International Conference on Learning Representations Β· 1682 papers
DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Bodhisattwa Prasad Majumder (Allen Institute for AI), Peter Clark (Allen Institute for AI)
CodeTransformerLarge Language ModelTextBenchmark
π― What it does: This paper proposes the DISCOVERYBENCH benchmark for evaluating the capabilities of large language models in automated data-driven scientific discovery (from data generation to hypothesis validation).
π― What it does: This paper proposes a Discrete Copula Diffusion (DCD) method that integrates discrete diffusion models with autoregressive Copula models during the inference phase to compensate for the lack of interdependence caused by the independent treatment of variables in each step of the diffusion model.
Discrete Latent Plans via Semantic Skill Abstractions
Haobin Jiang (Peking University), Zongqing Lu (Peking University)
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerAuto EncoderContrastive LearningMultimodality
π― What it does: This paper proposes LADS, a hierarchical method for learning language-conditioned discrete implicit plans through semantic skill abstraction.
Discretization-invariance? On the Discretization Mismatch Errors in Neural Operators
Wenhan Gao (Stony Brook University), Yi Liu (Stony Brook University)
CodeConvolutional Neural NetworkTabularPhysics Related
π― What it does: Theoretical and experimental research on discretization mismatch errors in neural operators is conducted, and a cross-resolution learning pipeline called CROP is proposed to eliminate this error.
DiSK: Differentially Private Optimizer with Simplified Kalman Filter for Noise Reduction
Xinwei Zhang (University of Southern California), Vahab Mirrokni (Google Research)
CodeOptimizationSafty and PrivacyConvolutional Neural NetworkTransformerSupervised Fine-TuningImageText
π― What it does: A differentially private optimizer based on simplified Kalman filtering (DiSK) is designed to significantly reduce the impact of DP noise on gradients during large-scale training.
DisPose: Disentangling Pose Guidance for Controllable Human Image Animation
Hongxiang Li (Peking University), Long Chen (Hong Kong University of Science and Technology)
CodeGenerationPose EstimationDiffusion modelVideo
π― What it does: Proposes the DisPose plugin, which achieves plug-and-play control without dense input by decoupling skeletal poses into motion fields and keypoint correspondences, significantly improving the quality of human image animation.
Dist Loss: Enhancing Regression in Few-Shot Region through Distribution Distance Constraint
Guangkun Nie (Peking University), Shenda Hong (Peking University)
CodeImageTime SeriesElectrocardiogram
π― What it does: Proposes Dist Loss, a loss function that simultaneously optimizes the distance between the predicted distribution and the label distribution in imbalanced regression tasks;
π― What it does: This paper proposes a distance metric based on tree slicingβDistance-based Tree-Sliced Wasserstein (Db-TSW)βfor comparing probability distributions in Euclidean space.
π― What it does: A teacher-student knowledge distillation framework called DistillHGNN is designed and implemented, utilizing HGNN+MLP to generate soft labels and high-order structural information, training a lightweight TinyGCN+MLP student model, which significantly improves inference speed and memory efficiency while maintaining high accuracy.
π― What it does: In the dataset distillation task, this paper proposes compressing large-scale datasets into small-scale synthetic data and parameterizing these synthetic instances through neural fields;
π― What it does: A model is proposed that simultaneously learns environment dynamics and policies within a Transformer, and utilizes MPC for planning to achieve sample-efficient meta-RL without parameter updates.
Distilling Structural Representations into Protein Sequence Models
Jeffrey Ouyang-Zhang (University of Texas at Austin), Daniel Jesus Diaz
CodeKnowledge DistillationProtein Structure PredictionGraph Neural NetworkTransformerAuto EncoderTextBiomedical Data
π― What it does: A protein language model called ISM is proposed, which only uses sequence input and can implicitly capture structural information, achieving better performance on various structure-related tasks.
Distributed Speculative Inference (DSI): Speculation Parallelism for Provably Faster Lossless Language Model Inference
Nadav Timor (Weizmann Institute of Science), David Harel (Weizmann Institute of Science)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: A new distributed inference algorithm called Distributed Speculative Inference (DSI) is proposed, which breaks the sequential limitations of traditional Speculative Inference (SI) by parallelizing the verification process, achieving sustainable acceleration of language model inference in a multi-GPU environment.
