π― What it does: This paper proposes a single-step diffusion model RFfusion based on Rectified Flow for efficient image fusion, balancing fusion quality and speed.
Efficient Speech Language Modeling via Energy Distance in Continuous Latent Space
Zhengrui Ma (Chinese Academy of Sciences), Min zhang
CodeGenerationData SynthesisComputational EfficiencyTransformerLarge Language ModelAudio
π― What it does: A new framework for speech language modeling in continuous latent space, called SLED, is proposed, which uses Energy Distance as the training objective and directly performs autoregressive modeling on continuous acoustic vectors.
π― What it does: This paper proposes an efficient method for training low-rank recurrent neural networks (RNNs), including online recursive least squares (RLS) learning based on a neural engineering framework, using continuous orthogonal matching pursuit (OMP) to find minimal networks, mapping infinite-width low-rank RNNs to Gaussian processes (GP) and optimizing basis function distributions through maximum marginal likelihood, as well as an active learning sampling strategy based on maximum predictive variance.
Efficient Training-Free Online Routing for High-Volume Multi-LLM Serving
Fangzhou Wu (University of Wisconsin-Madison), Sandeep Silwal (University of Wisconsin-Madison)
CodeLarge Language ModelBenchmark
π― What it does: A training-free online multi-LLM routing algorithm is proposed, which can efficiently allocate requests to the best model under high query volume and limited token budget.
π― What it does: Two efficient multilingual CLIP downstream fine-tuning strategies (TaPCL and CiPCL) are proposed, achieving significant reductions in computational costs while maintaining multilingual capabilities.
EfficientNav: Towards On-Device Object-Goal Navigation with Navigation Map Caching and Retrieval
Zebin Yang (Peking University), Meng Li (Peking University)
CodeOptimizationComputational EfficiencyRobotic IntelligenceTransformerLarge Language ModelMultimodalityRetrieval-Augmented Generation
π― What it does: A zero-range target navigation system, EfficientNav, has been implemented on local devices, using a small LLM as a planner. It compresses and reuses navigation map information through discrete KV caching, attention clustering, and semantic-aware retrieval to complete path planning from the starting point to the designated object.
π― What it does: EfficientVLA is proposed, a fully training-independent and structured inference acceleration framework designed for Diffusion-based Vision-Language-Action (VLA) models. It achieves significant improvements in inference speed and computational cost by pruning language module layers, selecting task-aware visual tokens, and caching intermediate features of the diffusion action head.
Effortless, Simulation-Efficient Bayesian Inference using Tabular Foundation Models
Julius Vetter (University of TΓΌbingen), Jakob H. Macke (Max Planck Institute for Intelligent Systems)
CodeTransformerTabularBenchmark
π― What it does: A training-free simulation-based Bayesian inference method NPE-PFN is proposed, using the pre-trained TabPFN as a conditional density estimator to perform autoregressive inference of multi-dimensional posteriors.
Yancheng Zhang (Institute of Artificial Intelligence University of Central Florida), Chen Chen (Institute of Artificial Intelligence University of Central Florida)
π― What it does: An exchangeable 2D/3D Gaussian scattering framework, EGGS, is proposed for achieving high-quality appearance rendering and precise geometric reconstruction simultaneously.
EgoDTM: Towards 3D-Aware Egocentric Video-Language Pretraining
Boshen Xu (Renmin University of China), Qin Jin (Renmin University of China)
CodeDepth EstimationRetrievalRobotic IntelligenceTransformerLarge Language ModelContrastive LearningVideoTextMultimodality
π― What it does: This paper proposes EgoDTM, a first-person video-language pre-training model that integrates a lightweight 3D visual decoder and spatially aware subtitles.
CodeObject DetectionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVideoMultimodalityChain-of-Thought
π― What it does: Proposes the EgoThinker framework, constructs a large-scale egocentric QA dataset EgoRe-5M, and achieves significant improvements in first-person video reasoning and fine-grained spatiotemporal localization through two-stage supervised fine-tuning and reinforcement learning training.
Elevating Visual Perception in Multimodal LLMs with Visual Embedding Distillation
Jitesh Jain (Georgia Tech), Jianwei Yang (Meta Superintelligence Labs)
CodeKnowledge DistillationRepresentation LearningTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: This paper proposes a multimodal large language model (MLLM) training framework named VisPer-LM, which enhances visual perception capabilities by optimizing predictive embedding of intermediate layer representations in LLM during the pre-training phase, distilling knowledge from expert visual perception encoders into the hidden representations of LLM.
