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.

Capturing Complex Spatial-Temporal Dependencies in Traffic Forecasting: A Self-Attention Approach

Benchmarking Multi-Step Legal Reasoning and Analyzing Chain-of-Thought Effects in Large Language Models

Towards Fine-Grained Interpretability: Counterfactual Explanations for Misclassification with Saliency Partition

Versatile and Risk-Sensitive Cardiac Diagnosis via Graph-Based ECG Signal Representation

TimeFlow: Towards Stochastic-Aware and Efficient Time Series Generation via Flow Matching Modeling

Thinker: Training LLMs in Hierarchical Thinking for Deep Search via Multi-Turn Interaction

Toward Practical BCI: A Real-time Wireless Imagined Speech EEG Decoding System

Computational Blueprints: Generating Isomorphic Mathematics Problems with Large Language Models

Lightweight Diffusion-based Framework for Online Imagined Speech Decoding in Aphasia

Neurophysiological Characteristics of Adaptive Reasoning for Creative Problem-Solving Strategy

DANS-KGC: Diffusion Based Adaptive Negative Sampling for Knowledge Graph Completion

Data Descriptions from Large Language Models with Influence Estimation

SparseRM: A Lightweight Preference Modeling with Sparse Autoencoder

Toward Robust EEG-based Intention Decoding during Misarticulated Speech in Aphasia

Confidence-Aware Neural Decoding of Overt Speech from EEG: Toward Robust Brain-Computer Interfaces

WaterMod: Modular Token-Rank Partitioning for Probability-Balanced LLM Watermarking

GAMA: A Neural Neighborhood Search Method with Graph-aware Multi-modal Attention for Vehicle Routing Problem

Alignment-Aware Quantization for LLM Safety

Operational machine learning for remote spectroscopic detection of CH$_{4}$ point sources

Towards AI-Assisted Generation of Military Training Scenarios

ResearchRubrics: A Benchmark of Prompts and Rubrics For Evaluating Deep Research Agents

AIA Forecaster: Technical Report

Making LLMs Reliable When It Matters Most: A Five-Layer Architecture for High-Stakes Decisions

AI-Driven Contribution Evaluation and Conflict Resolution: A Framework & Design for Group Workload Investigation

Beyond Fact Retrieval: Episodic Memory for RAG with Generative Semantic Workspaces

Think Before You Retrieve: Learning Test-Time Adaptive Search with Small Language Models

Procedural Knowledge Improves Agentic LLM Workflows

Beyond Correctness: Confidence-Aware Reward Modeling for Enhancing Large Language Model Reasoning

Agentic Educational Content Generation for African Languages on Edge Devices

Analysing Environmental Efficiency in AI for X-Ray Diagnosis

SWE-Compass: Towards Unified Evaluation of Agentic Coding Abilities for Large Language Models

Perceptually Aligning Representations of Music via Noise-Augmented Autoencoders

Rethinking Metrics and Diffusion Architecture for 3D Point Cloud Generation

DeepEyesV2: Toward Agentic Multimodal Model

MERaLiON-SER: Robust Speech Emotion Recognition Model for English and SEA Languages

Minimal and Mechanistic Conditions for Behavioral Self-Awareness in LLMs

multiMentalRoBERTa: A Fine-tuned Multiclass Classifier for Mental Health Disorder

Addressing divergent representations from causal interventions on neural networks

Alternative Fairness and Accuracy Optimization in Criminal Justice

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

Conversational Collective Intelligence (CCI) using Hyperchat AI in a Real-world Forecasting Task

Explaining Human Choice Probabilities with Simple Vector Representations

Development of the Bioinspired Tendon-Driven DexHand 021 with Proprioceptive Compliance Control

GraphCliff: Short-Long Range Gating for Subtle Differences but Critical Changes

AutoAdv: Automated Adversarial Prompting for Multi-Turn Jailbreaking of Large Language Models

MM-UNet: Morph Mamba U-shaped Convolutional Networks for Retinal Vessel Segmentation

RobustFSM: Submodular Maximization in Federated Setting with Malicious Clients

Tool Zero: Training Tool-Augmented LLMs via Pure RL from Scratch

Reg-DPO: SFT-Regularized Direct Preference Optimization with GT-Pair for Improving Video Generation

Reasoning Planning for Language Models

Aligning Brain Signals with Multimodal Speech and Vision Embeddings

Calibrating and Rotating: A Unified Framework for Weight Conditioning in PEFT

FedAdamW: A Communication-Efficient Optimizer with Convergence and Generalization Guarantees for Federated Large Models

Beyond Data Scarcity Optimizing R3GAN for Medical Image Generation from Small Datasets

Inside CORE-KG: Evaluating Structured Prompting and Coreference Resolution for Knowledge Graphs

Learning from N-Tuple Data with M Positive Instances: Unbiased Risk Estimation and Theoretical Guarantees

Universal Spectral Tokenization via Self-Supervised Panchromatic Representation Learning

DynaSpec: Context-aware Dynamic Speculative Sampling for Large-Vocabulary Language Models

Meronymic Ontology Extraction via Large Language Models

Robust Driving Control for Autonomous Vehicles: An Intelligent General-sum Constrained Adversarial Reinforcement Learning Approach

DPCformer: An Interpretable Deep Learning Model for Genomic Prediction in Crops

The Markovian Thinker

scUnified: An AI-Ready Standardized Resource for Single-Cell RNA Sequencing Analysis

Data-Centric Elastic Pipeline Parallelism for Efficient Long-Context LLM Training

Security-aware Semantic-driven ISAC via Paired Adversarial Residual Networks

TimeMosaic: Temporal Heterogeneity Guided Time Series Forecasting via Adaptive Granularity Patch and Segment-wise Decoding

