Daily Arxiv

전 세계에서 발간되는 인공지능 관련 논문을 정리하는 페이지 입니다.
본 페이지는 Google Gemini를 활용해 요약 정리하며, 비영리로 운영 됩니다.
논문에 대한 저작권은 저자 및 해당 기관에 있으며, 공유 시 출처만 명기하면 됩니다.

Batch Acquisition Function Evaluations and Decouple Optimizer Updates for Faster Bayesian Optimization

Physics-Informed Neural Networks for Nonlinear Output Regulation

Unlocking the Forgery Detection Potential of Vanilla MLLMs: A Novel Training-Free Pipeline

Generalized Denoising Diffusion Codebook Models (gDDCM): Tokenizing images using a pre-trained diffusion model

Rethinking Saliency Maps: A Cognitive Human Aligned Taxonomy and Evaluation Framework for Explanations

Dimension vs. Precision: A Comparative Analysis of Autoencoders and Quantization for Efficient Vector Retrieval on BEIR SciFact

UNSEEN: Enhancing Dataset Pruning from a Generalization Perspective

PFAvatar: Pose-Fusion 3D Personalized Avatar Reconstruction from Real-World Outfit-of-the-Day Photos

Optimal Look-back Horizon for Time Series Forecasting in Federated Learning

Fine-Grained Representation for Lane Topology Reasoning

VULPO: Context-Aware Vulnerability Detection via On-Policy LLM Optimization

A Meta-Heuristic Load Balancer for Cloud Computing Systems

Embedding Explainable AI in NHS Clinical Safety: The Explainability-Enabled Clinical Safety Framework (ECSF)

MedBuild AI: An Agent-Based Hybrid Intelligence Framework for Reshaping Agency in Healthcare Infrastructure Planning through Generative Design for Medical Architecture

MOON Embedding: Multimodal Representation Learning for E-commerce Search Advertising

Efficient Reinforcement Learning for Zero-Shot Coordination in Evolving Games

Who Gets the Reward, Who Gets the Blame? Evaluation-Aligned Training Signals for Multi-LLM Agents

A Style is Worth One Code: Unlocking Code-to-Style Image Generation with Discrete Style Space

LocalBench: Benchmarking LLMs on County-Level Local Knowledge and Reasoning

LoopTool: Closing the Data-Training Loop for Robust LLM Tool Calls

Tele-LLM-Hub: Building Context-Aware Multi-Agent LLM Systems for Telecom Networks

FGM-HD: Boosting Generation Diversity of Fractal Generative Models through Hausdorff Dimension Induction

Energy Consumption of Dataframe Libraries for End-to-End Deep Learning Pipelines:A Comparative Analysis

Resilient by Design -- Active Inference for Distributed Continuum Intelligence

Resource Efficient Sleep Staging via Multi-Level Masking and Prompt Learning

Physics-Informed Neural Networks for Real-Time Gas Crossover Prediction in PEM Electrolyzers: First Application with Multi-Membrane Validation

MusRec: Zero-Shot Text-to-Music Editing via Rectified Flow and Diffusion Transformers

Leveraging LLM-based agents for social science research: insights from citation network simulations

Deploying Rapid Damage Assessments from sUAS Imagery for Disaster Response

DRIP: Defending Prompt Injection via Token-wise Representation Editing and Residual Instruction Fusion

Clone Deterministic 3D Worlds

Posterior Sampling by Combining Diffusion Models with Annealed Langevin Dynamics

MMEdge: Accelerating On-device Multimodal Inference via Pipelined Sensing and Encoding

FinVet: A Collaborative Framework of RAG and External Fact-Checking Agents for Financial Misinformation Detection

From Reasoning LLMs to BERT: A Two-Stage Distillation Framework for Search Relevance

PromptGuard at BLP-2025 Task 1: A Few-Shot Classification Framework Using Majority Voting and Keyword Similarity for Bengali Hate Speech Detection

A More Realistic Evaluation of Cross-Frequency Transfer Learning and Foundation Forecasting Models

Towards Sharper Object Boundaries in Self-Supervised Depth Estimation

Patent Language Model Pretraining with ModernBERT

Retrosynthesis Planning via Worst-path Policy Optimisation in Tree-structured MDPs

