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.

Something's Fishy In The Data Lake: A Critical Re-evaluation of Table Union Search Benchmarks

Breaking the Ceiling: Exploring the Potential of Jailbreak Attacks through Expanding Strategy Space

SageAttention2++: A More Efficient Implementation of SageAttention2

FCKT: Fine-Grained Cross-Task Knowledge Transfer with Semantic Contrastive Learning for Targeted Sentiment Analysis

Towards Conversational Development Environments: Using Theory-of-Mind and Multi-Agent Architectures for Requirements Refinement

Cooperation of Experts: Fusing Heterogeneous Information with Large Margin

RSCF: Relation-Semantics Consistent Filter for Entity Embedding of Knowledge Graph

CogniBench: A Legal-inspired Framework and Dataset for Assessing Cognitive Faithfulness of Large Language Models

VLM Can Be a Good Assistant: Enhancing Embodied Visual Tracking with Self-Improving Vision-Language Models

In-context Language Learning for Endangered Languages in Speech Recognition

OpenS2V-Nexus: A Detailed Benchmark and Million-Scale Dataset for Subject-to-Video Generation

Inference-time Alignment in Continuous Space

Incentivizing Strong Reasoning from Weak Supervision

JEDI: Latent End-to-end Diffusion Mitigates Agent-Human Performance Asymmetry in Model-Based Reinforcement Learning

A Comprehensive Real-World Assessment of Audio Watermarking Algorithms: Will They Survive Neural Codecs?

AgentRecBench: Benchmarking LLM Agent-based Personalized Recommender Systems

Towards Large Reasoning Models for Agriculture

VTool-R1: VLMs Learn to Think with Images via Reinforcement Learning on Multimodal Tool Use

OptiMindTune: A Multi-Agent Framework for Intelligent Hyperparameter Optimization

FastMamba: A High-Speed and Efficient Mamba Accelerator on FPGA with Accurate Quantization

Moderating Harm: Benchmarking Large Language Models for Cyberbullying Detection in YouTube Comments

ALPS: Attention Localization and Pruning Strategy for Efficient Alignment of Large Language Models

Deep Video Discovery: Agentic Search with Tool Use for Long-form Video Understanding

Towards Practical Defect-Focused Automated Code Review

EVM-Fusion: An Explainable Vision Mamba Architecture with Neural Algorithmic Fusion

PAEFF: Precise Alignment and Enhanced Gated Feature Fusion for Face-Voice Association

Text Generation Beyond Discrete Token Sampling

AKRMap: Adaptive Kernel Regression for Trustworthy Visualization of Cross-Modal Embeddings

Information Science Principles of Machine Learning: A Causal Chain Meta-Framework Based on Formalized Information Mapping

Advancing Sequential Numerical Prediction in Autoregressive Models

Towards Visuospatial Cognition via Hierarchical Fusion of Visual Experts

Visuospatial Cognitive Assistant

Exploring Criteria of Loss Reweighting to Enhance LLM Unlearning

Search and Refine During Think: Autonomous Retrieval-Augmented Reasoning of LLMs

Learning When to Think: Shaping Adaptive Reasoning in R1-Style Models via Multi-Stage RL

LibIQ: Toward Real-Time Spectrum Classification in O-RAN dApps

Continuous Thought Machines

Robust Localization, Mapping, and Navigation for Quadruped Robots

Addressing Concept Mislabeling in Concept Bottleneck Models Through Preference Optimization

Automating tumor-infiltrating lymphocyte assessment in breast cancer histopathology images using QuPath: a transparent and accessible machine learning pipeline

Progressive Language-guided Visual Learning for Multi-Task Visual Grounding

A Comprehensive Survey in LLM(-Agent) Full Stack Safety: Data, Training and Deployment

Generative Framework for Personalized Persuasion: Inferring Causal, Counterfactual, and Latent Knowledge

Adjoint Sampling: Highly Scalable Diffusion Samplers via Adjoint Matching

Layers at Similar Depths Generate Similar Activations Across LLM Architectures

A Knowledge-guided Adversarial Defense for Resisting Malicious Visual Manipulation

SynWorld: Virtual Scenario Synthesis for Agentic Action Knowledge Refinement

Token embeddings violate the manifold hypothesis

HumanAesExpert: Advancing a Multi-Modality Foundation Model for Human Image Aesthetic Assessment

Improving User Behavior Prediction: Leveraging Annotator Metadata in Supervised Machine Learning Models

Experience Retrieval-Augmentation with Electronic Health Records Enables Accurate Discharge QA

From Head to Tail: Towards Balanced Representation in Large Vision-Language Models through Adaptive Data Calibration

Language-Enhanced Representation Learning for Single-Cell Transcriptomics

TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are Stronger Zero-Shot Forecaster

When Trust Collides: Decoding Human-LLM Cooperation Dynamics through the Prisoner's Dilemma

WISE: A World Knowledge-Informed Semantic Evaluation for Text-to-Image Generation

Optimal Output Feedback Learning Control for Discrete-Time Linear Quadratic Regulation

Wanda++: Pruning Large Language Models via Regional Gradients

LINGOLY-TOO: Disentangling Reasoning from Knowledge with Templatised Orthographic Obfuscation

Deficient Excitation in Parameter Learning

Interpreting CLIP with Hierarchical Sparse Autoencoders

BatteryLife: A Comprehensive Dataset and Benchmark for Battery Life Prediction

Self-Taught Agentic Long Context Understanding

ReQFlow: Rectified Quaternion Flow for Efficient and High-Quality Protein Backbone Generation

How Do LLMs Perform Two-Hop Reasoning in Context?

