Daily Arxiv

世界中で発行される人工知能関連の論文をまとめるページです。
このページはGoogle Geminiを活用して要約し、非営利で運営しています。
論文の著作権は著者および関連機関にあり、共有する際は出典を明記してください。

Watermarking and Anomaly Detection in Machine Learning Models for LORA RF Fingerprinting

Vulnerable Agent Identification in Large-Scale Multi-Agent Reinforcement Learning

Sea-ing Through Scattered Rays: Revisiting the Image Formation Model for Realistic Underwater Image Generation

DPANet: Dual Pyramid Attention Network for Multivariate Time Series Forecasting

MeanFlowSE: one-step generative speech enhancement via conditional mean flow

Empathy-R1: A Chain-of-Empathy and Reinforcement Learning Framework for Long-Form Mental Health Support

Threat Modeling for Enhancing Security of IoT Audio Classification Devices under a Secure Protocols Framework

AToken: A Unified Tokenizer for Vision

TGPO: Tree-Guided Preference Optimization for Robust Web Agent Reinforcement Learning

Comprehensive Evaluation of CNN-Based Audio Tagging Models on Resource-Constrained Devices

MapAnything: Universal Feed-Forward Metric 3D Reconstruction

Improving Anomalous Sound Detection with Attribute-aware Representation from Domain-adaptive Pre-training

Hardness, Structural Knowledge, and Opportunity: An Analytical Framework for Modular Performance Modeling

Benchmark of stylistic variation in LLM-generated texts

SWE-Effi: Re-Evaluating Software AI Agent System Effectiveness Under Resource Constraints

Structure Matters: Brain Graph Augmentation via Learnable Edge Masking for Data-efficient Psychiatric Diagnosis

DischargeSim: A Simulation Benchmark for Educational Doctor-Patient Communication at Discharge

Riemannian Batch Normalization: A Gyro Approach

On the Security of Tool-Invocation Prompts for LLM-Based Agentic Systems: An Empirical Risk Assessment

MIDOG 2025: Mitotic Figure Detection with Attention-Guided False Positive Correction

Do Retrieval Augmented Language Models Know When They Don't Know?

LongCat-Flash Technical Report

MedCOD: Enhancing English-to-Spanish Medical Translation of Large Language Models Using Enriched Chain-of-Dictionary Framework

Middo: Model-Informed Dynamic Data Optimization for Enhanced LLM Fine-Tuning via Closed-Loop Learning

PVPO: Pre-Estimated Value-Based Policy Optimization for Agentic Reasoning

CORE-RAG: Lossless Compression for Retrieval-Augmented LLMs via Reinforcement Learning

OpenWHO: A Document-Level Parallel Corpus for Health Translation in Low-Resource Languages

Subjective Behaviors and Preferences in LLM: Language of Browsing

Using Natural Language for Human-Robot Collaboration in the Real World

RegionMed-CLIP: A Region-Aware Multimodal Contrastive Learning Pre-trained Model for Medical Image Understanding

Causal2Vec: Improving Decoder-only LLMs as Versatile Embedding Models

VLA-Mark: A cross modal watermark for large vision-language alignment model

Deformable Dynamic Convolution for Accurate yet Efficient Spatio-Temporal Traffic Prediction

Deep Reinforcement Learning with Gradient Eligibility Traces

Generating Moving 3D Soundscapes with Latent Diffusion Models

Capturing Polysemanticity with PRISM: A Multi-Concept Feature Description Framework

Discrete Diffusion in Large Language and Multimodal Models: A Survey

DualEdit: Dual Editing for Knowledge Updating in Vision-Language Models

Algorithmic Fairness: Not a Purely Technical but Socio-Technical Property

OptiScene: LLM-driven Indoor Scene Layout Generation via Scaled Human-aligned Data Synthesis and Multi-Stage Preference Optimization

Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward

AS-ASR: A Lightweight Framework for Aphasia-Specific Automatic Speech Recognition

LLMs Can Compensate for Deficiencies in Visual Representations

Spatial Understanding from Videos: Structured Prompts Meet Simulation Data

Emergent Abilities of Large Language Models under Continued Pretraining for Language Adaptation

Cross-Attention Speculative Decoding

Beyond Linear Steering: Unified Multi-Attribute Control for Language Models

Noise-Robustness Through Noise: Asymmetric LoRA Adaption with Poisoning Expert

SEMMA: A Semantic Aware Knowledge Graph Foundation Model

AmpleHate: Amplifying the Attention for Versatile Implicit Hate Detection

Fairness-in-the-Workflow: How Machine Learning Practitioners at Big Tech Companies Approach Fairness in Recommender Systems

GRE Suite: Geo-localization Inference via Fine-Tuned Vision-Language Models and Enhanced Reasoning Chains

A Survey of Large Language Models for Data Challenges in Graphs

CLEAR: A Clinically-Grounded Tabular Framework for Radiology Report Evaluation

Creative Preference Optimization

MUG-Eval: A Proxy Evaluation Framework for Multilingual Generation Capabilities in Any Language

Search and Refine During Think: Facilitating Knowledge Refinement for Improved Retrieval-Augmented Reasoning

Space Group Equivariant Crystal Diffusion

Schreier-Coset Graph Propagation

Examining Deployment and Refinement of the VIOLA-AI Intracranial Hemorrhage Model Using an Interactive NeoMedSys Platform

StreamBridge: Turning Your Offline Video Large Language Model into a Proactive Streaming Assistant

ConCISE: Confidence-guided Compression in Step-by-step Efficient Reasoning

AttentionDrop: A Novel Regularization Method for Transformer Models

MigGPT: Harnessing Large Language Models for Automated Migration of Out-of-Tree Linux Kernel Patches Across Versions