π― What it does: A distribution backtracking distillation (DisBack) method is proposed, which introduces the entire convergence trajectory of the teacher model into score distillation, thereby accelerating and improving the convergence speed and quality of single-step generative models.
Distribution-Free Data Uncertainty for Neural Network Regression
Domokos M. Kelen (HUN-REN SZTAKI), Andras A Benczur
CodeOptimizationTabular
π― What it does: This paper proposes an unbiased sample approximation of the nonparametric Continuous Ranked Probability Score (CRPS) to train non-deterministic neural networks for direct sampling and learning arbitrary distributions in regression tasks, thereby achieving distribution-independent uncertainty quantification.
Distributional Associations vs In-Context Reasoning: A Study of Feed-forward and Attention Layers
Lei Chen (New York University), Alberto Bietti (Flatiron Institute)
CodeTransformerLarge Language ModelText
π― What it does: This study investigates the different learning mechanisms of the feed-forward layer and the attention layer in Transformers regarding distributed associations and contextual reasoning, and explains their dynamics through experiments and theory.
DistRL: An Asynchronous Distributed Reinforcement Learning Framework for On-Device Control Agent
Taiyi Wang (University of Cambridge), Kun Shao (AI Centre)
CodeRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningMultimodality
π― What it does: Designed and implemented DISTRL, an asynchronous distributed reinforcement learning framework for efficiently fine-tuning multimodal large language models online on mobile devices, ultimately training device control agents with better control capabilities.
Divergence-enhanced Knowledge-guided Context Optimization for Visual-Language Prompt Tuning
Yilun Li (Capital Normal University), Wei Song (Capital Normal University)
CodeClassificationOptimizationTransformerPrompt EngineeringVision Language ModelImage
π― What it does: This paper proposes knowledge-guided context optimization (DeKg) based on HilbertβSchmidt independence criterion, which reduces bias towards pre-trained knowledge by regularizing the independence of learned prompts and pre-trained prompts, thereby enhancing the learning of task-specific knowledge in downstream tasks.
Diversity Empowers Intelligence: Integrating Expertise of Software Engineering Agents
Kexun Zhang (Salesforce AI Research), Caiming Xiong (Salesforce AI Research)
CodeOptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIText
π― What it does: A framework called DEI (Diversity Empowered Intelligence) is proposed to integrate and collaborate multiple Software Engineering (SWE) agents, thereby enhancing the automatic repair rate of GitHub issues.
Divide and Translate: Compositional First-Order Logic Translation and Verification for Complex Logical Reasoning
Hyun Ryu (Korea Advanced Institute of Science and Technology), Eunho Yang (Korea Advanced Institute of Science and Technology)
CodeLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: A neural-symbolic framework named CLOVER is proposed, which utilizes LLM to decompose sentences through logical dependency structures and gradually translate them into first-order logic formulas, followed by verification using a SAT solver, thereby achieving complex logical reasoning.
DLEFT-MKC: Dynamic Late Fusion Multiple Kernel Clustering with Robust Tensor Learning via Min-Max Optimization
Yi Zhang (National University of Defense Technology), En Zhu (National University of Defense Technology)
CodeOptimizationTabular
π― What it does: A multi-kernel clustering method DLEFT-MKC based on dynamic late fusion and tensor learning is proposed, utilizing min-max optimization to achieve clustering consistency and robustness;
Do Deep Neural Network Solutions Form a Star Domain?
Ankit Sonthalia (TΓΌbingen AI Center UniversitΓ€t TΓΌbingen), Seong Joon Oh (TΓΌbingen AI Center UniversitΓ€t TΓΌbingen)
CodeOptimizationSupervised Fine-TuningImage
π― What it does: The Star Domain Conjecture is proposed, suggesting that the solution set of deep neural networks forms a star domain rather than a simple convex set, and introduces the Starlight algorithm to find star models.
Do Egocentric Video-Language Models Truly Understand Hand-Object Interactions?