Embedding Principle of Homogeneous Neural Network for Classification Problem
Jiahan Zhang (Shanghai Jiao Tong University), Tao Luo (Shanghai Jiao Tong University)
CodeClassificationOptimizationTabular
π― What it does: This study proves that in homogeneous (bias-free, positive 1-homogeneous activation) neural networks, linear isometric transformations using neuron or channel splitting can embed the KKT points of the minimum norm maximum margin problem of smaller networks into the corresponding KKT points of larger networks. Moreover, during the gradient flow training process, this mapping preserves the trajectory and direction limits.
π― What it does: This paper proposes the E AD 3 paradigm, which combines human driving cognition (EEG) with visual features through contrastive learning to train a driving thought model and integrate it into various end-to-end autonomous driving frameworks to enhance planning performance.
Runling Long (Harbin Institute of Technology), Liqiang Nie (City University of Hong Kong)
CodeObject DetectionRobotic IntelligenceTransformerLarge Language ModelAgentic AIImageMultimodality
π― What it does: This paper proposes the task of 'Embodied Crowd Counting', utilizing drones to actively explore large outdoor scenes and perform crowd counting.
Emergence and Evolution of Interpretable Concepts in Diffusion Models
Berk Tinaz (University of Southern California), Mahdi Soltanolkotabi (University of Southern California)
CodeGenerationExplainability and InterpretabilityDiffusion modelAuto EncoderImageText
π― What it does: This study investigates the internal mechanisms of diffusion models, analyzes the evolution of concepts during the generation process, and utilizes sparse autoencoders to achieve interpretable editing control.
Emergence of Linear Truth Encodings in Language Models
Shauli Ravfogel (New York University), Alberto Bietti (Flatiron Institute)
CodeTransformerText
π― What it does: A single-layer Transformer toy model was constructed, and a synthetic dataset was designed in accordance with the Truth Co-occurrence Hypothesis to study the generation mechanism of linear truth subspaces.
Empirical Study on Robustness and Resilience in Cooperative Multi-Agent Reinforcement Learning
Simin Li (Beihang University), Xianglong Liu (Beihang University)
CodeRobotic IntelligenceReinforcement Learning
π― What it does: Large-scale experiments were conducted in real environments to systematically evaluate the relationship between cooperation, robustness, and resilience, and to explore the impact of hyperparameters on all three;
Enabling Differentially Private Federated Learning for Speech Recognition: Benchmarks, Adaptive Optimizers, and Gradient Clipping
Martin Pelikan (Apple), Tatiana Likhomanenko (Apple)
CodeRecognitionFederated LearningSafty and PrivacyTransformerBenchmarkAudio
π― What it does: This paper studies the combination of federated learning and differential privacy in speech recognition tasks, proposing a training strategy based on hierarchical gradient clipping and hierarchical gradient normalization, and providing theoretical convergence analysis and practical benchmarks.
CodeGenerationComputational EfficiencyTransformerLarge Language ModelDiffusion modelText
π― What it does: Proposes the E2D2 encoding-decoding diffusion architecture, which splits the diffusion model into an encoder (handling clean context) and a lightweight decoder (multi-step denoising), significantly improving inference and training efficiency.
CodeGenerationData SynthesisOptimizationDrug DiscoveryGraph Neural NetworkDiffusion modelGraphPhysics Related
π― What it does: This paper proposes a framework that directly embeds physical energy into the loss function, implemented in both continuous and discrete physical systems, primarily targeting regression and generation tasks for molecular and spin systems.
π― What it does: Proposes the Energy Matching framework, which combines flow matching with energy-based models to achieve the OT path from noise to data and forms a Boltzmann distribution at the data approximation.
π― What it does: An enhanced self-distillation framework is proposed, using a lightweight ANN branch to guide SNN training, balancing rate-coded gradients and significantly reducing BPTT costs.
Enhancing Deep Batch Active Learning for Regression with Imperfect Data Guided Selection
Yinjie Min (Nankai University), Yongdao Zhou (Nankai University)
CodeOptimizationData-Centric LearningTabularTime SeriesFinance Related
π― What it does: This paper proposes to improve active learning for regression tasks by using auxiliary data through density ratio weighted loss estimation, overcoming the challenge of estimating prediction uncertainty when there are no labels.