Hybrid Temporal Differential Consistency Autoencoder for Efficient and Sustainable Anomaly Detection in Cyber-Physical Systems

Who is Responsible When AI Fails? Mapping Causes, Entities, and Consequences of AI Privacy and Ethical Incidents

No Black Box Anymore: Demystifying Clinical Predictive Modeling with Temporal-Feature Cross Attention Mechanism

Negotiative Alignment: Embracing Disagreement to Achieve Fairer Outcomes - Insights from Urban Studies

MT-RewardTree: A Comprehensive Framework for Advancing LLM-Based Machine Translation via Reward Modeling

Pruning the Paradox: How CLIP's Most Informative Heads Enhance Performance While Amplifying Bias

KatFishNet: Detecting LLM-Generated Korean Text through Linguistic Feature Analysis

SuPreME: A Supervised Pre-training Framework for Multimodal ECG Representation Learning

Sparsity May Be All You Need: Sparse Random Parameter Adaptation

Neural Networks for Learnable and Scalable Influence Estimation of Instruction Fine-Tuning Data

"It Felt Like I Was Left in the Dark": Exploring Information Needs and Design Opportunities for Family Caregivers of Older Adult Patients in Critical Care Settings

Gradient Alignment in Physics-informed Neural Networks: A Second-Order Optimization Perspective

Advances in Multimodal Adaptation and Generalization: From Traditional Approaches to Foundation Models

A Layered Multi-Expert Framework for Long-Context Mental Health Assessments

Efficient Real-time Refinement of Language Model Text Generation

FLOAT: Generative Motion Latent Flow Matching for Audio-driven Talking Portrait

Dynamic Neural Curiosity Enhances Learning Flexibility for Autonomous Goal Discovery

Bayesian Concept Bottleneck Models with LLM Priors

G2D2: Gradient-Guided Discrete Diffusion for Inverse Problem Solving

SeCodePLT: A Unified Platform for Evaluating the Security of Code GenAI

DiRW: Path-Aware Digraph Learning for Heterophily

Towards Interactive and Learnable Cooperative Driving Automation: a Large Language Model-Driven Decision-Making Framework

DynamicNER: A Dynamic, Multilingual, and Fine-Grained Dataset for LLM-based Named Entity Recognition

CrackSCF: Lightweight Cascaded Fusion Network for Robust and Efficient Structural Crack Segmentation

ConfReady: A RAG based Assistant and Dataset for Conference Checklist Responses

FOVAL: Calibration-Free and Subject-Invariant Fixation Depth Estimation Across Diverse Eye-Tracking Datasets

Assessing invariance to affine transformations in image quality metrics

The Great AI Witch Hunt: Reviewers Perception and (Mis)Conception of Generative AI in Research Writing

Database-Augmented Query Representation for Information Retrieval

Two Is Better Than One: Aligned Representation Pairs for Anomaly Detection

BBScoreV2: Learning Time-Evolution and Latent Alignment from Stochastic Representation

Beyond Pixels: Enhancing LIME with Hierarchical Features and Segmentation Foundation Models

Spatio-Temporal Anomaly Detection with Graph Networks for Data Quality Monitoring of the Hadron Calorimeter

Understanding AI Evaluation Patterns: How Different GPT Models Assess Vision-Language Descriptions

Online Robust Planning under Model Uncertainty: A Sample-Based Approach

HiPhO: How Far Are (M)LLMs from Humans in the Latest High School Physics Olympiad Benchmark?

Middo: Model-Informed Dynamic Data Optimization for Enhanced LLM Fine-Tuning via Closed-Loop Learning

Created by

Haebom

作者

Zinan Tang, Xin Gao, Qizhi Pei, Zhuoshi Pan, Mengzhang Cai, Jiang Wu, Conghui He, Lijun Wu

概要

本稿では、大規模言語モデル（LLM）の地図微調整（SFT）のための新しいフレームワークであるMiddoを提案します。 Middoは、既存の静的データセット管理スキームの限界を克服するために、モデルのパフォーマンスの変化に応じて自ら進化する動的データ最適化システムを構築しています。 3つの軸（損失パターン、埋め込みクラスターダイナミクス、自己整合スコア）に基づいてモデルのパフォーマンスを診断し、不適切なサンプルを識別して意味論的な整合性を維持しながら、教育的に貴重なトレーニングデータに変換します。これらの動的学習原理により、モデルのパフォーマンスが向上するにつれて、データ最適化プロセスも継続的に進化します。いくつかのベンチマーク実験の結果、Middoは従来のデータセットサイズを維持しながら平均7.15％の精度向上を達成し、シードデータの品質を向上させ、LLMのパフォーマンスを向上させることを示しています。これは、データとモデルの動的人間-AI共振化による持続可能なLLMトレーニングのための新しいパラダイムを提示します。コードとデータセットは公開されています。

Takeaways、Limitations

•

Takeaways：

◦

モデルのパフォーマンスが向上するにつれて、データセットを動的に最適化する新しいフレームワークを提示します。

◦

従来のSFT方式のLimitationsである静的なデータセット管理の問題を解決。

◦

平均7.15％の精度向上という実験結果により性能改善効果を実証。

◦

データとモデルの共振化による持続可能なLLMトレーニングパラダイムの提示

◦

コードとデータセットの開示による研究の再現性と拡張性の確保

•

Limitations：

◦

現時点では実験結果のみが提示されており、フレームワークの詳細な実装とアルゴリズムの詳細な説明は不足しています。

◦

さまざまなLLMとデータセットの一般化の可能性に関する追加の検証が必要です。

◦

自己整合スコアの定義と測定方法の明確な説明の欠如。

◦

モデルのパフォーマンスの向上がデータセットの定性的な向上によるものか、他の要因によるものなのかを明確に分析する必要があります。

Made with Slashpage