Boshen Xu (Renmin University of China), Qin Jin (Renmin University of China)
CodeRecognitionRetrievalDomain AdaptationTransformerLarge Language ModelContrastive LearningVideoTextBenchmark
π― What it does: This study investigates the limitations of first-person perspective video-text pre-training in understanding human-object interactions (HOI) and proposes the EgoHOIBench benchmark to evaluate the model's grasp of fine-grained semantic variations in HOI. Additionally, it introduces the asynchronous contrastive learning objective EgoNCE++, which generates hard negative samples for V2T using LLM or vocabulary, while maintaining noun-centric clustering on the T2V side to enhance the model's ability to distinguish verbs.
Do I Know This Entity? Knowledge Awareness and Hallucinations in Language Models
Javier Ferrando (Universitat Politècnica de Catalunya), Neel Nanda (ETH Zürich)
CodeTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderText
π― What it does: This paper utilizes Sparse Autoencoders (SAE) to identify linear directions in the representation space of large language models (Gemma 2 2B/9B, Llama 3.1 8B) that can distinguish between known and unknown entities, and verifies that these directions have a causal impact on knowledge rejection and hallucination behaviors in chat models; it also discovers an 'unknown' direction that can differentiate between the model's correct and incorrect responses to queries, indicating internal uncertainty signals within the model.
Do Large Language Models Truly Understand Geometric Structures?
Xiaofeng Wang (Shanghai Jiao Tong University), Rui Wang (Shanghai Jiao Tong University)
CodeRecognitionTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: The GeomRel dataset is proposed, focusing on geometric relationship recognition, evaluating LLM's understanding of geometric structures, and revealing its limitations through experiments.
Do LLMs ``know'' internally when they follow instructions?
Juyeon Heo (University of Cambridge), Jaya Narain (Apple)
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Identify a single dimension related to instruction following within LLMs and enhance adherence rates through representation engineering.
Do LLMs estimate uncertainty well in instruction-following?
Juyeon Heo (University of Cambridge), Jaya Narain (Apple)
CodeTransformerLarge Language ModelTextBenchmark
π― What it does: A systematic evaluation of the ability of large language models to estimate uncertainty in instruction-following tasks was conducted, and a dual-version (Controlled and Realistic) controlled benchmark was proposed;
Do Stochastic, Feel Noiseless: Stable Stochastic Optimization via a Double Momentum Mechanism
Tehila Dahan (Technion), Kfir Yehuda Levy
CodeOptimizationConvolutional Neural NetworkImage
π― What it does: A dual momentum mechanism is proposed, combining Anytime-SGD and STORM, resulting in a stochastic gradient descent algorithm with stable convergence rates that is insensitive to learning rates, along with its accelerated version.
Do Vision & Language Decoders use Images and Text equally? How Self-consistent are their Explanations?
Letitia Parcalabescu (Heidelberg University), Anette Frank (Heidelberg University)
CodeExplainability and InterpretabilityTransformerVision Language ModelTextMultimodality
π― What it does: This paper studies the differences in the utilization of image and text information by visual language model (VLM) decoders when answering and generating explanations, and evaluates their self-consistency.
π― What it does: This paper systematically studies the essence of WGAN's success by constructing a discrete GAN model and utilizing the computable Wasserstein distance. It proves that when the discriminator is a convolutional network, WGAN actually optimizes the block-level Wasserstein distance rather than the entire image-level distance.
DocMIA: Document-Level Membership Inference Attacks against DocVQA Models
Khanh Nguyen (Computer Vision Center, Universitat Autonoma de Barcelona), Dimosthenis Karatzas (Computer Vision Center, Universitat Autonoma de Barcelona)
π― What it does: This study investigates document-level membership inference attacks on the DocVQA model, designing both white-box and black-box attack methods without auxiliary data.
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: This paper demonstrates the shortcomings of current LLM refusal training in tense generalization by rewriting harmful requests into the past tense and conducts a systematic evaluation of multiple mainstream LLMs.
CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: A new security assessment method is proposed, utilizing 'Response Guided Question Augmentation (ReG-QA)' to generate semantically relevant natural questions on aligned LLMs, detecting the model's robustness in natural scenarios.
DoF: A Diffusion Factorization Framework for Offline Multi-Agent Reinforcement Learning
Chao Li (Xiamen University), Siqi Shen (Xiamen University)
CodeReinforcement LearningDiffusion model
π― What it does: Designed and implemented the DoF (Diffusion Factorization Framework), which decomposes collective diffusion models into individual diffusion models through noise and data factorization, achieving offline multi-agent reinforcement learning.
