ICML 2026 Papers — Page 19
International Conference on Machine Learning · 6554 papers
Enhancing LLM Training via Spectral Clipping
Xiaowen Jiang (CISPA Helmholtz Center for Information Security), Sebastian U Stich (CISPA Helmholtz Center for Information Security)
OptimizationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the SPECTRA framework, which achieves strict control over the spectral norm by performing spectral clipping on the weight update matrix during LLM training, with optional pre-spectral clipping on the original gradient, thereby improving training stability and final generalization performance.
Enhancing LLMs for Graph Tasks via Graph-aware LoRA Generation
Junshu Sun (Chinese Academy of Sciences), Shuhui Wang (Chinese Academy of Sciences)
ClassificationRepresentation LearningDrug DiscoveryGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraph
🎯 What it does: This paper proposes a graph information injection method at the weight level, i.e., after extracting the global graph representation using a graph neural network (GNN), the weights updated by low-rank LoRA are directly injected into the hidden layers of a large language model (LLM). This approach injects the global structural information of the graph into the model without altering the model architecture, thereby enhancing the zero-shot transfer capability for graph tasks.
Enhancing Membership Inference Attacks on Diffusion Models from a Frequency-Domain Perspective
Puwei Lian (Southeast University), Bingkun BAO
Safty and PrivacyAdversarial AttackDiffusion modelImage
🎯 What it does: Studied the impact of high-frequency defects in the frequency domain of diffusion models on membership inference attacks, and proposed a general high-frequency filter to enhance the attack effectiveness.
Enhancing Multi-Modal LLMs Reasoning via Difficulty-Aware Group Normalization
Jinghan Li (University of Science and Technology of China), Xiangnan He (University of Science and Technology of China)
Computational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes a difficulty-aware grouping normalization method called Durian, which improves the training stability and inference performance of multi-modal LLMs under the RLVR framework.
Enhancing Neural Theorem Proving via High-Quality Proof Selection and Verifier Feedback
Xiaoxue Zhu (Zhejiang University), Yongwang Zhao (Zhejiang University)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes a neural theorem proving framework based on data centers, significantly improving model performance on Isabelle through high-quality data selection and validator feedback.
Enhancing Numerical Prediction in LLMs via Smooth MMD Alignment
Zhuo Zuo (Sichuan University), Xianggen Liu (Sichuan University)
ClassificationRecommendation SystemOptimizationData-Centric LearningTransformerLarge Language ModelMixture of ExpertsContrastive LearningTextTabularBenchmark
🎯 What it does: Proposed a novel numerical prediction training loss called SMMD, which constructs a distance kernel based on a numerical subword vocabulary and uses MMD to match distributions, while applying graph Laplacian smoothing regularization to the prediction-target residual to improve the accuracy of LLMs in numerical outputs.
Enhancing Protein-Protein Interaction Prediction with Hierarchical Motif-based Multimodal Protein Embedding
Zaifei YANG (Hong Kong University of Science and Technology), James Kwok (Hong Kong University of Science and Technology)
Drug DiscoveryProtein Structure PredictionGraph Neural NetworkTransformerSupervised Fine-TuningContrastive LearningMultimodalityGraphBiomedical Data
🎯 What it does: A hierarchical multi-modal protein encoder, MMM-PPI, was constructed to improve the prediction of protein-protein interactions (PPI).
Enhancing Reasoning for Diffusion LLMs via Distribution Matching Policy Optimization
Yuchen Zhu (Georgia Institute of Technology), Yongxin Chen (Georgia Institute of Technology)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningDiffusion modelContrastive LearningText
🎯 What it does: Proposes the Distribution Matching Policy Optimization (DMPO) framework for fine-tuning diffusion large language models (dLLM) using reinforcement learning, aiming to align the model's policy distribution with a reward-biased distribution;
Enhancing Train-Free Infinite-Frame Generation for Consistent Long Videos
Xiaokun Feng (University of Chinese Academy of Sciences), Kaiqi Huang (University of Chinese Academy of Sciences)
GenerationTransformerPrompt EngineeringDiffusion modelScore-based ModelVideoRetrieval-Augmented Generation
🎯 What it does: Proposes an untrained, infinite-frame video generation method called MIGA, which can generate long videos under fixed memory.
EnsembleVLA: Ensemble Learning for Vision-Language Action Models
Mingchen Song (Harbin Institute Of Technology), Weili Guan (Harbin Institute Of Technology)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerVision-Language-Action ModelDiffusion modelScore-based ModelFlow-based ModelMultimodalityOrdinary Differential Equation
🎯 What it does: Propose the EnsembleVLA framework, which integrates distributions from different VLA (Vision-Language-Action) models at multiple levels through an energy-based model, and further introduces adaptive weights and confidence gating for fine-grained action correction after integration.
Ensembling Sparse Autoencoders
Soham Gadgil (University of Washington), Su-In Lee (University of Washington)
Explainability and InterpretabilityRepresentation LearningTransformerMixture of ExpertsAuto EncoderContrastive LearningText
🎯 What it does: This paper studies integrating sparse autoencoders (SAE) to improve the reconstruction quality and interpretability of language model activations.
Entangled No More: Multi-Domain Decoupling for Robust Dynamic Graph Neural Networks
Youda Mo (South China Normal University), Quanlong Guan (Jinan University)
Anomaly DetectionGraph Neural NetworkTransformerContrastive LearningGraphTime Series
🎯 What it does: Proposes a multi-domain decoupling framework named DeR-Mamba, aiming to enhance the robustness of dynamic graph neural networks in environments with incomplete structures, noise perturbations, and temporal redundancies.
