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.

AI/ML in 3GPP 5G Advanced -- Services and Architecture

CookAnything: A Framework for Flexible and Consistent Multi-Step Recipe Image Generation

Dynamic Correction of Erroneous State Estimates via Diffusion Bayesian Exploration

Concept-Guided Backdoor Attack on Vision Language Models

On the Holographic Geometry of Deterministic Computation

Decoding inner speech with an end-to-end brain-to-text neural interface

Physics-informed Neural Operator Learning for Nonlinear Grad-Shafranov Equation

MindEval: Benchmarking Language Models on Multi-turn Mental Health Support

LAET: A Layer-wise Adaptive Ensemble Tuning Framework for Pretrained Language Models

FAST-CAD: A Fairness-Aware Framework for Non-Contact Stroke Diagnosis

Designing LLM-based Multi-Agent Systems for Software Engineering Tasks: Quality Attributes, Design Patterns and Rationale

SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control

MOSS: Efficient and Accurate FP8 LLM Training with Microscaling and Automatic Scaling

Data-Augmented Deep Learning for Downhole Depth Sensing and Field Validation

Chinese Discharge Drug Recommendation in Metabolic Diseases with Large Language Models

Analysing Moral Bias in Finetuned LLMs through Mechanistic Interpretability

Orders in Chaos: Enhancing Large-Scale MoE LLM Serving with Data Movement Forecasting

Reinforce-Ada: An Adaptive Sampling Framework under Non-linear RL Objectives

TempoControl: Temporal Attention Guidance for Text-to-Video Models

AortaDiff: A Unified Multitask Diffusion Framework For Contrast-Free AAA Imaging

SoREX: Towards Self-Explainable Social Recommendation with Relevant Ego-Path Extraction

The AI Productivity Index (APEX)

Uncovering Grounding IDs: How External Cues Shape Multimodal Binding

Pushing Toward the Simplex Vertices: A Simple Remedy for Code Collapse in Smoothed Vector Quantization

V-CECE: Visual Counterfactual Explanations via Conceptual Edits

Momentum-constrained Hybrid Heuristic Trajectory Optimization Framework with Residual-enhanced DRL for Visually Impaired Scenarios

HARP: Hallucination Detection via Reasoning Subspace Projection

Rethinking Sparse Autoencoders: Select-and-Project for Fairness and Control from Encoder Features Alone

IPA: An Information-Reconstructive Input Projection Framework for Efficient Foundation Model Adaptation

Sparse but Wrong: Incorrect L0 Leads to Incorrect Features in Sparse Autoencoders

ORFuzz: Fuzzing the "Other Side" of LLM Safety -- Testing Over-Refusal

Retro-Expert: Collaborative Reasoning for Interpretable Retrosynthesis

A Survey on Diffusion Language Models

CodeNER: Code Prompting for Named Entity Recognition

ReSem3D: Refinable 3D Spatial Constraints via Fine-Grained Semantic Grounding for Generalizable Robotic Manipulation

VERIRAG: A Post-Retrieval Auditing of Scientific Study Summaries

SustainDiffusion: Optimising the Social and Environmental Sustainability of Stable Diffusion Models

LittleBit: Ultra Low-Bit Quantization via Latent Factorization

Real-Time Execution of Action Chunking Flow Policies

AURA: A Diagnostic Framework for Tracking User Satisfaction of Interactive Planning Agents

Enhancing SPARQL Query Rewriting for Complex Ontology Alignments

Exploring Ordinal Bias in Action Recognition for Instructional Videos

Robust Weight Imprinting: Insights from Neural Collapse and Proxy-Based Aggregation

Experiments with Large Language Models on Retrieval-Augmented Generation for Closed-Source Simulation Software

SAT: Dynamic Spatial Aptitude Training for Multimodal Language Models

Large Language Models: An Applied Econometric Framework

Edge-Only Universal Adversarial Attacks in Distributed Learning

Image-Guided Semantic Pseudo-LiDAR Point Generation for 3D Object Detection

Detecting the Future: All-at-Once Event Sequence Forecasting with Horizon Matching

