Page 32 - Paper Library - AiPaper

Specificity, length and luck drive gene rankings in association studies

Gene Specificity in Association StudiesRare Variant Burden TestingGene Prioritization for Quantitative TraitsUK Biobank Data AnalysisDifferences in Trait Biology

This study analyzes 209 quantitative trait association studies, revealing systematic differences in gene prioritization between GWAS and rare variant burden tests. It proposes criteria based on trait importance and specificity, highlighting their distinct impacts on trait biology

Convergent genome evolution shaped the emergence of terrestrial animals

Published:11/12/2025

Convergent Genome EvolutionGenomic Adaptations in Terrestrial AnimalsAnimal Terrestrialization EventsPatterns of Gene Gain and LossComparative Genomics across Animal Phyla

This study analyzes 154 genomes from 21 animal phyla to uncover the convergence and contingency in terrestrialization events, revealing unique gene patterns yet recurrent adaptive functions, crucial for life on land, while establishing a timeline for these transitions.

OnePiece: Bringing Context Engineering and Reasoning to Industrial Cascade Ranking System

Published:9/23/2025

Industrial Ranking SystemContext Engineering and ReasoningTransformer-based Recommendation ModelsMulti-Task Training ApproachUser Feedback Chain Supervision

OnePiece integrates context engineering and multistep reasoning into industrial ranking systems, enhancing existing Transformer models. Key innovations include structured context engineering and progressive multitask training, leading to significant performance improvements in

VGGT: Visual Geometry Grounded Transformer

Published:3/15/2025

Visual Geometry Grounded Transformer3D Attribute InferenceMulti-View Depth EstimationDense Point Cloud ReconstructionCamera Parameter Estimation

VGGT is a feedforward neural network that directly infers 3D scene attributes from one or multiple views, achieving stateoftheart results in various 3D tasks while enhancing downstream task performance.

GPR: Towards a Generative Pre-trained One-Model Paradigm for Large-Scale Advertising Recommendation

Published:11/13/2025

Generative Recommendation SystemsAdvertising Recommendation OptimizationUnified Generative Model FrameworkMulti-Stage Joint Training StrategyHeterogeneous Hierarchical Decoder

This paper presents GPR (Generative Pretrained Recommender), a novel framework that redefines advertising recommendation as an endtoend generative task, overcoming issues of objective misalignment and error propagation, while enhancing semantic alignment and modeling consisten

010

Sparse Meets Dense: Unified Generative Recommendations with Cascaded Sparse-Dense Representations

Published:3/4/2025

Sparse-Dense Recommendation ModelGenerative Recommendation SystemsCascaded Sparse-Dense RepresentationsUser Interaction Sequence ModelingOnline Recommendation System Optimization

The study introduces the COBRA framework, which integrates sparse semantic IDs and dense vectors through alternating generation. This endtoend training enhances dynamic optimization of representations, effectively capturing semantic and collaborative insights from useritem int

Fast Video Generation with Sliding Tile Attention

Published:2/7/2025

Sliding Tile Attention MechanismVideo Diffusion Generation ModelsEfficient Attention MechanismHunyuanVideoComputational Efficiency Optimization

The study introduces Sliding Tile Attention (STA) to reduce computational bottlenecks in video generation, achieving 58.79% Model FLOPs Utilization while decreasing latency to 501 seconds without quality loss, demonstrating significant efficiency improvements over existing method

Sparse VideoGen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity

Published:2/4/2025

Acceleration of Diffusion ModelsVideo Generation TransformerSpatial-Temporal SparsityEfficient Inference FrameworkDynamic Sparse Patterns

Sparse VideoGen (SVG) enhances video generation efficiency by leveraging the inherent sparsity of 3D attention, classifying attention heads into spatial and temporal types. It achieves up to 2.33x acceleration while maintaining generation quality, with opensource code available.

Rethinking the Evaluation for Conversational Recommendation in the Era of Large Language Models

Published:5/22/2023

LLM-based Conversational Recommender SystemsConversational Recommendation Evaluation MethodsInteractive Evaluation ApproachesApplication of ChatGPT in Recommendation SystemsLLM-based User Simulators

This paper reveals limitations in current evaluation methods for conversational recommender systems (CRSs) and proposes the iEvaLM approach using LLMbased user simulators, which shows significant improvements and emphasizes explainability in experiments on two public datasets.

