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, 요약본 공유 시 출처만 명기하면 됩니다.
This service is supported by Google Gemini.

CLOVER: Closed-Loop Value Estimation and Ranking for End-to-End Autonomous Driving Planning

Logging Policy Design for Off-Policy Evaluation

TFGN: Task-Free, Replay-Free Continual Pre-Training Without Catastrophic Forgetting at LLM Scale

Unlocking Complex Visual Generation via Closed-Loop Verified Reasoning

Do Coding Agents Understand Least-Privilege Authorization?

Does RAG Know When Retrieval Is Wrong? Diagnosing Context Compliance under Knowledge Conflict

Agentic Recommender System with Hierarchical Belief-State Memory

Beyond Binary: Reframing GUI Critique as Continuous Semantic Alignment

ICED: Concept-level Machine Unlearning via Interpretable Concept Decomposition

Active Learners as Efficient PRP Rerankers

PanoWorld: Towards Spatial Supersensing in 360$^\circ$ Panorama World

Protocol-Driven Development: Governing Generated Software Through Invariants and Continuous Evidence

Controllable Quantum Memory Capacity in Quantum Reservoir Networks with Tunable partial-SWAPs

ODRPO: Ordinal Decompositions of Discrete Rewards for Robust Policy Optimization

Representing Higher-Order Networks: A Survey of Graph-Based Frameworks

Beyond GRPO and On-Policy Distillation: An Empirical Sparse-to-Dense Reward Principle for Language-Model Post-Training

Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models

SURGE: Surrogate Gradient Adaptation in Binary Neural Networks

BEACON: A Multimodal Dataset for Learning Behavioral Fingerprints from Gameplay Data

HYPERPOSE: Hyperbolic Kinematic Phase-Space Attention for 3D Human Pose Estimation

Swarm Skills: A Portable, Self-Evolving Multi-Agent System Specification for Coordination Engineering

Key-Value Means: Transformers with Expandable Block-Recurrent Compressed Memory

Sparsity Moves Computation: How FFN Architecture Reshapes Attention in Small Transformers

ShadowMerge: A Novel Poisoning Attack on Graph-Based Agent Memory via Relation-Channel Conflicts

Fitting Is Not Enough: Smoothness in Extremely Quantized LLMs

AIPO: Learning to Reason from Active Interaction

When Language Overwrites Vision: Over-Alignment and Geometric Debiasing in Vision-Language Models

PerFlow: Physics-Embedded Rectified Flow for Efficient Reconstruction and Uncertainty Quantification of Spatiotemporal Dynamics

BFORE: Butterfly-Firefly Optimized Retinex Enhancement for Low-Light Image Quality Improvement

Trojan Hippo: Weaponizing Agent Memory for Data Exfiltration

Skills as Verifiable Artifacts: A Trust Schema and a Biconditional Correctness Criterion for Human-in-the-Loop Agent Runtimes

Graph Construction and Matching for Imperative Programs using Neural and Structural Methods

CAP: Controllable Alignment Prompting for Unlearning in LLMs

Region-Grounded Report Generation for 3D Medical Imaging: A Fine-Grained Dataset and Graph-Enhanced Framework

DPrivBench: Benchmarking LLMs' Reasoning for Differential Privacy

Don't Retrieve, Navigate: Distilling Enterprise Knowledge into Navigable Agent Skills for QA and RAG

Hardware Utilization and Inference Performance of Edge Object Detection Under Fault Injection

KV Cache Offloading for Context-Intensive Tasks

DMax: Aggressive Parallel Decoding for dLLMs

Convergent Representations of Linguistic Constructions in Human and Artificial Neural Systems

Large Language Models as Optimization Controllers: Adaptive Continuation for SIMP Topology Optimization

IndicSafe: A Benchmark for Evaluating Multilingual LLM Safety in South Asia

FormulaCode: Evaluating Agentic Optimization on Large Codebases

Topology-Preserving Polygon Augmentation for Segmentation in Structured Visual Domains

A Unified View of Score-Based and Drifting Models

Stock Market Prediction Using Node Transformer Architecture Integrated with BERT Sentiment Analysis

Benchmark of Benchmarks: Unpacking Influence and Code Repository Quality in LLM Safety Benchmarks

Flowette: Flow Matching with Graphette Priors for Graph Generation

Brain-OF: An Omnifunctional Foundation Model for fMRI, EEG and MEG

Long Range Frequency Tuning for QML

Golden Layers and Where to Find Them: Improved Knowledge Editing for Large Language Models Via Layer Gradient Analysis

OmniVL-Guard: Towards Unified Vision-Language Forgery Detection and Grounding via Balanced RL

Improve Large Language Model Systems with User Logs

Antidistillation Fingerprinting

Decouple Searching from Training: Scaling Data Mixing via Model Merging for Large Language Model Pre-training

ExplainerPFN: Towards tabular foundation models for model-free zero-shot feature importance estimations

"Unlimited Realm of Exploration and Experimentation": Methods and Motivations of AI-Generated Sexual Content Creators

Structure-BiEval: A Self-Supervised, Dual-Track Framework for Decoupling Structure and Content in LLM Evaluation for Web Information Systems

SafeGPT: Preventing Data Leakage and Unethical Outputs in Enterprise LLM Use

LASER: Language Model Regression for Semi-Structured Workflow Resource and Runtime Estimation

EMFusion: An Uncertainty-Aware Conditional Diffusion Framework for Frequency-Selective EMF Forecasting in Wireless Networks

Drawback of Enforcing Equivariance and its Compensation via the Lens of Expressive Power