EnterpriseOps-Gym: Environments and Evaluations for Stateful Agentic Planning and Tool Use in Enterprise Settings
Shiva Krishna Reddy Malay (ServiceNow Research), Sai Rajeswar (ServiceNow Research)
Autonomous DrivingOptimizationFederated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningAdversarial AttackHyperparameter SearchData-Centric LearningRobotic IntelligenceMeta LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionTransformerLarge Language ModelPrompt EngineeringTextTabularSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the ENTERPRISEOPS-GYM benchmark for evaluating LLM agents that perform long-term, stateful planning and tool usage in enterprise environments;
EntRAG: Entity-Centric Retrieval-Augmented Generation for Knowledge-based Visual Question Answering
Yiheng Hu (University of New South Wales), Wenjie Zhang (University of New South Wales)
GenerationRetrievalRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose EntRAG, a retrieval-augmented generation framework centered on entities, for knowledge-driven visual question answering;
EntroKV: Entropy-Guided Dynamic Budget Allocation for KV-Cache Compression
Wenhao Gao (Beijing University of Posts and Telecommunications), Xiaojie Wang (Beijing University of Posts and Telecommunications)
CompressionOptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposed a dynamic budget allocation framework called EntroKV for KV cache compression, aiming to address the memory bottleneck problem in long-context large language model (LLM) services.
Entropic Mirror Monte Carlo
Anas Cherradi (Sorbonne Université, Université Paris Cité, CNRS, Laboratoire de Probabilités, Statistique et Modélisation, LPSM), Julien Stoehr (CEREMADE, Université ParisDauphine, Université PSL, CNRS)
OptimizationFederated LearningComputational EfficiencyRepresentation LearningScore-based ModelMultimodalityTabularTime SeriesStochastic Differential Equation
🎯 What it does: Proposes the Entropic Mirror Monte Carlo (EM2C) algorithm, combining Entropic Mirror Descent with Markov transitions to achieve adaptive importance sampling and expectation estimation on high-dimensional multi-modal target distributions.
Entropy-Aware Dynamic KV Cache Sparsification for Autoregressive Image Generation and Editing
Tong Tong (Beijing University of Posts and Telecommunications), Alex Jinpeng Wang (Central South University)
GenerationTransformerVision Language ModelDiffusion modelImageText
🎯 What it does: Proposes SparseAR, a training-agnostic entropy-aware dynamic KV cache sparsification method, aimed at improving the inference efficiency of autoregressive image generation and editing.
Entropy-Aware On-Policy Distillation of Language Models
Woogyeol Jin (KAIST AI), Kimin Lee (KAIST AI)
Knowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningText
🎯 What it does: This paper proposes a new adversarial knowledge distillation method called Entropy-Aware On-Policy Distillation (EOPD), which improves the distillation process of language models on self-generated trajectories by using forward KL in high-entropy positions and backward KL in low-entropy positions.
Entropy-aware Span-Constrained Optimal Transport for Robust Cross-Tokenizer Knowledge Distillation
Zhi-Ping Liu (Nanjing University), Xinghao Chen (Huawei)
Knowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the E-SCOT framework in cross-tokenizer knowledge distillation, viewing distillation as a sparse optimal transport problem, leveraging a vocabulary-agnostic baseline metric, span-anchored lexical alignment, and adaptive reweighting based on R'-Enn entropy, to achieve more reliable alignment and information transfer between teacher and student.
Entropy-informed Decoding: Adaptive Information-Driven Branching
Benjamin Patrick Evans (JP Morgan AI Research), Leo Ardon (JP Morgan AI Research)
GenerationOptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes a decoding method called EDEN based on entropy information adaptive branching, allowing LLMs to dynamically allocate search width during generation according to the output distribution entropy.
Envisioning Beyond the Few: Disentangled Semantics and Primitives for Few-Shot Atypical Layout-to-Image Generation
Nan Bao (Beihang University), Jia Li (Beihang University)
GenerationAnomaly DetectionConvolutional Neural NetworkTransformerPrompt EngineeringDiffusion modelContrastive LearningImage
🎯 What it does: Propose a representation-driven framework based on semantic-primitive decoupling for few-shot abnormal layout to image generation
Envy-Free Allocation of Indivisible Goods via Noisy Queries
Zihan Li (Meta), Warut Suksompong (National University Of Singapore)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyReinforcement LearningContrastive LearningGaussian SplattingStochastic Differential Equation
🎯 What it does: This paper studies the problem of fairly distributing indivisible items between two agents, where the agents' value functions can only be obtained through noisy queries. We provide theoretical upper and lower bounds on the number of queries required to find an envy-free allocation under Gaussian noise, additive values, and value ranges [0, 1], and prove that the optimal scale is ~m^{2.5}/Δ^{2} (Δ is the lower bound of the optimal negative envy).
ePC: Fast and Deep Predictive Coding in Digital Simulation
Cédric Goemaere (IDLab, Ghent University - imec), Thomas Demeester (IDLab, Ghent University - imec)
ClassificationOptimizationComputational EfficiencyConvolutional Neural NetworkTransformerAuto EncoderImage
🎯 What it does: Reformulate the traditional state-based predictive coding (sPC) and propose error-based predictive coding (ePC) to eliminate the problem of exponential signal decay in digital simulations and achieve fast convergence in deep networks.
EPiC: Efficient Video Camera Control Learning with Precise Anchor-Video Guidance
Zun Wang (University of North Carolina), Mohit Bansal (University of North Carolina)
GenerationComputational EfficiencyRobotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelContrastive LearningOptical FlowVideoPoint Cloud
🎯 What it does: Construct aligned anchor videos based on the visibility mask of the first frame, and implement efficient and accurate video camera control learning by adding a lightweight Anchor‑ControlNet on a frozen main model.