Refining Text Generation for Realistic Conversational Recommendation via Direct Preference Optimization

Published:8/27/2025

Conversational Recommender SystemsDirect Preference OptimizationLLM Applications in RecommendationUser Preference ExtractionText Generation Refinement

This paper introduces an improved Conversational Recommender System method using Large Language Models to generate dialogue summaries and recommendations, capturing both explicit and implicit user preferences. Direct Preference Optimization (DPO) is employed to ensure rich conten

Sparse VideoGen2: Accelerate Video Generation with Sparse Attention via Semantic-Aware Permutation

Published:5/25/2025

Sparse Attention Video GenerationDiffusion TransformersSemantic-Aware PermutationTraining-Free FrameworkVideo Generation Acceleration

This paper presents SVG2, a trainingfree framework that enhances critical token identification accuracy through semanticaware permutation, reducing computation waste and addressing efficiency bottlenecks in sparse attention for video generation, achieving up to 2.30x accelerati

Galaxy: A Cognition-Centered Framework for Proactive, Privacy-Preserving, and Self-Evolving LLM Agents

Published:8/6/2025

Cognition-Centered Intelligent Personal Assistant FrameworkPrivacy-Preserving Self-Evolving LLM AgentsProactive Behavior in Intelligent Personal AssistantsUnified Cognitive Architecture and System DesignMultidimensional Capability Generation

The paper presents , a cognitioncentered framework for proactive, selfevolving, and privacypreserving LLM agents, integrating cognitive architecture with system design. Experimental results demonstrate its superior performance across various benchmarks.

Modeling shopper interest broadness with entropy-driven dialogue policy in the context of arbitrarily large product catalogs

Published:9/8/2025

Conversational Recommender SystemsLLM-based Recommendation SystemsUser Interest Broadness ModelingEntropy-Driven Dialogue PolicyDynamic Exploration and Exploitation in Recommendations

This study introduces a conversational recommender system strategy that balances exploration and exploitation by modeling user interest breadth through retrieval score entropy, dynamically adjusting dialogue policies based on query specificity to enhance user experience.

S$^2$Edit: Text-Guided Image Editing with Precise Semantic and Spatial Control

Published:7/7/2025

Text-Guided Image EditingPrecise Semantic and Spatial ControlFine-Tuning of Diffusion ModelsPersonalized Image EditingIdentity Information Embedding

^2

Edit is a novel textguided image editing method using a pretrained diffusion model, embedding identity into learnable tokens while ensuring semantic disentanglement and spatial focus through object masks for precise localized editing.

CCL24-Eval 任办系统报告 : 基于大型语言模型的中文空间语义评测

Chinese Spatial Semantic Evaluation Based on Large Language ModeEvaluation of Spatial Semantic Understanding AbilityImpact of Prompting Strategies on Task PerformanceEvaluation of ERNIE-4 ModelEntity and Role Recognition

This study evaluates large language models' spatial semantic understanding through tasks like entity and role recognition. Three prompt strategies were tested, revealing ERNIE4 performed best with a 1shot vanilla prompt, ranking sixth with an accuracy of 56.20%.

System Report for CCL24-Eval Task 3: Chinese Spatial Semantic Understanding Based on In-Context Learning and Chain of Thought Strategy

Chinese Spatial Semantic UnderstandingChain of Thought StrategyIn-Context LearningSpatial Information Entity RecognitionSpatial Information Anomaly Detection

This report details the team's methods for the SpaCE2024 evaluation, utilizing incontext learning and chain of thought strategies, achieving top accuracy in five tasks with an overall accuracy of 0.6024, securing first place.

Virus-inspired lipopeptide-derived nucleic acid delivery to cartilage for osteoarthritis therapy

Published:10/16/2025

Virus-Inspired Lipopeptide NanoparticlesOptimized Gene Delivery VectorsCartilage-Targeted Gene TherapyROS-Responsive Nano-in-Gel SystemOsteoarthritis Treatment

This study developed virusinspired lipopeptide nanoparticles (VPN) for RNA delivery in osteoarthritis therapy. The optimized VPN2 showed a 2.5fold increase in transfection efficiency and effectively alleviated cartilage degeneration in mice, highlighting its potential for cart

Identify Bitter Peptides by Using Deep Representation Learning Features

Published:7/17/2022

Bitter Peptide Identification using Deep Representation LearningSequence Embedding TechniquesBidirectional LSTM (BiLSTM)Light Gradient Boosting Machine (LGBM)Bitter Peptide Prediction Method

This study introduces the iBitterDRLF method, employing deep learning techniques to enhance the identification of bitter peptides, significantly improving palatability outcomes in related products.

Identification and prediction of milk-derived bitter taste peptides based on peptidomics technology and machine learning method

Published:9/1/2023

Bitter Peptide Prediction Model for Dairy ProductsPeptidomics-based Bitter Peptide ScreeningApplication of LightGBM in Bitter Peptide IdentificationBitter Receptor Activity Validation

This study developed a screening workflow for bitter taste peptides using peptidomics and machine learning, achieving 90.3% accuracy with the novel CPMBP model. Among 724 distinct peptides, 180 potential bitter peptides were identified, with three confirmed to activate the human

iBitter-Stack: A Multi-Representation Ensemble Learning Model for Accurate Bitter Peptide Identification

Published:9/19/2025

Bitter Peptide RecognitionMulti-Representation Ensemble Learning ModelBioinformatics Methods

The iBitterStack framework enhances bitter peptide identification accuracy by integrating Protein Language Model embeddings and handcrafted physicochemical features, utilizing various machine learning classifiers, achieving 96.09% accuracy in independent tests.

621 - 640 / 980

Papers