ThinkGuard: Deliberative Slow Thinking Leads to Cautious Guardrails

MUDDFormer: Breaking Residual Bottlenecks in Transformers via Multiway Dynamic Dense Connections

ReLearn: Unlearning via Learning for Large Language Models

Revisiting Weak-to-Strong Generalization in Theory and Practice: Reverse KL vs. Forward KL

A Physics-Informed Machine Learning Framework for Safe and Optimal Control of Autonomous Systems

Non-Markovian Discrete Diffusion with Causal Language Models

CoSER: Coordinating LLM-Based Persona Simulation of Established Roles

TransMLA: Migrating GQA Models to MLA with Full DeepSeek Compatibility and Speedup

UniDB: A Unified Diffusion Bridge Framework via Stochastic Optimal Control

Beyond External Monitors: Enhancing Transparency of Large Language Models for Easier Monitoring

ExpProof : Operationalizing Explanations for Confidential Models with ZKPs

Advancing Reasoning in Large Language Models: Promising Methods and Approaches

Path Planning for Masked Diffusion Model Sampling

LV-XAttn: Distributed Cross-Attention for Long Visual Inputs in Multimodal Large Language Models

Robust LLM Alignment via Distributionally Robust Direct Preference Optimization

Improving Rule-based Reasoning in LLMs using Neurosymbolic Representations

Message-Passing GNNs Fail to Approximate Sparse Triangular Factorizations

LoRA-One: One-Step Full Gradient Could Suffice for Fine-Tuning Large Language Models, Provably and Efficiently

Distribution-aware Fairness Learning in Medical Image Segmentation From A Control-Theoretic Perspective

A Checks-and-Balances Framework for Context-Aware Ethical AI Alignment

Risk-Informed Diffusion Transformer for Long-Tail Trajectory Prediction in the Crash Scenario

Can Large Language Models Be Trusted as Evolutionary Optimizers for Network-Structured Combinatorial Problems?

Redundancy Principles for MLLMs Benchmarks

K-COMP: Retrieval-Augmented Medical Domain Question Answering With Knowledge-Injected Compressor

An Innovative Data-Driven and Adaptive Reinforcement Learning Approach for Context-Aware Prescriptive Process Monitoring

Diffusion Adversarial Post-Training for One-Step Video Generation

Gender-Neutral Large Language Models for Medical Applications: Reducing Bias in PubMed Abstracts

PRMBench: A Fine-grained and Challenging Benchmark for Process-Level Reward Models

Constraint-Adaptive Policy Switching for Offline Safe Reinforcement Learning

Revisiting In-Context Learning with Long Context Language Models

Energy and polarization based on-line interference mitigation in radio interferometry

How to Synthesize Text Data without Model Collapse?

Preference Adaptive and Sequential Text-to-Image Generation

Sample Efficient Robot Learning in Supervised Effect Prediction Tasks

Federated Continual Graph Learning

RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts

Created by

Haebom

저자

Hjalmar Wijk, Tao Lin, Joel Becker, Sami Jawhar, Neev Parikh, Thomas Broadley, Lawrence Chan, Michael Chen, Josh Clymer, Jai Dhyani, Elena Ericheva, Katharyn Garcia, Brian Goodrich, Nikola Jurkovic, Holden Karnofsky, Megan Kinniment, Aron Lajko, Seraphina Nix, Lucas Sato, William Saunders, Maksym Taran, Ben West, Elizabeth Barnes

개요

본 논문은 AI 연구 개발(R&D) 능력 평가를 위한 새로운 벤치마크인 RE-Bench(Research Engineering Benchmark, v1)를 소개합니다. RE-Bench는 7가지의 어려운 개방형 ML 연구 엔지니어링 환경과 61명의 전문가가 8시간 동안 시도한 71개의 데이터로 구성됩니다. 논문은 8시간 동안 전문가들이 상당한 진전을 이루었으며(82%가 0점이 아닌 점수를 달성, 24%가 기준 해결책을 충족 또는 초과), 여러 최첨단 모델과 인간 전문가의 성능을 비교 분석합니다. 2시간의 시간 제한에서는 최고의 AI 에이전트가 인간 전문가보다 4배 높은 점수를 달성하지만, 8시간 이상의 시간이 주어지면 인간의 성과가 AI를 능가하며, 32시간의 시간 제한에서는 인간이 AI보다 2배 높은 점수를 기록합니다. AI 에이전트는 많은 ML 주제에 대한 상당한 전문 지식을 가지고 있으며(예: 인간 전문가보다 빠른 맞춤형 Triton 커널 작성), 인간보다 10배 이상 빠르고 저렴하게 솔루션을 생성하고 테스트할 수 있습니다. RE-Bench의 평가 환경, 인간 전문가 데이터, 분석 코드 및 에이전트 경로는 오픈소스로 공개되어 향후 연구에 활용될 수 있습니다.

시사점, 한계점

•

시사점:

◦

현존 최첨단 AI 모델의 ML 연구 개발 능력에 대한 현실적인 평가 기준을 제시합니다.

◦

인간 전문가와 AI 에이전트의 성능을 직접 비교하여 각각의 강점과 약점을 보여줍니다.

◦

AI 에이전트의 높은 속도와 효율성을 확인하며, AI를 활용한 연구 개발 가속화 가능성을 제시합니다.

◦

오픈소스 데이터셋과 평가 환경 제공을 통해 향후 연구 발전에 기여합니다.

•

한계점:

◦

벤치마크의 과제가 특정 ML 분야에 치우쳐져 일반화에 한계가 있을 수 있습니다.

◦

참여한 인간 전문가의 수와 다양성이 제한적일 수 있습니다.

◦

장기간의 연구 개발 과정 전체를 반영하지 못하고 단기적인 성과에 집중되어 있습니다.

◦

AI 에이전트의 성능 향상에 따른 벤치마크의 지속적인 업데이트 필요성이 존재합니다.

Made with Slashpage