π― What it does: This paper views unconstrained adversarial attacks as a preference alignment problem and proposes a two-stage framework called APA: first aligning visual consistency using LoRA and a differentiable similarity reward, and then enhancing attack effectiveness through dual-path attack guidance and diffusion enhancement, significantly improving black-box transfer rates.
π― What it does: This paper proposes a robust pooling method based on the principal singular vector, RS-Pool, to enhance the adversarial robustness of GNNs in graph classification tasks.
Enhancing GUI Agent with Uncertainty-Aware Self-Trained Evaluator
Gongwei Chen (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningContrastive LearningMultimodalitySequential
π― What it does: A framework for Uncertainty-Aware Reinforcement Self-Training (URST) is proposed, which trains a GUI trajectory evaluator using a lightweight multimodal large language model (MLLM) that can self-generate reasoning and judgments and continuously improve.
π― What it does: This paper addresses the problem of thermal infrared image enhancement by proposing the Progressive Prompt Fusion Network (PPFN) and Selective Progressive Training (SPT) methods, which can simultaneously handle noise, blur, and low contrast degradation in single or composite degradation scenarios.
Enhancing Interpretability in Deep Reinforcement Learning through Semantic Clustering
Liang Zhang (University of Arizona), Adarsh Pyarelal (University of Arizona)
CodeExplainability and InterpretabilityReinforcement LearningImage
π― What it does: This study investigates the semantic clustering characteristics in deep reinforcement learning (DRL) and proposes a semantic clustering module that combines feature dimensionality reduction with online clustering to enhance the interpretability of DRL models.
Enhancing LLM Watermark Resilience Against Both Scrubbing and Spoofing Attacks
Huanming Shen (University of Electronic Science and Technology of China), Xiaojun Wan (Peking University)
CodeGenerationAdversarial AttackTransformerLarge Language ModelText
π― What it does: By introducing an equivalent texture key mechanism and the SEEK scheme, the robustness of LLM watermarking against text rewriting (scrubbing) and forgery (spoofing) attacks is enhanced.
π― What it does: A sample selection strategy called Early Cutting is proposed, which recalibrates and removes easily mislearned noisy samples (Mislabeled Easy Examples, MEEs) that are mislearned in the early stages of learning, improving the model's generalization performance on noisy labeled data.
Enhancing the Maximum Effective Window for Long-Term Time Series Forecasting
Jiahui Zhang (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
CodeTransformerTime Series
π― What it does: This paper proposes the Maximum Effective Window (MEW) metric and constructs two modules: the Information Bottleneck Filter (IBF) and the Hybrid Transformer-Mamba (HTM), to enhance the Transformer model's utilization of long-window time series. Based on this, the PIH model is designed to achieve a longer backtracking window and state-of-the-art (SOTA) prediction performance.
CodeTransformerLarge Language ModelReinforcement LearningVision Language ModelTextMultimodality
π― What it does: Design and implement the Self-Consistency Sampling (SCS) method, which uses truncated resampling and visual perturbations to generate consistency rewards, thereby enhancing the reliability and accuracy of reward-based reinforcement learning in multi-choice tasks of multimodal large language models.
Enhancing Training Data Attribution with Representational Optimization
Weiwei Sun (Carnegie Mellon University), Yiming Yang (Carnegie Mellon University)
CodeExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A scalable training data attribution method called AirRep is proposed, which measures the impact of training data on model predictions by learning representations aligned with tasks and models.
CodeRecognitionGenerationData-Centric LearningTransformerVision Language ModelImageTextMultimodality
π― What it does: A method for uncertainty-guided dropout decoding of visual tokens during the inference phase is proposed, which enhances the reliability of large visual-language models (LVLMs) by removing high-uncertainty tokens.
π― What it does: This paper proposes a prediction error-based unlabeled environment inference method called DynaInfer, which automatically assigns environment labels to trajectories in multi-environment dynamic system learning and trains a generalizable dynamic model.
π― What it does: A method called EnzyControl is proposed, which can achieve both functional site retention and substrate specificity control during the enzyme scaffold generation process.
π― What it does: Proposes the EPA framework, which achieves lossless and degradation-resistant video frame interpolation by utilizing the semantic perception features of visual foundation models and event-guided alignment.
π― What it does: A method is proposed to detect AI-generated images by estimating empirical uncertainty through weight perturbation, utilizing the prior knowledge of large-scale visual models.