Pure Vision Language Action (VLA) Models: A Comprehensive Survey

UniPixel: Unified Object Referring and Segmentation for Pixel-Level Visual Reasoning

Assisting the Grading of a Handwritten General Chemistry Exam with Artificial Intelligence

Privacy-Preserving Personalization in Education: A Federated Recommender System for Student Performance Prediction

UNO: Unifying One-stage Video Scene Graph Generation via Object-Centric Visual Representation Learning

Normality and the Turing Test

GUARD: Guideline Upholding Test through Adaptive Role-play and Jailbreak Diagnostics for LLMs

Cross-Platform E-Commerce Product Categorization and Recategorization: A Multimodal Hierarchical Classification Approach

ANO : Faster is Better in Noisy Landscape

Generative Medical Event Models Improve with Scale

ComoRAG: A Cognitive-Inspired Memory-Organized RAG for Stateful Long Narrative Reasoning

LLMCARE: early detection of cognitive impairment via transformer models enhanced by LLM-generated synthetic data

Oblivionis: A Lightweight Learning and Unlearning Framework for Federated Large Language Models

Rethinking Tokenization for Rich Morphology: The Dominance of Unigram over BPE and Morphological Alignment

Discovering Spatial Correlations of Earth Observations for weather forecasting by using Graph Structure Learning

Forecasting When to Forecast: Accelerating Diffusion Models with Confidence-Gated Taylor

Universal Neurons in GPT-2: Emergence, Persistence, and Functional Impact

Embedding-Aware Quantum-Classical SVMs for Scalable Quantum Machine Learning

RaGS: Unleashing 3D Gaussian Splatting from 4D Radar and Monocular Cues for 3D Object Detection

Controllable Hybrid Captioner for Improved Long-form Video Understanding

Steering Out-of-Distribution Generalization with Concept Ablation Fine-Tuning

When Person Re-Identification Meets Event Camera: A Benchmark Dataset and An Attribute-guided Re-Identification Framework

Evaluating LLM-based Workflows for Switched-Mode Power Supply Design

Monitoring Risks in Test-Time Adaptation

White-Basilisk: A Hybrid Model for Code Vulnerability Detection

Bridging the Plausibility-Validity Gap by Fine-Tuning a Reasoning-Enhanced LLM for Chemical Synthesis and Discovery

CoRe: Benchmarking LLMs Code Reasoning Capabilities through Static Analysis Tasks

DP-Fusion: Token-Level Differentially Private Inference for Large Language Models

Meta SecAlign: A Secure Foundation LLM Against Prompt Injection Attacks

GPT, But Backwards: Exactly Inverting Language Model Outputs

Visual Structures Helps Visual Reasoning: Addressing the Binding Problem in VLMs

ReCode: Updating Code API Knowledge with Reinforcement Learning

Time-Prompt: Integrated Heterogeneous Prompts for Unlocking LLMs in Time Series Forecasting

Sekai: A Video Dataset towards World Exploration

Inverse-Free Wilson Loops for Transformers: A Practical Diagnostic for Invariance and Order Sensitivity

Created by

Haebom

저자

Edward Y. Chang, Ethan Y. Chang

개요

대규모 언어 모델은 실질적으로 중요한 무해한 수정에도 답변을 변경할 수 있음. RAG 출력은 구절이 재정렬될 때 바뀌고, 미세 조정은 사전 훈련에서 학습된 불변성을 약화시키며, 토론 또는 사고 사슬 프롬프트는 경로 종속적인 경로를 따르며, 컴파일러 융합 또는 재정렬은 결정 경계 근처의 로짓을 방해함. 이러한 실패는 의도된 불변성을 위반하고, 지속적인 통합을 깨뜨리며, 팀이 속도를 위해 안전성을 타협하도록 함. WILSON은 내부 표현에 대한 간단한 루프 및 재정렬 검사를 시스템 신호로 변환하는 최소한의 사후 진단 도구 모음임. WILSON은 JVPs 및 Hutchinson 프로브로 계산된 위치 및 레이어에 대한 역 없는 곡률 맵과 재정렬 위험을 플래그하는 활성화 수준의 교환기를 결합함. 신호는 계산 비용이 저렴하고, 표준 Transformer에 대해 모델 불가지론적이며, 오케스트레이터를 위해 임계값 및 CSV 아티팩트로 내보내짐. 이를 통해 RAG를 순서 효과로부터 보호하고, 미세 조정 회귀를 포착하고, 토론 경로 및 긴 다중 턴 컨텍스트를 안정화하고, 배포에서 융합 또는 재정렬을 제어하는 등의 구체적인 작업을 수행할 수 있음. 즉, WILSON은 실패를 예상하고 안전한 최적화를 승인하여 모델 아키텍처나 훈련을 변경하지 않고도 신뢰성과 처리량을 함께 개선하는 데 도움이 됨.

시사점, 한계점

•

시사점:

◦

WILSON은 대규모 언어 모델의 안정성을 향상시키기 위한 사후 진단 도구임.

◦

모델 아키텍처나 훈련을 변경하지 않고 안전한 최적화를 가능하게 함.

◦

RAG의 순서 효과, 미세 조정 회귀, 토론 경로 불안정성 등을 해결하는 데 도움을 줌.

•

한계점:

◦

WILSON의 구체적인 성능 및 효과에 대한 추가 정보가 필요함.

◦

다른 모델 및 작업에 대한 일반화 가능성에 대한 평가가 필요함.

◦

WILSON의 실제 적용 및 통합에 대한 구체적인 사례 연구가 부족함.

Made with Slashpage