π― What it does: A domain-guided (DoG) transfer learning method is proposed, which fine-tunes a pre-trained diffusion model on the target domain and guides generation during the sampling process.
π― What it does: This paper proposes the infinite-horizon RMAB model PREF-RMAB, which can only observe preference feedback without scalar rewards, and designs an online algorithm DOPL that can explore, collect preference information, and make direct decisions in an unknown environment, theoretically proving its sublinear regret.
DOTS: Learning to Reason Dynamically in LLMs via Optimal Reasoning Trajectories Search
Murong Yue (George Mason University), Dong Yu (Tencent AI Lab)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
π― What it does: The DOTS method is proposed, allowing LLMs to autonomously plan the optimal reasoning path before answering; by defining atomic reasoning actions, searching for the best action trajectory, and training external or internal planners through supervised fine-tuning, the reasoning ability of LLMs is enhanced.
Doubly robust identification of treatment effects from multiple environments
Piersilvio De Bartolomeis (ETH Zurich), Fanny Yang (University of Michigan)
CodeTabular
π― What it does: A method named RAMEN is proposed, which utilizes multiple heterogeneous data sources to achieve unbiased estimation of treatment effects (ATE) without a complete causal graph by identifying nodes with invariant conditional distributions.
DPaI: Differentiable Pruning at Initialization with Node-Path Balance Principle
Lichuan Xiang (University of Warwick), Hongkai Wen (University of Warwick)
CodeOptimizationConvolutional Neural NetworkImage
π― What it does: A differentiable initialization pruning method based on node-path balance, DPaI, is proposed, which directly obtains a sparse subnetwork during the network initialization phase.
DRESSing Up LLM: Efficient Stylized Question-Answering via Style Subspace Editing
Xinyu Ma (Peking University), Yasha Wang (Peking University)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: The DRESS method is proposed, utilizing untrained representation editing techniques to enable large language models to output answers in a specified style when responding to questions.
π― What it does: A unified Transformer framework called DriveTransformer is designed, integrating perception, prediction, mapping, and planning for E2E autonomous driving tasks, achieving task parallelism, sparse representation, and streaming processing.
DRL: Decomposed Representation Learning for Tabular Anomaly Detection
Hangting Ye (Jilin University), Yi Chang (Jilin University)
CodeAnomaly DetectionRepresentation LearningTabularFinance Related
π― What it does: A framework for anomaly detection in tabular data based on Decomposed Representation Learning (DRL) is proposed, which maps the representations of positive samples to a constrained latent space using fixed orthogonal basis vectors, and enhances the distinction between positive and anomalous classes through decomposition loss and separation constraints.
DSBench: How Far Are Data Science Agents from Becoming Data Science Experts?
Liqiang Jing (University of Texas at Dallas), Dong Yu (Tencent AI Lab)
CodeLarge Language ModelAgentic AIMultimodalityTabularBenchmarkFinance Related
π― What it does: Proposed and implemented the DSBench data science benchmark, collecting 466 data analysis tasks from ModelOff and 74 data modeling tasks from Kaggle, designed a complete evaluation workflow, and made the data and code publicly available;
DSPO: Direct Score Preference Optimization for Diffusion Model Alignment
Huaisheng Zhu (Pennsylvania State University), Vasant G Honavar (Pennsylvania State University)
CodeGenerationOptimizationReinforcement Learning from Human FeedbackDiffusion modelImageText
π― What it does: This paper proposes and implements an algorithm called DSPO, which directly uses score matching for fine-tuning text-to-image diffusion models, aiming to make the generated images more aligned with human preferences.
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This study implements a dual-process learning method that coexists structured context learning and weight learning, utilizing active forgetting and temporary forgetting techniques to maintain the model's flexible adaptation to new and old vocabulary.
DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads
Guangxuan Xiao (Massachusetts Institute of Technology), Song Han (Massachusetts Institute of Technology)
CodeRetrievalCompressionOptimizationComputational EfficiencyTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: The DuoAttention framework is proposed for inference in long-context large language models, which reduces memory and latency by identifying retrieval heads and streaming heads, using full KV caching only for retrieval heads.