EpiCache: Episodic KV Cache Management for Long-Term Conversation on Resource-Constrained Environments
Minsoo Kim (Apple), Minsik Cho (Apple)
CompressionComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose a training-agnostic KV cache management framework called EPICACHE, which achieves efficient inference for long conversations in resource-constrained environments through block-level prefilling and episodic clustering.
EpiCoCo: De Novo Epitope Generation via MHC-Context Co-Modeling and Contrastive Affinity Guidance
Haoyang Luan (Shanghai Jiao Tong University), Yang Yang (Shanghai Jiao Tong University)
GenerationDrug DiscoveryTransformerPrompt EngineeringDiffusion modelContrastive LearningBiomedical Data
🎯 What it does: Proposed the EpiCoCo framework, which can jointly generate high-affinity epitope sequences and structures in the MHC context.
Episodic Memory-Guided Controllable Experience Synthesis for Reinforcement Learning
Xiao Ma (Zhejiang Sci Tech University), Wu-Jun Li (Nanjing University)
Data SynthesisTransformerReinforcement LearningDiffusion modelScore-based ModelContrastive LearningTabularSequential
🎯 What it does: Propose an experience recall guided controllable diffusion model (EMCES) for synthesizing high-quality reinforcement learning experiences.
Epistemic Gain, Aleatoric Cost: Uncertainty Decomposition in Multi-Agent Debate for Math Reasoning
Dan Qiao (Chinese University of Hong Kong), Baoxiang Wang (Chinese University of Hong Kong)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextChain-of-Thought
🎯 What it does: This paper quantifies and distinguishes the epistemic gains and aleatoric costs in multi-agent debate (MAD) through a Bayesian uncertainty decomposition framework, and on this basis, designs an uncertainty-guided multi-agent reinforcement learning algorithm called UMAD, aiming to achieve higher epistemic information utilization and lower generation instability during the debate process.
Epistemic Uncertainty Quantification for Pre-trained VLMs via Riemannian Flow Matching
Li Ju (Uppsala University), Prashant Singh (Uppsala University)
Anomaly DetectionRepresentation LearningData-Centric LearningVision Language ModelScore-based ModelFlow-based ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose a framework called REPVLM based on Riemannian flow matching, which is used to estimate the probability density of pre-trained vision-language models (VLMs) in their embedding space, thereby obtaining the model's awareness of uncertainty in its representations.
EpiTwin: Spatiotemporal Graph Transformers for Epileptic sEEG Signal Reconstruction
Jingbo Yang (Tianjin University), Xiaofei Wang (Tianjin University)
RestorationGraph Neural NetworkTransformerGenerative Adversarial NetworkContrastive LearningTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: Propose a conditional spatial graph Transformer (EpiTwin) for reconstructing neural signals under sparse and irregular sEEG electrode sampling.
EPS3D: End-to-End Feed-Forward 3D Panoptic Segmentation
Runsong Zhu (Chinese University of Hong Kong), Chi-Wing Fu (Chinese University of Hong Kong)
SegmentationKnowledge DistillationTransformerGaussian SplattingImagePoint Cloud
🎯 What it does: Proposes an open-source vocabulary 3D panoramic segmentation framework called EPS3D, which is end-to-end and designed for multi-view uncalibrated images. It can simultaneously predict semantic, instance, and geometric features, and achieve perspective-consistent 3D segmentation through feature fusion integration;
EPSVec: Efficient and Private Synthetic Data Generation via Dataset Vectors
Amin Banayeeanzade (University of Southern California), Sai Praneeth Karimireddy (University of Southern California)
Data SynthesisSafty and PrivacyComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringGaussian SplattingText
🎯 What it does: Propose the EPSVEC method: Injecting a Dataset Vector with differential privacy into the activation space of LLMs to generate high-quality, privacy-compliant synthetic text in one go.
EqGINO: Equivariant Geometry-Informed Fourier Neural Operators for 3D PDEs
Sungwon Kim (KAIST), Chanyoung Park (KAIST)
Computational EfficiencyGraph Neural NetworkTransformerDiffusion modelAuto EncoderContrastive LearningPoint CloudMeshGraphBenchmarkPhysics Related
🎯 What it does: Propose a Fourier neural operator framework called EqGINO that combines equi-deformable geometric information for fast, coordinate-invariant approximation solutions of 3D PDEs
Equalized Generative Treatment: Matching f-divergences for Fairness in Generative Models
Alexandre Verine (école normale supérieure), Florian Le Bronnec (RIKEN Center for Computational Science)
GenerationData SynthesisFederated LearningSafty and PrivacyDiffusion modelGenerative Adversarial NetworkContrastive LearningImageText
🎯 What it does: Propose a new fairness definition called Equalized Generative Treatment (EGT), and provide theoretical analysis and experimental verification.
EquiCAD: A Geometric Equivariant Neural Network for 3D Shape Classification
Yonghao Su (Sichuan University), Xianggen Liu (Sichuan University)
ClassificationConvolutional Neural NetworkGraph Neural NetworkContrastive LearningMeshGraph
🎯 What it does: Propose the EquiCAD framework for classifying 3D shapes represented by CAD B-rep;
Equilibrium Pricing in Oligopolistic Data Markets
Bhaskar Ray Chaudhury (University of Illinois at Urbana-Champaign), Jiaxin Song (University of Illinois at Urbana-Champaign)
OptimizationFederated LearningReinforcement Learning from Human FeedbackReinforcement LearningTabularTime SeriesFinance Related
🎯 What it does: Studies the price game between budget-constrained buyers and strategic sellers in oligarchic data markets, exploring the existence of equilibrium pricing and the conditions for approximate equilibrium.