π― What it does: Designed the Equilibrium Policy Generalization (EPG) framework for training pursuit-avoidance game strategies with zero-shot generalization performance under different graph structures.
Equivariance Everywhere All At Once: A Recipe for Graph Foundation Models
Ben Finkelshtein (University of Oxford), Ron Levie (Technion Israel Institute of Technology)
CodeClassificationGraph Neural NetworkGraph
π― What it does: A universal graph foundational model framework based on triple symmetry (node, label permutation equivalence, feature permutation invariance) is proposed, and its universal approximation capability on multi-set functions is proven.
Rohit Gandikota (Northeastern University), David Bau (Northeastern University)
CodeTransformerLarge Language ModelText
π― What it does: The paper proposes a method for achieving concept-level 'forgetting' through the self-classification ability of the language model itself, namely Erasing Language Memory (ELM).
Error Broadcast and Decorrelation as a Potential Artificial and Natural Learning Mechanism
Mete Erdogan (Koc University), Alper Tunga Erdogan
CodeOptimizationConvolutional Neural NetworkImage
π― What it does: A new neural network learning framework called Error Broadcast and Decorrelation (EBD) is proposed, which addresses the credit assignment problem by minimizing the correlation between output error and layer activations across layers.
ESCA: Contextualizing Embodied Agents via Scene-Graph Generation
Jiani Huang (University of Pennsylvania), Mayur Naik (University of Pennsylvania)
CodeRecognitionObject DetectionRobotic IntelligenceTransformerLarge Language ModelVision Language ModelContrastive LearningVideoTextMultimodality
π― What it does: The ESCA framework is proposed, utilizing the open-source CLIP model SGClip to generate scene graphs to enhance the perception and planning of multimodal large language models (MLLMs) in embodied agents.
Escaping saddle points without Lipschitz smoothness: the power of nonlinear preconditioning
Alexander Bodard (KU Leuven), Panagiotis Patrinos (KU Leuven)
CodeOptimization
π― What it does: This paper studies the convergence and saddle point escape properties of nonlinear preconditioned gradient descent in non-convex optimization, and proposes a new smoothness condition, proving its validity in key problems such as phase recovery and matrix decomposition.
π― What it does: Proposes the ESCORT framework, which combines particle evolution with SVGD, correlation projection, and temporal consistency to improve the belief representation of POMDP.
π― What it does: This paper proposes and implements an attention-aware inverse planning framework to infer cognitive biases from human behavior trajectories.
Themistoklis Haris (Boston University), Charalampos Tsourakakis (Boston University)
CodeGraph
π― What it does: This paper proposes two local algorithms that estimate the hitting time between any two points by performing a small number of random walks only near the target node.
CodeTransformerLarge Language ModelSupervised Fine-TuningTextSequentialBenchmark
π― What it does: The EvaLearn benchmark is proposed, which evaluates the learning ability and efficiency of LLMs in a sequential manner across six types of tasks with 182 sequences (each containing 7 questions, totaling 648 challenging questions) and provides five comprehensive learning metrics.
Evaluating LLM-contaminated Crowdsourcing Data Without Ground Truth
Yichi Zhang (Rutgers University), Yang Liu (University of California)
CodeTransformerLarge Language ModelReinforcement LearningText
π― What it does: This paper proposes a conditional evaluation mechanism that utilizes 'peer prediction' and low-cost signals generated by LLMs in the absence of true labels, aimed at identifying low-effort, LLM-assisted cheating crowd workers.
Swadesh Sistla (University of Washington), Max Kleiman-Weiner (University of Washington)
CodeOptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextBenchmarkChain-of-Thought
π― What it does: This paper studies the performance of large language models (LLMs) in open-source games, exploring their understanding of code, strategy generation, and evolutionary balance.
Evaluating Robustness of Monocular Depth Estimation with Procedural Scene Perturbations
John Nugent, Jia Deng (Princeton University)
CodeDepth EstimationPoint CloudBenchmark
π― What it does: Systematically evaluate the robustness of monocular depth estimation models under different environmental changes through programmatically generated 3D scenes and various perturbations (camera, objects, materials, lighting, etc.).
Jiacheng Fu (University of Science and Technology of China), Zhiwei Xiong (University of Science and Technology of China)
CodeDepth EstimationPoint Cloud
π― What it does: The first HDR structured light 3D measurement framework specifically designed for event cameras is proposed, achieving accurate 3D reconstruction in extremely high dynamic range scenes.