CodeSegmentationDomain AdaptationAutonomous DrivingLarge Language ModelImage
π― What it does: The DynAlign framework is proposed, achieving automatic alignment of different label spaces and fine-grained segmentation in unsupervised cross-domain segmentation.
Dynamic Assortment Selection and Pricing with Censored Preference Feedback
Jung-hun Kim (Seoul National University), Min-hwan Oh (Seoul National University)
CodeRecommendation SystemOptimizationTabular
π― What it does: A dynamic multi-product selection and pricing framework based on the truncated multinomial logit (C-MNL) model is proposed, addressing the issue of buyers filtering and only providing sparse feedback on purchases under price decisions.
π― What it does: By combining 3D Gaussian point clouds with deformable fields, DG-Mesh can reconstruct temporally consistent, high-quality meshes from dynamic videos and track vertex motion, supporting topological changes.
Dynamic Mixture of Experts: An Auto-Tuning Approach for Efficient Transformer Models
Yongxin Guo (Chinese University of Hong Kong), Tao Lin (Westlake University)
CodeOptimizationComputational EfficiencyTransformerMixture of ExpertsMultimodality
π― What it does: A dynamic mixture of experts (DYNMOE) framework is proposed, which can automatically determine the number of experts and the number of experts to be activated for each token during the training process.
Dynamic Multimodal Evaluation with Flexible Complexity by Vision-Language Bootstrapping
Yue Yang (Shanghai AI Laboratory), Wenqi Shao (University of Hong Kong)
CodeGenerationData SynthesisTransformerLarge Language ModelVision Language ModelImageTextMultimodality
π― What it does: This paper proposes a dynamic multimodal evaluation protocol called Vision-Language Bootstrapping (VLB), which generates new evaluation samples through adaptive enhancement of images and text, aiming to reduce data contamination and increase evaluation complexity.
Dynamic-LLaVA: Efficient Multimodal Large Language Models via Dynamic Vision-language Context Sparsification
Wenxuan Huang (East China Normal University), Shaohui Lin (Key Laboratory of Advanced Theory and Application in Statistics and Data Science)
CodeGenerationComputational EfficiencyTransformerLarge Language ModelVision Language ModelTextMultimodality
π― What it does: The Dynamic-LLaVA framework is proposed to achieve dynamic visual-language context sparsification for multi-modal large language models (MLLM), significantly reducing computational and memory overhead during inference.
Dynamic-SUPERB Phase-2: A Collaboratively Expanding Benchmark for Measuring the Capabilities of Spoken Language Models with 180 Tasks
Chien-yu Huang (National Taiwan University), Hung-yi Lee (National Taiwan University)
CodeClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningMultimodalityBenchmarkAudio
π― What it does: Released and maintained the Dynamic-SUPERB Phase-2 evaluation benchmark, which includes 180 instruction-based tasks covering speech, music, and general audio, and established a task classification system and automated evaluation process; simultaneously conducted a systematic evaluation of various public models on this benchmark.
Dynamical Diffusion: Learning Temporal Dynamics with Diffusion Models
Xingzhuo Guo (Tsinghua University), Mingsheng Long (Tsinghua University)
CodeDiffusion modelVideoTime Series
π― What it does: A new framework called Dynamical Diffusion (DyDiff) is proposed for learning temporal dynamics, particularly applying diffusion models in time prediction tasks.
π― What it does: This paper proposes Dynamic Prompt Tuning (DynaPrompt), which enhances the zero-shot generalization performance of the CLIP model under distribution shifts by adaptively selecting, updating, and appending prompts through an online prompt buffer.
Dysca: A Dynamic and Scalable Benchmark for Evaluating Perception Ability of LVLMs
Jie Zhang (Chinese Academy of Sciences), Xilin Chen (Chinese Academy of Sciences)
CodeData SynthesisAdversarial AttackTransformerVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
π― What it does: A dynamic and scalable benchmark named Dysca has been developed to evaluate the perceptual capabilities of large visual language models (LVLMs), which includes 20 perceptual sub-tasks, 4 types of image scenes (Clean, Corruption, Print-Attacking, Adversarial-Attacking), and 3 types of question formats (multiple choice, true/false, open-ended), generating a total of 617K visual-text question-answer pairs.