Equilibrium Propagation for Non-Conservative Systems
Antonino Emanuele Scurria (Université libre de Bruxelles), Serge Massar (Université libre de Bruxelles)
ClassificationOptimizationComputational EfficiencyRepresentation LearningConvolutional Neural NetworkDiffusion modelContrastive LearningImagePhysics Related
🎯 What it does: Extend Equilibrium Propagation (EP) to non-conservative systems, proposing two algorithms (Asymmetric EP and Dyadic EP) to achieve precise gradient learning.
Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning
Benhao Huang (Carnegie Mellon University), J Zico Kolter (Carnegie Mellon University)
OptimizationComputational EfficiencyRepresentation LearningTransformerTabularSequential
🎯 What it does: Proposes Equilibrium Reasoners (EqR), a model that achieves scalable reasoning by learning task-conditioned attractors.
Equivalence of Context and Parameter Updates in Modern Transformer Blocks
Adrian Goldwaser (University of Cambridge), Benoit Dherin (Google Research)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningImageText
🎯 What it does: Investigated and demonstrated that the influence of given context in modern Transformers (such as Gemma, Llama, Falcon, etc.) can be fully achieved by performing rank-1 patching on MLP weights and RMSNorm scales, thereby equating the context to parameter updates.
Equivariant Covariance Tensors: Guaranteed SPD Uncertainty for Tensor-Valued Geometric Learning
Ruihan Liu (Fudan University), Qingchao Jiang (East China University of Science and Technology)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkContrastive LearningPoint CloudGraphTabularPhysics Related
🎯 What it does: Proposed a full covariance uncertainty framework for predicting symmetric rank-2 tensors in E(3)-equivariant networks, utilizing matrix exponential mapping to ensure covariance is symmetric positive definite, and achieving robust training through Log-Euclidean loss.
Equivariant Latent Alignment via Flow Matching under Group Symmetries
Sunghyun Kim (Seoul National University), Joonseok Lee (Seoul National University)
GenerationData SynthesisRepresentation LearningTransformerDiffusion modelScore-based ModelFlow-based ModelAuto EncoderImagePoint CloudMesh
🎯 What it does: Propose the Residual Latent Flow (RLF) method, which corrects mismatches in equivariant representations in the latent space through flow matching, thereby improving the quality of novel view synthesis.
Equivariant Neural Networks for General Linear Symmetries on Lie Algebras
Chankyo Kim (University of Michigan), Maani Ghaffari (University of Michigan)
OptimizationComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerAuto EncoderContrastive LearningGaussian SplattingPoint CloudMeshTabularTime SeriesBenchmarkPhysics Related
🎯 What it does: Proposed Reductive Lie Neurons (ReLNs), achieving fully equivariant neural networks under the adjoint action of GL(n), and unifying the processing of vector and matrix features.
ERAlign: Energy-based Representation Alignment of GNNs and LLMs on Text-attributed Graphs
Xianlin Zeng (China RongTong Academy of Sciences Group Corporation Limited), Xiangyu Chen (China RongTong Academy of Sciences Group Corporation Limited)
ClassificationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextMultimodalityGraph
🎯 What it does: Propose the ERAlign framework, which aligns hierarchical representations of graph neural networks (GNN) and large language models (LLM) in a shared latent space using an energy model
Erased but Not Forgotten: How Backdoors Compromise Concept Erasure
Tobias Braun (Technical University of Darmstadt), Anna Rohrbach (Technical University of Darmstadt)
GenerationSafty and PrivacyKnowledge DistillationAdversarial AttackTransformerPrompt EngineeringDiffusion modelScore-based ModelImageTextMultimodality
🎯 What it does: This paper studies how to bypass the concept erasure mechanism in text-to-image diffusion models through backdoor attacks, proposing the Erasure Evasion Backdoor (EEB) attack, and evaluates its robustness against various erasure methods.
ERGeoBench: A Comprehensive Benchmark for Embodied Reasoning and Geo-localization in Multimodal Large Language Models
Kaiwen Xue (Beijing University of Posts and Telecommunications), Haoran Luo (Nanyang Technological University)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes ERGeoBench, a comprehensive benchmark for evaluating the embedded reasoning and geolocation capabilities of multimodal large language models.
Error Amplification Limits ANN-to-SNN Conversion in Continuous Control
Zijie Xu (Peking University), Zhaofei Yu (Peking University)
Computational EfficiencySpiking Neural NetworkReinforcement LearningTabularTime Series
🎯 What it does: This paper studies the conversion of pre-trained artificial neural networks (ANN) into spiking neural networks (SNN), and analyzes the error amplification problem in continuous control tasks.
Error Analysis of Discrete Flow with Generator Matching
Zhengyan Wan (East China Normal University), Guang Cheng (University of California, Los Angeles)
GenerationData SynthesisOptimizationComputational EfficiencyTransformerDiffusion modelScore-based ModelFlow-based ModelRectified FlowGenerative Adversarial NetworkContrastive LearningTabularTime SeriesSequentialStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper provides the first non-asymptotic error analysis framework for discrete flow models, systematically deriving the error propagation from generator matching to sample estimation;
Error Propagation in Dynamic Programming: From Stochastic Control to American Option Pricing
Andrea Della Vecchia (EPFL), Damir Filipovic
OptimizationTabularFinance Related
🎯 What it does: Studied the discrete-time stochastic optimal control (SOC) problem, proposed a dynamic programming algorithm based on kernel ridge regression (KRR) (KRR-DP), and provided theoretical analysis of error propagation and convergence guarantees.