Every Rollout Counts: Optimal Resource Allocation for Efficient Test-Time Scaling
Xinglin Wang (Beijing Institute of Technology), Kan Li (Beijing Institute of Technology)
CodeOptimizationReinforcement LearningTabular
π― What it does: A theoretical optimal resource allocation method for search during testing based on a Bayesian probability model is proposed, and a direction-oriented resource allocation method (DORA) is implemented to achieve higher accuracy under different computational powers.
Shigui Li (South China University of Technology), Delu Zeng (South China University of Technology)
CodeGenerationOptimizationDiffusion modelImage
π― What it does: This paper proposes EVODiff, which optimizes the inference process of diffusion models using an information-theoretic framework, achieving reference-free conditional variance optimization by reducing conditional entropy.
π― What it does: This study investigates the 'Fitness Evaluation Bias' (FEB) present in Evolutionary Multi-View Classification (EMVC) and proposes the introduction of an Evolutionary Navigator (EN) and a knowledge distillation loss based on Wasserstein distance (ENL) to eliminate this bias, thereby enhancing the performance of EMVC.
CodeOptimizationTransformerReinforcement LearningVision Language ModelDiffusion modelMultimodality
π― What it does: The EvolvedGRPO framework is proposed, which enhances the reasoning ability of large visual language models through progressive instruction evolution.
Exact and Linear Convergence for Federated Learning under Arbitrary Client Participation is Attainable
Bicheng Ying (Google), Haibo Yang (Rochester Institute of Technology)
CodeOptimizationFederated LearningTabular
π― What it does: A Federated Learning algorithm named FOCUS is proposed, which can achieve precise linear convergence under the participation of any clients;
Boaz Lavon (Tel Aviv University), Lior Wolf (Tel Aviv University)
CodeGenerationAI Code AssistantTransformerLarge Language ModelReinforcement LearningAgentic AIText
π― What it does: This paper proposes a real-time reasoning method that uses execution signals during code generationβExecution-Guided Classifier-Free Guidance (EG-CFG). It dynamically guides the language model to generate executable code by generating and executing candidate code line by line.
Explaining Similarity in Vision-Language Encoders with Weighted Banzhaf Interactions
Hubert Baniecki (University of Warsaw), Przemyslaw Biecek (University of Warsaw)
CodeExplainability and InterpretabilityVision Language ModelImageMultimodality
π― What it does: A second-order interaction interpretation method based on weighted Banzhaf interaction (FIXLIP) is proposed to reveal the similarity prediction of visual-language encoders.
π― What it does: A novel neural-inspired CNN called Early Vision Networks (EVNets) is proposed, which combines subcortical processing and VOneBlock, demonstrating its ability to enhance robustness against visual perturbations and alignment with primate vision.
π― What it does: This study investigates the dynamic sparsity in Einsum expressions and proposes a hybrid algorithm that dynamically switches between dense and sparse representations based on the average density of the remaining tensors during execution.
Exploiting LLMs for Automatic Hypothesis Assessment via a Logit-Based Calibrated Prior
Yue Gong (Amazon Web Services), Raul Castro Fernandez (University of Chicago)
CodeTransformerLarge Language ModelTabular
π― What it does: Utilize large language models (LLM) to construct prior distributions for the correlation coefficients of variable pairs to achieve automatic hypothesis evaluation.
Exploiting the Asymmetric Uncertainty Structure of Pre-trained VLMs on the Unit Hypersphere
Li Ju (Uppsala University), Prashant Singh (Uppsala University)
CodeRetrievalRepresentation LearningVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: This paper proposes AsymVLM, a posterior probability adaptation framework that utilizes a pre-trained visual language model (VLM) to probabilistically embed text on the unit hypersphere, enabling cross-modal retrieval and uncertainty quantification.
Exploring Polyglot Harmony: On Multilingual Data Allocation for Large Language Models Pretraining
Ping Guo (ByteDance), Yin Zheng (ByteDance)
CodeOptimizationTransformerLarge Language ModelText
π― What it does: The CLIMB framework is proposed, which models the cross-lingual interaction-aware language ratio to predict the validation loss of multilingual pre-trained models and automatically find the optimal language allocation.
π― What it does: A systematic study of the phenomenon of performance degradation over training time in self-supervised learning for dense tasks is conducted, and a label-free evaluation metric DSE based on inter-class margin and effective dimension is proposed to predict and mitigate this degradation.