Efficient Decoding Methods for Language Models on Encrypted Data

Spectral Masking and Interpolation Attack (SMIA): A Black-box Adversarial Attack against Voice Authentication and Anti-Spoofing Systems

DepthVision: Enabling Robust Vision-Language Models with GAN-Based LiDAR-to-RGB Synthesis for Autonomous Driving

Automatic Differentiation of Agent-Based Models

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

Skill-Aligned Fairness in Multi-Agent Learning for Collaboration in Healthcare

Spatial Policy: Guiding Visuomotor Robotic Manipulation with Spatial-Aware Modeling and Reasoning

RynnEC: Bringing MLLMs into Embodied World

LENS: Learning to Segment Anything with Unified Reinforced Reasoning

MOON: Generative MLLM-based Multimodal Representation Learning for E-commerce Product Understanding

DIVER: A Multi-Stage Approach for Reasoning-intensive Information Retrieval

Explaining Similarity in Vision-Language Encoders with Weighted Banzhaf Interactions

Dynamic User-controllable Privacy-preserving Few-shot Sensing Framework

LLMDistill4Ads: Using Cross-Encoders to Distill from LLM Signals for Advertiser Keyphrase Recommendations

SlotMatch: Distilling Object-Centric Representations for Unsupervised Video Segmentation

GAIS: Frame-Level Gated Audio-Visual Integration with Semantic Variance-Scaled Perturbation for Text-Video Retrieval

GMAT: Grounded Multi-Agent Clinical Description Generation for Text Encoder in Vision-Language MIL for Whole Slide Image Classification

AgentArmor: Enforcing Program Analysis on Agent Runtime Trace to Defend Against Prompt Injection

Cross-Domain Few-Shot Learning with Coalescent Projections and Latent Space Reservation

FastDINOv2: Frequency Based Curriculum Learning Improves Robustness and Training Speed

OptScale: Probabilistic Optimality for Inference-time Scaling

EvoLM: In Search of Lost Language Model Training Dynamics

Synthetic Geology: Structural Geology Meets Deep Learning

Foundation Models in Medical Imaging: A Review and Outlook

FairDICE: Fairness-Driven Offline Multi-Objective Reinforcement Learning

OG-VLA: Orthographic Image Generation for 3D-Aware Vision-Language Action Model

Scaling Textual Gradients via Sampling-Based Momentum

MCTSr-Zero: Self-Reflective Psychological Counseling Dialogues Generation via Principles and Adaptive Exploration

In-context Language Learning for Endangered Languages in Speech Recognition

SpecEdge: Scalable Edge-Assisted Serving Framework for Interactive LLMs

Preference Learning with Lie Detectors can Induce Honesty or Evasion

Anti-adversarial Learning: Desensitizing Prompts for Large Language Models

AdCare-VLM: Towards a Unified and Pre-aligned Latent Representation for Healthcare Video Understanding

Non-Uniform Class-Wise Coreset Selection for Vision Model Fine-tuning

ACoRN: Noise-Robust Abstractive Compression in Retrieval-Augmented Language Models

A Survey of Cross-domain Graph Learning: Progress and Future Directions

Availability-aware Sensor Fusion via Unified Canonical Space

ElementaryNet: A Non-Strategic Neural Network for Predicting Human Behavior in Normal-Form Games

1-Lipschitz Network Initialization for Certifiably Robust Classification Applications: A Decay Problem

Personalized Image Generation for Recommendations Beyond Catalogs

IPAD: Inverse Prompt for AI Detection - A Robust and Interpretable LLM-Generated Text Detector

MoM: Linear Sequence Modeling with Mixture-of-Memories

Rethinking Token-wise Feature Caching: Accelerating Diffusion Transformers with Dual Feature Caching

Automated Materials Discovery Platform Realized: Scanning Probe Microscopy of Combinatorial Libraries

Higher-Order Transformers With Kronecker-Structured Attention

Optimizing Federated Learning by Entropy-Based Client Selection

VisAidMath: Benchmarking Visual-Aided Mathematical Reasoning

Can Machines Think Like Humans? A Behavioral Evaluation of LLM Agents in Dictator Games