Eagle: Exploring The Design Space for Multimodal LLMs with Mixture of Encoders
Min Shi (Georgia Tech), Guilin Liu (NVIDIA)
CodeTransformerLarge Language ModelMixture of ExpertsVision Language ModelImageTextMultimodality
π― What it does: The system evaluates and constructs a multi-modal large language model (MLLM) framework called Eagle, which combines multiple visual encoders and enhances the model's perceptual capabilities by gradually adding visual experts and pre-alignment techniques.
π― What it does: An improved method (D2PO) is proposed by incorporating a time decay factor Ξ³ into Direct Preference Optimization (DPO), allowing early tokens to have a greater weight during the alignment process through dynamic weighting, thereby reducing length bias and enhancing the model's alignment effectiveness.
π― What it does: A method is proposed to accelerate the training of Rectified Flow and Diffusion models by extending the distance between noise and sample paths.
π― What it does: A cardiac ultrasound (ECHO) video generation model called ECHOPulse has been developed based on ECG time series signals, capable of quickly generating high-quality, ECG-synchronized cardiac ultrasound videos without expert annotations.
π― What it does: Proposes two edge-based graph prompt tuning methods, EdgePrompt and EdgePrompt+, for adapting downstream tasks on pre-trained GNN models without the need for fine-tuning.
EDiT: A Local-SGD-Based Efficient Distributed Training Method for Large Language Models
Jialiang Cheng (Ant Group), Jian Sha (Ant Group)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: For the distributed training of large-scale language models, two efficient distributed training methods, EDIt and A-EDIt, are proposed, integrating Local SGD with model sharding, and introducing techniques such as hierarchical synchronization, prefetching, and pseudo-gradient penalties to enhance efficiency and stability.
π― What it does: This paper proposes a framework called ARAM, which utilizes an acceptance-rejection method and enhanced MDP to achieve action-constrained learning based on unconstrained RL algorithms.
CodeKnowledge DistillationRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningAgentic AISequential
π― What it does: A method for active imitation learning based on Random Network Distillation (RND), called RND-DAgger, is proposed to reduce expert intervention and improve learning efficiency.
Efficient and Context-Aware Label Propagation for Zero-/Few-Shot Training-Free Adaptation of Vision-Language Model
Yushu Li (South China University of Technology), Xun Xu (Institute for Infocomm Research A*STAR)
CodeClassificationDomain AdaptationComputational EfficiencyGraph Neural NetworkVision Language ModelImageMultimodality
π― What it does: A graph-based label propagation method is proposed for zero-shot/few-shot training adaptation of visual-language models, aimed at improving label efficiency and inference efficiency.
π― What it does: This paper proposes an efficient and verifiable causal discovery algorithm based on the causal Markov assumption, which can identify latent variables and the complete causal structure in polynomial time, even in the presence of latent variables and complex causal relationships.
Efficient Automated Circuit Discovery in Transformers using Contextual Decomposition
Aliyah R. Hsu (University of California Berkeley), Bin Yu (University of California Berkeley)
CodeTransformerText
π― What it does: A context decomposition method suitable for Transformers (CD-T) is proposed, which enables automated circuit discovery and can quickly identify computational subgraphs responsible for specific tasks in the model without the need for training or manual examples.
Efficient Biological Data Acquisition through Inference Set Design
Ihor Neporozhnii (University of Toronto), Jason Hartford (Valence Labs)
CodeDrug DiscoveryImageBiomedical Data
π― What it does: This study proposes a method for efficiently acquiring biological data through inference set design, aiming to achieve system target accuracy by selecting the smallest candidate set, thereby reducing experimental costs in drug discovery.
π― What it does: This paper proposes an Efficient Cross-Episode Transformers (ECET) model for online meta reinforcement learning, which achieves rapid task adaptation by utilizing cross-episode and intra-episode experiences.
Efficient Dictionary Learning with Switch Sparse Autoencoders
Anish Mudide (Massachusetts Institute of Technology), Christian Schroeder de Witt (University of Oxford)
CodeOptimizationComputational EfficiencyMixture of ExpertsAuto EncoderText
π― What it does: A Switch Sparse Autoencoder (Switch SAE) is proposed, which splits the training of sparse autoencoders into multiple small experts through a Mixture of Experts architecture, significantly reducing FLOPs and memory overhead while maintaining interpretable features.