π― What it does: Proposed and implemented the Endpoint Conditional Stochastic Interpolation (ECSI) model, which unifies and expands the design space of the diffusion bridge model, enhances sampling efficiency and image quality, and addresses the issue of insufficient conditional diversity in one-to-many translations.
π― What it does: Exploring the optimal trade-off of multi-task learning through mode connectivity, the EXTRA method is proposed, optimizing the curve rather than the endpoints, using NURBS curves and sequentially aware objectives;
Extracting task-relevant preserved dynamics from contrastive aligned neural recordings
Yiqi Jiang (Stanford University), Mark Schnitzer (Stanford University)
CodeRepresentation LearningContrastive LearningMultimodalityTime Series
π― What it does: Proposes the CANDY framework, which uses contrastive learning and linear dynamical systems for joint training to align neural embeddings across sessions and extract task-relevant conservative dynamics;
Extragradient Method for $(L_0, L_1)$-Lipschitz Root-finding Problems
Sayantan Choudhury (Johns Hopkins University), Nicolas Loizou (Johns Hopkins University)
CodeOptimizationReinforcement Learning
π― What it does: This paper studies the Extragradient (EG) method under the Ξ±-symmetric (L, Lβ, Lβ)-Lipschitz condition, proposes an adaptive step size strategy, and provides convergence analysis for strongly monotone, monotone, and weak Minty problems, achieving linear or sublinear convergence rates.
π― What it does: The FACE framework is proposed, which maps the continuous embeddings of collaborative filtering models to discrete semantic tokens in the vocabulary of pre-trained LLMs, achieving semantic alignment through contrastive learning.
Dipkamal Bhusal (Rochester Institute of Technology), Nidhi Rastogi (Rochester Institute of Technology)
CodeOptimizationExplainability and InterpretabilityConvolutional Neural NetworkImage
π― What it does: A new conceptual explanation framework called FACE is proposed, which utilizes NMF and incorporates KL divergence regularization to ensure that the extracted concepts are consistent with the model predictions, thereby improving the reliability of the explanations.
π― What it does: A recommendation framework called PreferGrow based on discrete diffusion is proposed, which directly models the relative preference ratios between items by 'diluting' user preferences during the forward process (replacing liked items with negative samples) and gradually 'restoring' user preferences during the backward process.
π― What it does: A framework (FRG) is proposed to achieve controllable high-confidence fairness in representation learning, ensuring that the disparity in downstream tasks does not exceed a user-specified threshold Ξ΅, with a confidence level of 1-Ξ΄.
FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models
Zihao Fu (Chinese University of Hong Kong), Chris Russell (University of Oxford)
CodeGenerationDiffusion modelImageText
π― What it does: This paper proposes FairImagen, a post-processing debiasing method that eliminates social biases such as gender and race in text-to-image generation directly in the CLIP embedding space of Stable Diffusion through FairPCA projection, noise injection, and cross-attribute joint projection, without the need to retrain the model or modify prompts.
Ronen Gradwohl (University of Haifa), Moshe Tennenholtz (Technion - Israel Institute of Technology)
CodeClassificationRecommendation SystemOptimizationTabularFinance Related
π― What it does: This paper studies the impact of multiple institutions using fair algorithms on the overall fairness of the ecosystem in a competitive environment. It demonstrates that even if a single algorithm meets fairness standards, the overall ecosystem may still be unfair, and explores mechanisms by which fairness adjustments may lead to a decrease in overall fairness.
π― What it does: The FairNet framework is proposed, which utilizes an internal bias detector to dynamically trigger the LoRA module, making representation-level fairness corrections only for instances predicted to have potential bias, thereby improving the performance of disadvantaged groups while maintaining overall accuracy.
π― What it does: Proposes Faithful Dynamic Imitation Learning (FaithDaIL), which learns agents through dynamic regression minimization and unbiased expert imitation targets under real-time human intervention.
Kiljae Lee (Ohio State University), Yuan Zhang (Ohio State University)
CodeExplainability and InterpretabilityAdversarial AttackTabular
π― What it does: This paper proposes the Faithful Group Shapley Value (FGSV) and provides an algorithm that can directly and quickly approximate FGSV; it defends against group-level data valuation attacks using FGSV, addressing the vulnerability of traditional Group Shapley Value to shell company attacks.