Error Propagation Mechanisms and Compensation Strategies for Quantized Diffusion Models
Songwei Liu (ByteDance Inc), Xing Mei (ByteDance Inc)
GenerationDiffusion modelScore-based ModelImageVideo
🎯 What it does: Proposed and implemented a cumulative error compensation method for quantized diffusion models based on error propagation theory, called TCEC
Error-Driven Graph Augmentation for Mesh-Based PDE Surrogates
Xuan Minh Vuong Nguyen (Safran AI Research), Elie Hachem (Mines Paris PSL)
OptimizationExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkTransformerAuto EncoderContrastive LearningMeshGraphPhysics Related
🎯 What it does: Proposes MiSe-GNN, which dynamically generates multi-scale gain edges based on model-predicted posterior errors, forming an error-driven graph enhancement mechanism to improve long-range information propagation in grid-based PDE surrogates.
Escaping Mode Collapse in LLM Generation via Geometric Regulation
Xin Du (Waseda University), Kumiko Tanaka-Ishii (Waseda University)
GenerationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Conduct a dynamic analysis of the generation process in large-scale language models, proposing to view mode collapse as geometric collapse, and evaluating the accessibility of trajectories by measuring the correlation dimension; based on this, we propose Reinforced Mode Regulation (RMR), which applies low-rank damping to the KV cache during inference to suppress self-sustaining dominant directions, thereby reducing mode collapse phenomena such as loops and repetitions.
Escaping the Diversity Trap in Robotic Manipulation via Anchor-Centric Adaptation
Yanzhe Chen (National University of Singapore), Mike Zheng Shou (National University of Singapore)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningVision-Language-Action ModelFlow-based ModelImageTextMultimodality
🎯 What it does: To address the adaptation of pre-trained Vision-Language-Action (VLA) models on real robotic platforms, the paper proposes an Anchor-Centric Adaptation (ACA) two-stage framework, and verifies its effectiveness under limited data budgets on a real robotic arm.
Escaping the Likelihood Trap: Geometric Diversity Optimization for Long-Form Image Captioning
Qingmei Tang (Shandong University), Wenyu Wang (Shandong University)
GenerationOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelDiffusion modelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes the Geo-RL framework to address the 'likelihood trap' in long-text image captioning by generating multiple captions for each image and maximizing their geometric volume in the semantic embedding space, thereby encouraging diverse and detail-rich descriptions;
Escaping the Mode: Multi-Answer Reinforcement Learning in LMs
Isha Puri (Massachusetts Institute of Technology), Yoon Kim (Massachusetts Institute of Technology)
Explainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBiomedical DataReview/Survey Paper
🎯 What it does: This paper proposes a multi-answer reinforcement learning framework that enables language models to generate multiple possible answers in one inference process and provide confidence estimates, addressing the issues of entropy collapse and lack of multi-answer support caused by traditional single-answer RL.
Escaping the Subspace Trap: The Role of Optimizer Geometry in Model Width Expansion
Jiabei Chen (Institute of Automation, Chinese Academy of Sciences), Kang Liu (Institute of Automation, Chinese Academy of Sciences)
OptimizationComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelMixture of ExpertsContrastive LearningText
🎯 What it does: Generate larger models from pre-trained seed models through Model Width Expansion (MWE), and investigate the Subspace Trap phenomenon that occurs during continued pre-training; propose using optimizer geometry (e.g., Muon) to escape the subspace trap; verify the effectiveness of this method across different model scales and families.
Escaping the Verifier: Learning to Reason via Demonstrations
Locke Cai (Together Ai), Ivan Provilkov (Together Ai)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Leveraging expert demonstration data, the paper proposes a Relative Adversarial Reinforcement Optimization (RARO) method based on inverse reinforcement learning, training large language models (LLMs) to perform reasoning without the need for a verifier.
Escaping Whack-a-Mole: Optimizing Documentation as Repo-Specific Playbooks for Coding Agents
Yutong Cheng (Virginia Tech), Wei Cheng (NEC Laboratories America)
OptimizationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose the DOCSEARCH framework, treating code documentation as a repository-specific operation manual for large language model code agents, and modeling the optimization problem as a black-box search;
Esoteric Language Models: A Family of Any-Order Diffusion LLMs
Subham Sekhar Sahoo (Cornell Tech), Arash Vahdat (NVIDIA)
GenerationData SynthesisComputational EfficiencyTransformerLarge Language ModelDiffusion modelAuto EncoderContrastive LearningText
🎯 What it does: Propose a language model framework called Eso-LMs that integrates autoregressive (AR) and masked diffusion (MDM) models, achieving KV cache and parallel generation during the diffusion phase, and enabling smooth interpolation between two modes through a tunable α0.
Estimating Continuous Treatment Effects with Two-Stage Kernel Ridge Regression
Seok-Jin Kim (Columbia University), Kaizheng Wang (Columbia University)
TabularBenchmark
🎯 What it does: Proposed a two-stage kernel ridge regression framework for estimating the causal effect function of continuous treatments in the presence of confounded observational data;
Estimating Correlation Clustering Cost in Node-Arrival Stream
Kaiwen Liu (Indiana University), Qin Zhang (Indiana University)
OptimizationFederated LearningComputational EfficiencyContrastive LearningGraph
🎯 What it does: Under the node arrival flow model, a streaming algorithm called CApprox4 is proposed to approximately compute the clustering cost, which can estimate the optimal clustering cost with sublinear space and a constant number of passes.
Estimating Tail Risks in Language Model Output Distributions
Rico Angell (New York University), He He (Columbia University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposed an efficient method for estimating the probability of rare harmful outputs based on importance sampling, used to assess the tail risk of large language models under given queries.