Generating Streamlining Constraints with Large Language Models

The Energy Cost of Artificial Intelligence Lifecycle in Communication Networks

KWT-Tiny: RISC-V Accelerated, Embedded Keyword Spotting Transformer

Benchmark on Drug Target Interaction Modeling from a Drug Structure Perspective

Virtual Human Generative Model: Masked Modeling Approach for Learning Human Characteristics

NAIST Academic Travelogue Dataset

SemCo: Toward Semantic Coherent Visual Relationship Forecasting

Multi-Agent Deep Research: Training Multi-Agent Systems with M-GRPO

Requirements for Aligned, Dynamic Resolution of Conflicts in Operational Constraints

CTRL-ALT-DECEIT: Sabotage Evaluations for Automated AI R&D

VSPO: Validating Semantic Pitfalls in Ontology via LLM-Based CQ Generation

Are We Asking the Right Questions? On Ambiguity in Natural Language Queries for Tabular Data Analysis

Task Memory Engine (TME): Enhancing State Awareness for Multi-Step LLM Agent Tasks

Created by

Haebom

저자

Ye Ye

개요

대규모 언어 모델(LLM)이 다단계 작업을 위한 자율 에이전트로 점점 더 많이 사용되고 있지만, 기존 프레임워크는 선형 프롬프트 연결이나 얕은 메모리 버퍼에 의존하여 작업 상태에 대한 구조적 이해를 유지하는 데 실패하는 경우가 많습니다. 이로 인해 취약한 성능, 빈번한 환각, 그리고 긴 범위의 일관성 저하가 발생합니다. 본 연구에서는 계층적 작업 메모리 트리(TMT)를 사용하여 작업 실행을 추적하는 경량의 구조적 메모리 모듈인 작업 메모리 엔진(TME)을 제안합니다. 트리의 각 노드는 작업 단계에 해당하며, 관련 입력, 출력, 상태 및 하위 작업 관계를 저장합니다. 활성 노드 경로를 기반으로 LLM 프롬프트를 동적으로 생성하는 프롬프트 합성 방법을 도입하여 실행 일관성과 상황적 근거를 크게 향상시킵니다. 다단계 에이전트 작업에 대한 사례 연구 및 비교 실험을 통해 TME가 최소한의 구현 오버헤드로 작업 완료 정확도를 높이고 해석 가능한 동작을 향상시킨다는 것을 보여줍니다. TME의 전체 구현은 https://github.com/biubiutomato/TME-Agent 에서 확인할 수 있습니다.

GitHub - biubiutomato/TME-Agent: TME: Structured memory engine for LLM agents to plan, rollback, and reason across multi-step tasks. DAG upgrade in progress.

TME: Structured memory engine for LLM agents to plan, rollback, and reason across multi-step tasks. DAG upgrade in progress. - biubiutomato/TME-Agent

시사점, 한계점

•

시사점:

◦

다단계 작업에서 LLM 에이전트의 성능과 해석 가능성을 향상시키는 경량의 구조적 메모리 모듈인 TME를 제시합니다.

◦

TMT를 이용한 계층적 작업 관리를 통해 작업 실행의 일관성과 상황적 근거를 향상시킵니다.

◦

최소한의 구현 오버헤드로 작업 완료 정확도를 높이고, 더욱 해석 가능한 에이전트 동작을 가능하게 합니다.

◦

공개된 GitHub 저장소를 통해 재현성과 확장성을 제공합니다.

•

한계점:

◦

제시된 TME의 성능이 다양한 종류의 다단계 작업과 LLM에 대해 얼마나 일반화될 수 있는지에 대한 추가적인 연구가 필요합니다.

◦

TMT의 계층적 구조 설계 및 최적화에 대한 추가적인 연구가 필요할 수 있습니다.

◦

매우 복잡하거나 장기간 실행되는 작업에 대한 TME의 확장성에 대한 추가적인 검증이 필요합니다.

◦

특정 작업에 대한 TMT의 구조 설계가 전적으로 사용자에게 의존하는 부분에 대한 개선이 필요할 수 있습니다.

Made with Slashpage