Efficient Discovery of Pareto Front for Multi-Objective Reinforcement Learning
Ruohong Liu (University of Oxford), Jiang Bian (Microsoft Research Asia)
CodeOptimizationReinforcement Learning
π― What it does: A two-stage C-MORL algorithm is proposed, which initializes training of multiple fixed preference policies using Pareto, and then extends the Pareto front and achieves policy allocation through constrained optimization.
Efficient Evolutionary Search Over Chemical Space with Large Language Models
Haorui Wang (Georgia Institute of Technology), Chao Zhang (Georgia Institute of Technology)
CodeOptimizationDrug DiscoveryTransformerLarge Language ModelPrompt EngineeringTextBiomedical Data
π― What it does: MOLLEO is proposed, a framework that integrates large language models (LLMs) into evolutionary algorithms for generating compounds that meet multiple attribute objectives.
Efficient Inference for Large Language Model-based Generative Recommendation
Xinyu Lin (National University of Singapore), Tat-Seng Chua (National University of Singapore)
CodeRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A Speculative Decoding (SD) acceleration framework called AtSpeed is proposed for LLM generative recommendation, which designs two strategies: strict Top-K verification and relaxed sampling verification, and achieves efficient draft-then-verify inference based on this.
π― What it does: An efficient mask autoencoder E-MAC based on density embedding is proposed for video object counting, along with the design of a Spatial Adaptive Mask (SAM) and Temporal Co-Fusion (TCF) module.
Efficient Model-Based Reinforcement Learning Through Optimistic Thompson Sampling
Jasmine Bayrooti (University of Cambridge), Amanda Prorok (University of Cambridge)
CodeRobotic IntelligenceReinforcement Learning
π― What it does: A HOT-GP algorithm based on joint Gaussian Process dynamics and reward models is proposed, achieving optimistic exploration in model-based reinforcement learning, thereby significantly improving sampling efficiency.
Efficient Neuron Segmentation in Electron Microscopy by Affinity-Guided Queries
Hang Chen (Tsinghua University), Xiaolin Hu (Tsinghua University)
CodeSegmentationConvolutional Neural NetworkImage
π― What it does: A query-based neuron segmentation method called AGQ is proposed, which directly predicts segmentation results on 3D electron microscopy images, eliminating the traditional watershed and clustering steps.
Efficient Perplexity Bound and Ratio Matching in Discrete Diffusion Language Models
Etrit Haxholli (MetaDialog Research), Eli Waxman (MetaDialog Research)
CodeGenerationOptimizationComputational EfficiencyTransformerLarge Language ModelDiffusion modelText
π― What it does: This paper proposes a new KL divergence theorem and a more compact perplexity upper bound in the discrete diffusion model, and replaces ratio matching with cross-entropy training to achieve more efficient language modeling.
Efficient Reinforcement Learning with Large Language Model Priors
Xue Yan (Institute of Automation, Chinese Academy of Science), Jun Wang (University College London)
CodeTransformerLarge Language ModelReinforcement LearningText
π― What it does: The study treats large language models (LLMs) as action prior distributions, implemented through variational inference and posterior sampling within a reinforcement learning framework, significantly enhancing the sampling efficiency for online and offline text decision tasks.
π― What it does: A training framework based on finite-dimensional distribution matching (FDM) is proposed for the efficient training of Neural Stochastic Differential Equations (Neural SDE).
Efficiently Democratizing Medical LLMs for 50 Languages via a Mixture of Language Family Experts
Guorui Zheng (Chinese University of Hong Kong), Benyou Wang (Chinese University of Hong Kong)
CodeLarge Language ModelMixture of ExpertsTextBiomedical Data
π― What it does: This paper constructs a high-quality medical dataset covering 12 high-resource languages and implements a large medical language model for 50 languages using a Mixture of Experts (MoE) architecture;
ELBOing Stein: Variational Bayes with Stein Mixture Inference
Ola RΓΈnning (University of Copenhagen), Thomas Hamelryck (University of Copenhagen)
CodeMixture of ExpertsTabularSequential
π― What it does: A new particle variational Bayesian method called Stein Mixture Inference (SMI) is proposed, which approximates the posterior by treating each particle as a component of a mixture model, avoiding the variance collapse problem of traditional SVGD.