Estimating the Empowerment of Language Model Agents
Jinyeop Song (Massachusetts Institute of Technology), Max Kleiman-Weiner (University of Washington)
Explainability and InterpretabilityRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIContrastive LearningText
🎯 What it does: Evaluate the capabilities of large language model agents in multi-turn text interactions, proposing a goal-free evaluation metric based on information theory—empowerment—and introducing the EELMA algorithm to estimate this metric.
Estimation of Treatment Effects Under Nonstationarity via the Truncated Policy Gradient Estimator
Ramesh Johari (Stanford University), Wenqian Xing (Stanford University)
Reinforcement LearningContrastive LearningTabularTime SeriesSequentialElectronic Health RecordsBenchmark
🎯 What it does: This paper proposes a new Truncated Policy Gradient (TPG) estimator to estimate the global average treatment effect (GATE) from a single random experimental trajectory in non-stationary Markov environments.
ETS: Energy-Guided Test-Time Scaling for Training-Free RL Alignment
Xiuyu Li (Renmin University of China), Ju Fan (Renmin University of China)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningScore-based ModelTextBenchmark
🎯 What it does: Proposes ETS, a training-free inference method that directly generates text by sampling from the optimal RL policy.
Euclean: Automated Geometry Problem Formalization with Unified Verification in Lean
Linbin Tang (Tsinghua University), Fan Yang (Microsoft Research)
OptimizationExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the EUCLEAN framework, which automatically converts natural language geometry problems into the MATHLIB formalization of Lean 4, completing a four-stage pipeline: constraint explicitation, configuration anchoring, mapping, and iterative repair.
Euler–Poincaré Neural Dynamics: A Geometric-Mechanics Framework for Scientific Simulation
Sungwoo Park (Korea University), Jiwoong Kim (Korea University)
Drug DiscoveryDiffusion modelContrastive LearningTime SeriesBiomedical DataPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose a framework for learning evolutionary operators based on geometric mechanics, called Euler-Poincaré Neural Dynamics (EPND), which reconstructs the traditional Koopman linear operator as a Lie group flow, and parameterizes the Euler-Poincaré equations through neural differentiable geometric fields.
Evaluating AI Grading on Real-World Handwritten College Mathematics: A Large-Scale Study Toward a Benchmark
Zhiqi Yu (University of California Irvine), Yifeng Yu (University of California Irvine)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringImageTextBenchmark
🎯 What it does: An automatic grading system combining OCR with large language models (LLM) was built and deployed at scale in the undergraduate calculus course at the University of California, Irvine, providing scores and feedback for thousands of handwritten homework assignments;
Evaluating and Explaining Prompt Sensitivity of LLMs Using Interactions
Ruiyang Qin (Tongji University), Wen Shen (Tongji University)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed an interaction-based metric called IPS to fine-grainedly evaluate the sensitivity of LLMs to prompts, and conducted experiments on 50 open-source models.
Evaluating and Rewarding LALMs for Expressive Role-Play TTS via Mean Continuation Log-Probability
Yong Ren (Institute of Automation, Chinese Academy of Sciences), Xuerui Yang (StepFun)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextMultimodalityAudio
🎯 What it does: Propose a 'Mean Continued Log Probability' (MCLP) based on LALM, which can serve both as an evaluation metric and as a reinforcement learning (RL) reward, to enhance the speech expression consistency in role-playing TTS.
Evaluating and Steering Modality Preferences in Multi-modal LLMs
Yu Zhang (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
Explainability and InterpretabilityRepresentation LearningAdversarial AttackTransformerLarge Language ModelPrompt EngineeringContrastive LearningImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Evaluate and guide the modality preferences of multimodal large language models.
Evaluating Bivariate Causal Statements Based on Mutual Compatibility
Erik Jahn (California Institute of Technology), Dominik Janzing (Amazon Research Tubingen)
OptimizationExplainability and InterpretabilityComputational EfficiencyLarge Language ModelPrompt EngineeringScore-based ModelContrastive LearningTextTabularReview/Survey PaperBenchmark
🎯 What it does: Evaluate and verify collections containing only binary causal statements (linear and graphical), using compatibility and incompatibility scores to detect internal consistency and measure credibility.
Evaluating Contextual Illegality: AI Compliance in Corporate Law Scenarios
Hilal Aka (Harvard Kennedy School, Georgetown University Law Center), Noam Kolt (Hebrew University)
ClassificationFederated LearningSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelAgentic AIPrompt EngineeringTextFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Evaluate the ability of large models to identify and reject contextually illegal requests in four corporate law scenarios.
Evaluating Language Models in Realistic Conversational Contexts
Ilija Subasic (Upwork Inc), Zhao Chen (Upwork Inc)
GenerationExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextRetrieval-Augmented Generation
🎯 What it does: Constructed the UPHELD dataset and used this dataset to evaluate the quality of large language models in real human-scale dialogues.
Evaluating LLM Uncertainty in Long-Form Generation Using Deterministic Ground Truth
Ido Amit (Technion), Ran El-Yaniv (Technion)
GenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the SALT benchmark for fine-grained uncertainty assessment in long-text generation, and conducted large-scale experiments on more than 50 LLMs;
Evaluating LLMs When They Do Not Know the Answer: Statistical Evaluation of Mathematical Reasoning via Comparative Signals
Zihan Dong (Rutgers University), Linjun Zhang (Rutgers University)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Propose a framework that utilizes pairwise comparison signals generated by LLMs as control variables, combined with semi-parametric inference to improve mathematical reasoning evaluation.