FALCON: An ML Framework for Fully Automated Layout-Constrained Analog Circuit Design
Asal Mehradfar (University of Southern California), Salman Avestimehr (University of Southern California)
CodeOptimizationGraph Neural NetworkTabular
π― What it does: The FALCON framework is proposed to achieve full automation of analog circuit design from performance specifications to complete layout constraints.
FALCON: Fine-grained Activation Manipulation by Contrastive Orthogonal Unalignment for Large Language Model
Jinwei Hu (University of Liverpool), Xiaowei Huang (University of Liverpool)
CodeOptimizationSafty and PrivacyTransformerLarge Language ModelContrastive LearningText
π― What it does: A fine-grained activation manipulation method for large language models, FALCON, is proposed, which utilizes information-theoretic metrics to select parameters and combines contrastive learning with gradient orthogonal projection to achieve secure knowledge forgetting.
FALQON: Accelerating LoRA Fine-tuning with Low-Bit Floating-Point Arithmetic
Kanghyun Choi (Seoul National University), Jinho Lee (Seoul National University)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes an FP8 quantization-based LoRA fine-tuning framework called FALQON, which directly integrates the LoRA adapter into the FP8 quantized model backbone, redesigning the forward/backward computation paths to eliminate the quantization overhead caused by small-sized tensors in traditional LoRA fine-tuning, thereby significantly accelerating the fine-tuning of large language models.
FANS: A Flatness-Aware Network Structure for Generalization in Offline Reinforcement Learning
Da Wang (Shanxi University), Jiye Liang (Shanxi University)
CodeReinforcement LearningTabularBenchmark
π― What it does: This paper proposes the FANS (Flatness-Aware Network Structure) framework, which achieves a preference for flat minimization through structured network design in offline reinforcement learning, thereby enhancing policy generalization and robustness.
Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning
Yeongbin Seo (Yonsei University), Jinyoung Yeo (Yonsei University)
CodeGenerationOptimizationComputational EfficiencyLarge Language ModelSupervised Fine-TuningDiffusion modelText
π― What it does: This paper addresses the long decoding window (LDW) problem of diffusion language models in open text generation, proposing two methods: Convolutional Decoding (Conv) and Rule-based Negative Sample Fine-tuning (R2FT), which significantly enhance the model's fluency and consistency while maintaining high speed and supporting bidirectional generation.
Fast Computation and Optimization for Opinion-Based Quantities of Friedkin-Johnsen Model
Haoxin Sun (Fudan University), Zhongzhi Zhang (Fudan University)
CodeOptimizationComputational EfficiencyGraph
π― What it does: This paper proposes a fast algorithm based on partial root forest sampling to estimate and optimize public opinion-related quantities under the Friedkin-Johnsen model;
π― What it does: A projection-free over-damped Langevin sampling framework OLLA is proposed, which can simultaneously handle non-convex equality and inequality constraints.
π― What it does: Developed two high-order inference algorithms for discrete diffusion models (ΞΈ-RK-2 and ΞΈ-Trapezoidal), achieving more efficient and accurate sampling;
π― What it does: A new kernel method training framework, EigenPro 4, is proposed, which utilizes a delayed projection strategy to achieve large-scale model training while maintaining the advantages of kernel methods.
Fast-Slow Thinking GRPO for Large Vision-Language Model Reasoning
Wenyi Xiao (Zhejiang University), Leilei Gan (Zhejiang University)
CodeOptimizationComputational EfficiencyTransformerReinforcement LearningVision Language ModelTextMultimodality
π― What it does: We propose FAST-GRPO, a fast-slow thinking framework for large visual language models (LVLM) that dynamically adjusts the depth of reasoning to avoid overthinking.
π― What it does: FastDINOv2 is proposed, achieving efficient pre-training of DINOv2 through two-stage frequency curriculum learning and Gaussian noise patching, significantly improving convergence speed while maintaining robustness.
π― What it does: Designed and implemented a trainable, hardware-friendly video sparse attention mechanism (VSA) to accelerate the training and inference of video diffusion Transformers.
FastLongSpeech: Enhancing Large Speech-Language Models for Efficient Long-Speech Processing
Shoutao Guo (Chinese Academy of Sciences), Yang Feng (Chinese Academy of Sciences)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelBenchmarkAudio
π― What it does: The FastLongSpeech framework is proposed, which enables large-scale speech-language models (LSLM) to efficiently handle long speech through iterative fusion and dynamic compression training.