time and memory.
Linearity of Relation Decoding in Transformer Language Models
Evan Hernandez (Massachusetts Institute of Technology), David Bau (Northeastern University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study investigates the linear approximability of relation decoding in Transformer language models and proposes a Linear Relation Embedding (LRE) framework.
Lion Secretly Solves a Constrained Optimization: As Lyapunov Predicts
Lizhang Chen (University of Texas at Austin), qiang liu
OptimizationImageTextOrdinary Differential Equation
🎯 What it does: This paper explains the essence of the Lion optimizer through theoretical analysis, proving its equivalence to minimizing the objective function under weighted decay constraints, and proposes a general family of algorithms called LionK;
Lipschitz Singularities in Diffusion Models
Zhantao Yang (Shanghai Jiao Tong University), Fan Cheng (Shanghai Jiao Tong University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: The study investigates the infinite Lipschitz singularity of diffusion models as time approaches 0, and proposes E-TSDM to eliminate this singularity through shared time conditions.
LipSim: A Provably Robust Perceptual Similarity Metric
Sara Ghazanfari (New York University), Siddharth Garg (New York University)
Knowledge DistillationRepresentation LearningAdversarial AttackTransformerContrastive LearningImage
🎯 What it does: This study investigates and proves the vulnerability of existing ViT-based perceptual similarity metrics (such as DreamSim) and proposes LipSim—a perceptual similarity metric that achieves provable robustness through 1-Lipschitz networks and projections.
Lipsum-FT: Robust Fine-Tuning of Zero-Shot Models Using Random Text Guidance
Giung Nam (KAIST AI), Juho Lee (KAIST AI)
Domain AdaptationTransformerSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: This paper proposes Lipsum-FT, a method for robust fine-tuning on visual-language models, and conducts experiments in two distribution shift scenarios, DomainNet and ImageNet, demonstrating its ability to improve the accuracy and uncertainty estimation of distribution shift data while maintaining or enhancing the original zero-shot performance.
LipVoicer: Generating Speech from Silent Videos Guided by Lip Reading
Yochai Yemini (Bar Ilan University), Ethan Fetaya (Bar Ilan University)
GenerationDiffusion modelVideoAudio
🎯 What it does: The LipVoicer framework is designed to generate high-quality, synchronized speech using silent videos and text predicted by a lip-reading model.
Listen, Think, and Understand
Yuan Gong (Massachusetts Institute of Technology), James R. Glass (Massachusetts Institute of Technology)
TransformerLarge Language ModelPrompt EngineeringMultimodalityAudio
🎯 What it does: A multi-modal large language model LTU has been developed, capable of completing audio question-answering tasks from closed to open format through audio perception, reasoning, and understanding.
LitCab: Lightweight Language Model Calibration over Short- and Long-form Responses
Xin Liu (University of Michigan), Lu Wang (University of Michigan)
GenerationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: A lightweight language model calibration method called LITCAB is proposed, which uses a linear layer to predict logit biases from the last hidden representation of the LM, enhancing the probability calibration of generated outputs.
LLaMA-Adapter: Efficient Fine-tuning of Large Language Models with Zero-initialized Attention
Renrui Zhang (Shanghai Artificial Intelligence Laboratory), Peng Gao (Shanghai Artificial Intelligence Laboratory)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelTextMultimodality
🎯 What it does: A lightweight adapter (LLaMA-Adapter) was developed on top of LLaMA 7B, achieving efficient fine-tuning for instruction following by inserting learnable prompts and zero-initialized attention mechanisms on top of a frozen model, training only 1.2M parameters in less than an hour; the method was also extended to multimodal applications, utilizing the CLIP image encoder to generate visual prompts and construct a multimodal LLM.
LLCP: Learning Latent Causal Processes for Reasoning-based Video Question Answer
Guangyi Chen (Carnegie Mellon University), Kun Zhang (Boston College)
Object TrackingAnomaly DetectionAutonomous DrivingRecurrent Neural NetworkAuto EncoderContrastive LearningVideoMultimodality
🎯 What it does: Proposes the LLCP framework, which utilizes a self-supervised temporal multivariate generative model to learn the underlying causal processes in videos, thereby achieving accident attribution and counterfactual prediction without relying on question-answer annotations.
Llemma: An Open Language Model for Mathematics
Zhangir Azerbayev (Princeton University), Sean Welleck (Carnegie Mellon University)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: An open-source large language model specifically designed for mathematics, LLEMMA, has been proposed and the 55B-token Proof Pile 2 dataset has been publicly released.
LLM Augmented LLMs: Expanding Capabilities through Composition
Rachit Bansal (Google Research India), Partha Talukdar (Google Research India)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The CALM framework is proposed, which achieves a combination of capabilities between a frozen Anchor LLM and a specialized Augmenting model by inserting cross-attention layers, supporting the completion of new tasks without modifying the original model weights.
LLM Blueprint: Enabling Text-to-Image Generation with Complex and Detailed Prompts
Hanan Gani (Mohamed Bin Zayed University of AI), Peter Wonka
GenerationTransformerLarge Language ModelDiffusion modelImageText
🎯 What it does: This paper proposes a two-stage text-to-image generation framework: first, a large language model (LLM) is used to extract a structured scene blueprint from long text prompts (including object layout, object descriptions, and background context), and an initial image is generated using a diffusion model conditioned on the layout; subsequently, through box-level iterative refinement, a CLIP multimodal-guided diffusion and image synthesis model is employed to gradually correct each object to ensure it fully matches the description.