Evaluating Newtonian Mechanics in Video Generative Models with Real Physical Systems
Antonios Tragoudaras (University of Amsterdam), Stratis Gavves (University of Amsterdam)
GenerationExplainability and InterpretabilityConvolutional Neural NetworkTransformerPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelContrastive LearningOptical FlowImageVideoTextBenchmarkPhysics Related
🎯 What it does: Designed and implemented a physics knowledge evaluation framework called Morpheus, used to quantify the physical consistency of video generation models in Newtonian mechanics scenarios; simultaneously collected and annotated 130 controlled experimental videos as a benchmark.
Evaluating Object-Centric Models beyond Object Discovery
Krishnakant Singh (TU Darmstadt), Stefan Roth (TU Darmstadt)
Explainability and InterpretabilityRepresentation LearningData-Centric LearningTransformerSupervised Fine-TuningVision Language ModelDiffusion modelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: A scalable evaluation framework is proposed: using vision instruction-tuned large language models (VLMs) as evaluators to perform zero-shot evaluation of the practicality of various Object-Centric Learning (OCL) models across multiple VQA tasks; a unified evaluation task and metric (AwGA) are jointly designed, and an improved eGQA dataset is constructed; additionally, a baseline called mFRESA based on multi-objective reconstruction is proposed to enhance the representation quality of OCL.
Evaluating Parameter Efficient Methods for RLVR
Qingyu Yin (Zhejiang University), Jinjin Gu (INSAIT, Sofia University St. Kliment Ohridski)
Computational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringMixture of ExpertsTextBenchmark
🎯 What it does: Systematic evaluation of 12+ parameter-efficient fine-tuning (PEFT) methods under the Reinforcement Learning with Verifiable Rewards (RLVR) framework;
Evaluating Relational Reasoning in LLMs with REL
Lukas Fesser (Harvard University), Marinka Zitnik (Eric and Wendy Schmidt Center at Broad Institute of MIT and Harvard)
Drug DiscoveryTransformerLarge Language ModelPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the Relational Complexity (RC) metric and constructed an interdisciplinary REL benchmark based on it, covering relational reasoning tasks in three fields: algebra, life sciences, and chemistry.
Evaluating Robustness of Reasoning Models on Parameterized Logical Problems
Naïm Es-sebbani (University of Artois), Zied Bouraoui (University of Caen Basse Normandie)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Constructed a diagnostic 2-SAT benchmark based on a parameterizable structured 2-CNF formula to evaluate the robustness of large language models in logical reasoning.
Evaluating Sample Utility for Efficient Data Selection by Mimicking Model Weights
Tzu-Heng Huang (University of Wisconsin-Madison), Javier Movellan (Apple Inc)
ClassificationData-Centric LearningTransformerContrastive LearningImageText
🎯 What it does: Propose Mimic Score, which evaluates data sample quality through geometric alignment between pre-trained model weights and sample gradients, and build Grad-Mimic, a two-stage data selection framework based on this.
Evaluating the Representation Space of Diffusion Models via Self-Supervised Principles
Xiao Li (University of Michigan), Qing Qu (Ohio State University)
GenerationRepresentation LearningSupervised Fine-TuningDiffusion modelContrastive LearningImage
🎯 What it does: This paper proposes an evaluation framework for the internal representation space of diffusion models based on the principle of self-supervised learning, and introduces the Invariant Contamination Ratio (ICR) metric to quantify invariance and residual contamination in the feature space.
Even Faster Kernel Matrix Linear Algebra via Density Estimation
Rikhav Shah (Massachusetts Institute of Technology), Haike Xu (Massachusetts Institute of Technology)
OptimizationComputational EfficiencyContrastive LearningGaussian SplattingTabularTime Series
🎯 What it does: Proposes a faster kernel matrix linear algebra algorithm using kernel density estimation (KDE), covering approximate matrix-vector multiplication, matrix multiplication, spectral norm (vertex eigenvalue), and the sum of all elements in the kernel matrix.
Event2Vec: Processing neuromorphic events directly by representations in vector space
Wei Fang (Yale University), Priyadarshini Panda (University of Southern California)
ClassificationConvolutional Neural NetworkSpiking Neural NetworkTransformerAuto EncoderContrastive LearningPoint CloudSequential
🎯 What it does: Designed and proposed an event representation called event2vec based on the vector space, enabling Transformers to directly process sparse asynchronous event streams.
Every Step Counts: Decoding Trajectories as Authorship Fingerprints of dLLMs
Qi Li (National University of Singapore), Xinchao Wang (National University of Singapore)
Explainability and InterpretabilityTransformerLarge Language ModelDiffusion modelGaussian SplattingText
🎯 What it does: Utilize the decoding trajectories of discrete diffusion large language models (dLLM) as model author fingerprints to achieve model attribution;
Evidential Copula Concept Embedding Models
Yanjie Qiu (Shanghai University), Wei Liu (Tongji University)
ClassificationExplainability and InterpretabilityRepresentation LearningConvolutional Neural NetworkContrastive LearningImageTabularBenchmark
🎯 What it does: Propose a new concept embedding model EC-CEM, which utilizes Gaussian Copula combined with evidential deep learning to simultaneously model the marginal uncertainty of concepts and dependencies between concepts, thereby achieving higher accuracy and interpretability.
Evidential Reasoning Advances Interpretable Real-World Disease Screening
Chenyu Lian (Hong Kong Polytechnic University), Jing Qin (Hong Kong Polytechnic University)
Anomaly DetectionExplainability and InterpretabilityTransformerSupervised Fine-TuningAuto EncoderContrastive LearningImageBiomedical DataElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: Proposed an evidence-based reasoning framework for disease screening called EviScreen, which improves screening performance and interpretability by utilizing region-level evidence from historical cases.