π― What it does: A label-free unsupervised core sample selection method called ELFS is proposed, which can select a high-quality subset of training samples without relying on true labels.
ELICIT: LLM Augmentation Via External In-context Capability
Futing Wang (Westlake University), Tao Lin (Westlake University)
CodeTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: A framework named ELICIT has been developed, which enhances the multi-task adaptability of large language models by externally storing task vectors and dynamically retrieving them during inference, without increasing the number of additional tokens.
Belinda Z. Li (Massachusetts Institute of Technology), Jacob Andreas (Massachusetts Institute of Technology)
CodeRecommendation SystemTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes an interactive task requirement mining framework called GATE, which actively guides users to express their preferences by generating free-form questions and examples, and uses the obtained information to train personalized prediction models.
Eliminating Position Bias of Language Models: A Mechanistic Approach
Ziqi Wang (University of Illinois Urbana-Champaign), Heng Ji (University of Illinois Urbana-Champaign)
CodeTransformerLarge Language ModelText
π― What it does: A training-agnostic, zero-shot PINE method is proposed, utilizing cross-document bidirectional attention and position reallocation to eliminate the positional information bias of language models.
EmbedLLM: Learning Compact Representations of Large Language Models
Richard Zhuang (University of California), Kannan Ramchandran (University of California)
CodeOptimizationRepresentation LearningTransformerLarge Language ModelText
π― What it does: Proposes the EmbedLLM framework, which learns compact vector representations of LLMs for multi-model management and downstream tasks.
Emergence of a High-Dimensional Abstraction Phase in Language Transformers
Emily Cheng (Universitat Pompeu Fabra), Marco Baroni (SISSA)
CodeTransformerLarge Language ModelText
π― What it does: The study investigates the trajectory of the intrinsic dimensionality of representations at various layers of the Transformer language model as the number of layers changes, and finds a commonly occurring high-dimensional peak corresponding to a key stage in abstract language processing.
CodeFederated LearningSafty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper studies the security vulnerabilities in Federated Instruction Tuning (FedIT). It first proposes a covert security attack method based on misaligned data and, on this basis, designs a post-processing defense scheme that fine-tunes the global model using automatically generated secure aligned data to repair the security deviations caused by the attack.
EMMA: Empowering Multi-modal Mamba with Structural and Hierarchical Alignment
Yifei Xing (Institute of Computing Technology Chinese Academy of Sciences), Yaowei Wang (Pengcheng Laboratory)
CodeImage TranslationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodality
π― What it does: By pixel-level alignment and multi-scale feature fusion, the quality of Mamba-based multimodal LLM visual representations is improved.
Donghwan Rho (Seoul National University), Jung Hee Cheon (Seoul National University)
CodeSafty and PrivacyComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes a homomorphic encryption-friendly Transformer architecture and achieves efficient personalized fine-tuning and inference in an encrypted environment.
Energy-based Backdoor Defense Against Federated Graph Learning
Guancheng Wan (Wuhan University), Mang Ye (Wuhan University)
CodeFederated LearningGraph Neural NetworkGraph
π― What it does: Designed and implemented the FedTGE framework, which utilizes an energy-based model to enable energy awareness in local graph neural networks for defense against backdoor attacks in federated graph learning.
Enhancing Clustered Federated Learning: Integration of Strategies and Improved Methodologies
Yongxin Guo (Chinese University of Hong Kong), Tao Lin (Westlake University)
CodeFederated LearningImage
π― What it does: Proposes the HCFL four-layer framework and the improved HCFL+, unifying and enhancing clustering federated learning methods to address data heterogeneity.
Enhancing Cognition and Explainability of Multimodal Foundation Models with Self-Synthesized Data
Yucheng Shi (University of Georgia), Ninghao Liu (University of Georgia)
CodeClassificationExplainability and InterpretabilitySupervised Fine-TuningContrastive LearningImageMultimodalityBiomedical Data
π― What it does: This paper proposes a visual fine-grained classification framework for self-synthesized interpretable answers, enhancing the recognition and explanation capabilities of large models through multi-round self-sampling and fine-tuning.
π― What it does: The study found that the initial random seed has a significant impact on the generation of text-to-image combinations, proposing to mine reliable seeds and use them to generate self-supervised data for fine-tuning, thereby significantly improving the accuracy of quantity and spatial relationship generation.