Neural Policy Composition from Free Energy Minimization

HAI-Eval: Measuring Human-AI Synergy in Collaborative Coding

Two-Dimensional Quantization for Geometry-Aware Audio Coding

Polynomial Neural Sheaf Diffusion: A Spectral Filtering Approach on Cellular Sheaves

LLM-EDT: Large Language Model Enhanced Cross-domain Sequential Recommendation with Dual-phase Training

DR Tulu: Reinforcement Learning with Evolving Rubrics for Deep Research

Sufficient Explanations in Databases and their Connections to Database Repairs

Weight Concentration Regularization for Improving Pruning Robustness Under High Sparsity

SemanticOpt: Towards LLM-Based Semantic Black-Box Optimization

Monotone and Separable Set Functions: Characterizations and Neural Models

SARVLM: A Vision Language Foundation Model for Semantic Understanding in SAR Imagery

A Model Can Help Itself: Reward-Free Self-Training for LLM Reasoning

ADMIT: Few-shot Knowledge Poisoning Attacks on RAG-based Fact Checking

UniShield: An Adaptive Multi-Agent Framework for Unified Forgery Image Detection and Localization

SAE-RNA: A Sparse Autoencoder Model for Interpreting RNA Language Model Representations

How to Train Your Advisor: Steering Black-Box LLMs with Advisor Models

NoiseShift: Resolution-Aware Noise Recalibration for Better Low-Resolution Image Generation

BioBlobs: Unsupervised Discovery of Functional Substructures for Protein Function Prediction

Dynamic-TreeRPO: Breaking the Independent Trajectory Bottleneck with Structured Sampling

Painless Activation Steering: An Automated, Lightweight Approach for Post-Training Large Language Models

Logic of Hypotheses: from Zero to Full Knowledge in Neurosymbolic Integration

Autoguided Online Data Curation for Diffusion Model Training

CUBE: Contrastive Understanding by Balanced Experiments

Generalized Policy Gradient with History-Aware Decision Transformer for Reliable Routing over Graph Signals

Beyond Binary Rewards: Training LMs to Reason About Their Uncertainty

CIS-BWE: Chaos-Informed Speech Bandwidth Extension

A Scalable Multi-LLM Collaboration System with Retrieval-based Selection and Exploration-Exploitation-Driven Enhancement

Blending Supervised and Reinforcement Fine-Tuning with Prefix Sampling

Deep Double Q-learning

The Hardness of Achieving Impact in AI for Social Impact Research: A Ground-Level View of Challenges & Opportunities

FAR: Function-preserving Attention Replacement for IMC-friendly Inference

ActiveDPO: Active Direct Preference Optimization for Sample-Efficient Alignment

COCO-Inpaint: A Benchmark for Detecting and Localizing Inpainting-Based Image Manipulations

Optimizing LLM Inference: Fluid-Guided Online Scheduling with Memory Constraints

Large Language Models Could Be Rote Learners

Do Chinese models speak Chinese languages?

TokenButler: Token Importance is Predictable

Seeing is Understanding: Unlocking Causal Attention into Modality-Mutual Attention for Multimodal LLMs

Tool-Star: Empowering LLM-Brained Multi-Tool Reasoner via Reinforcement Learning

Created by

Haebom

저자

Guanting Dong, Yifei Chen, Xiaoxi Li, Jiajie Jin, Hongjin Qian, Yutao Zhu, Hangyu Mao, Guorui Zhou, Zhicheng Dou, Ji-Rong Wen

개요

본 논문은 대규모 언어 모델(LLMs)의 다중 도구 협업 추론 능력 향상을 위한 강화학습 기반 프레임워크인 Tool-Star를 제안합니다. Tool-Star는 6가지 유형의 외부 도구를 자율적으로 호출하여 단계적 추론을 수행하며, 도구 사용 데이터 부족 문제를 해결하기 위해 도구 통합 프롬프트와 힌트 기반 샘플링을 결합한 데이터 합성 파이프라인을 제시합니다. 또한, LLM이 도구 호출 피드백을 통해 추론 패턴을 탐색하도록 유도하는 초기 미세조정과 계층적 보상 설계를 활용한 다중 도구 자기 비판 강화학습 알고리즘을 포함하는 2단계 학습 프레임워크를 제안합니다. 10개 이상의 어려운 추론 벤치마크 실험을 통해 Tool-Star의 효과와 효율성을 보여줍니다. 코드는 깃허브에서 공개됩니다.

시사점, 한계점

•

시사점:

◦

대규모 언어 모델의 다중 도구 협업 추론 능력 향상을 위한 효과적인 강화학습 기반 프레임워크 제시

◦

도구 사용 데이터 부족 문제 해결을 위한 효율적인 데이터 합성 파이프라인 제안

◦

초기 미세조정과 계층적 보상 설계를 활용한 2단계 학습 프레임워크를 통해 성능 향상

◦

다양한 추론 벤치마크에서 Tool-Star의 우수한 성능 검증

◦

오픈소스 코드 공개를 통한 연구의 재현성 및 확장성 증대

•

한계점:

◦

제안된 데이터 합성 파이프라인의 일반화 가능성에 대한 추가 연구 필요

◦

사용된 도구의 종류 및 수 제한에 따른 성능 저하 가능성

◦

다양한 유형의 추론 문제에 대한 Tool-Star의 일반화 성능 평가 필요

◦

계산 비용 및 학습 시간에 대한 추가적인 분석 필요

PDF 보기

Made with Slashpage