EVMbench: Evaluating AI Agents on Smart Contract Security
Justin Wang (OpenAI), Olivia Watkins (OpenAI)
AI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented the EVMBENCH evaluation framework to measure the detection, patching, and attack capabilities of AI agents in the field of smart contract security.
EvoC2F: Compiling Tool Orchestration for Efficient and Evolvable LLM Agents
Lei Wei (Alibaba International Digital Commerce Group), Bin Wang (Peking University)
Autonomous DrivingOptimizationFederated LearningComputational EfficiencyAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the EvoC2F framework, which converts tool scheduling into a compilable intermediate representation (Plan IR), achieving efficient and reliable LLM agent tool usage;
EvoCF: Multi-Agent Collaboration via Agentic Memory-Driven Evolutionary Counterfactual Planning
Haotian Chi (Jilin University), Haiyan Yin (Agency for Science, Technology and Research)
OptimizationRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the EvoCF framework, which enhances the robustness of multi-agent collaborative planning through agentic memory, evolutionary counterfactual planning, and symbolic constraints.
EvoClaw: Evaluating AI Agents on Continuous Software Evolution
Gangda Deng (University Of Southern California), Xingyao Wang (OpenHands)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: Built the DeepCommit pipeline to reorganize noisy commits into executable Milestone DAGs, and proposed the SWE-Milestone benchmark to evaluate the performance of LLM agents in continuous software evolution.
EvoEGF-Mol: Evolving Exponential Geodesic Flow for Structure-based Drug Design
Yaowei Jin (Lingang Laboratory), Qian Shi (Lingang Laboratory)
Drug DiscoveryGraph Neural NetworkTransformerDiffusion modelScore-based ModelFlow-based ModelGraphBiomedical DataBenchmarkStochastic Differential Equation
🎯 What it does: Under the information geometry framework, a probabilistic pathway for structure-based drug design is proposed, generating molecular coordinates and ion categories simultaneously using the e-geodesic flow of exponential family distributions.
EvoGM: Learning to Merge LLMs via Evolutionary Generative Optimization
Tao Jiang (Southern University of Science and Technology), Jianguo Zhang (Southern University of Science and Technology)
OptimizationKnowledge DistillationRepresentation LearningHyperparameter SearchTransformerLarge Language ModelReinforcement LearningMixture of ExpertsGenerative Adversarial NetworkContrastive LearningText
🎯 What it does: Propose a model merging framework called EvoGM based on evolutionary generative optimization, which dynamically searches for the optimal merging coefficients using a learnable generator to achieve untrained merging of multiple expert LLMs.
Evolution of Benchmark: Black-Box Optimization Benchmark Design through Large Language Model
Chen Wang (South China University of Technology), Yue-Jiao Gong (South China University of Technology)
OptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextTabularBenchmark
🎯 What it does: Proposes EoB, which utilizes large language models for program evolution to automatically generate benchmark functions that are both similar to the target task landscape and capable of distinguishing different BBO algorithms.
Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning
Xin Qiu (Cognizant AI Lab), Risto Miikkulainen (Cognizant AI Lab)
OptimizationComputational EfficiencyHyperparameter SearchReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: This paper is the first to directly apply evolutionary strategies (ES) to full-parameter fine-tuning of language models with billions of parameters, without any dimensionality reduction, demonstrating the feasibility of ES on large-scale LLMs;
Evolution Strategies at the Hyperscale
Bidipta Sarkar (University of Oxford), Jakob Nicolaus Foerster (University of Oxford)
OptimizationComputational EfficiencyHyperparameter SearchTransformerLarge Language ModelReinforcement LearningDiffusion modelContrastive LearningGaussian SplattingTextTabularTime SeriesSequential
🎯 What it does: Proposes an evolutionary strategy method called EGGROLL, which utilizes low-rank perturbation matrices to significantly improve the arithmetic intensity and training throughput of large-scale (billion-parameter) models, and provides theoretical analysis of convergence and linearization in high-dimensional spaces.
Evolving Interdependent Operators with Large Language Models for Multi-Objective Combinatorial Optimization
junhao qiu, Qingfu Zhang (City University of Hong Kong)
OptimizationTransformerLarge Language ModelPrompt EngineeringTextTabular
🎯 What it does: Propose an E2OC framework based on large language models (LLMs) that automatically co-evolves combinations of multiple neighborhood search operators to enhance the performance of multi-objective evolutionary algorithms (MOEAs).
Evolving Interpretable Constitutions for Multi-Agent Coordination
Ujwal Kumar (Shibaura Institute of Technology), Phan Xuan Tan (Shibaura Institute of Technology)
Autonomous DrivingOptimizationFederated LearningExplainability and InterpretabilityRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: Automatically discover interpretable constitutional rules in a multi-agent Grid-World through LLM-driven genetic programming evolution, achieving agent collaboration and survival;
Evolving Quantitative Reasoning through Self-Play in Digital Twin Markets
Tianmi Ma (Wuhan University of Technology), Joey Tianyi Zhou (Agency for Science, Technology and Research)
TransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextTabularTime SeriesFinance Related
🎯 What it does: This paper proposes a quantification reasoning framework driven by large language models (LLMs) in digital twin markets, achieved through self-play. The core idea is to position the LLM as a high-level planner, dynamically generating and iteratively evaluating quantification tools;
EVOLVING ROLLOUTS: Harnessing Historical Experience for Web Agent Evolution in Reinforcement Learning
Sinuo Wang (Adelaide University), Wangchunshu Zhou (OPPO)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Designed the EVOLVING ROLLOUTS framework, combining policy parameter optimization with the dynamic evolution of an experience repository, enabling Web agents to extract reusable experiences from zero-variance rollouts (fully successful or fully failed groups), significantly improving sample efficiency and task performance.