Variational Learning of Gaussian Process Latent Variable Models through Stochastic Gradient Annealed Importance Sampling

SOAP: Enhancing Spatio-Temporal Relation and Motion Information Capturing for Few-Shot Action Recognition

A Scene-aware Models Adaptation Scheme for Cross-scene Online Inference on Mobile Devices

Towards Data-efficient Customer Intent Recognition with Prompt-based Learning Paradigm

GTM: Simulating the World of Tools for AI Agents

Self-Transparency Failures in Expert-Persona LLMs: How Instruction-Following Overrides Honesty

Learning the Value of Value Learning

OpenMMReasoner: Pushing the Frontiers for Multimodal Reasoning with an Open and General Recipe

ToolMind Technical Report: A Large-Scale, Reasoning-Enhanced Tool-Use Dataset

Debate over Mixed-knowledge: A Robust Multi-Agent Reasoning Framework for Incomplete Knowledge Graph Question Answering

Spilling the Beans: Teaching LLMs to Self-Report Their Hidden Objectives

Counterfactual Reasoning for Steerable Pluralistic Value Alignment of Large Language Models

Generalized Parallel Scaling with Interdependent Generations

SOCK: A Benchmark for Measuring Self-Replication in Large Language Models

KNARsack: Teaching Neural Algorithmic Reasoners to Solve Pseudo-Polynomial Problems

IS-Bench: Evaluating Interactive Safety of VLM-Driven Embodied Agents in Daily Household Tasks

FedIFL: A federated cross-domain diagnostic framework for motor-driven systems with inconsistent fault modes

Enhancing Large Language Models through Neuro-Symbolic Integration and Ontological Reasoning

Rolling in the deep of cognitive and AI biases

Enhancing Retrieval-Augmented Generation with Entity Linking for Educational Platforms

Training-Time Action Conditioning for Efficient Real-Time Chunking

Whatever Remains Must Be True: Filtering Drives Reasoning in LLMs, Shaping Diversity

AQUA-Net: Adaptive Frequency Fusion and Illumination Aware Network for Underwater Image Enhancement

M4-RAG: A Massive-Scale Multilingual Multi-Cultural Multimodal RAG

MaxShapley: Towards Incentive-compatible Generative Search with Fair Context Attribution

Trusted AI Agents in the Cloud

Impugan: Learning Conditional Generative Models for Robust Data Imputation

Zoom in, Click out: Unlocking and Evaluating the Potential of Zooming for GUI Grounding

Measuring the Effect of Background on Classification and Feature Importance in Deep Learning for AV Perception

World Models That Know When They Don't Know: Controllable Video Generation with Calibrated Uncertainty

Natural Language Summarization Enables Multi-Repository Bug Localization by LLMs in Microservice Architectures

Neural Coherence : Find higher performance to out-of-distribution tasks from few samples

Sparse Attention Post-Training for Mechanistic Interpretability

Optimizing Medical Question-Answering Systems: A Comparative Study of Fine-Tuned and Zero-Shot Large Language Models with RAG Framework

NEAT: Neighborhood-Guided, Efficient, Autoregressive Set Transformer for 3D Molecular Generation

Phase-OTDR Event Detection Using Image-Based Data Transformation and Deep Learning

Approximation of Box Decomposition Algorithm for Fast Hypervolume-Based Multi-Objective Optimization

Probing the effectiveness of World Models for Spatial Reasoning through Test-time Scaling

3D Path Planning for Robot-assisted Vertebroplasty from Arbitrary Bi-plane X-ray via Differentiable Rendering

Mechanistic Interpretability of Antibody Language Models Using SAEs

Active Video Perception: Iterative Evidence Seeking for Agentic Long Video Understanding

Efficient Text Classification with Conformal In-Context Learning

Big Tech-Funded AI Papers Have Higher Citation Impact, Greater Insularity, and Larger Recency Bias

Bayesian Active Inference for Intelligent UAV Anti-Jamming and Adaptive Trajectory Planning

Faithfulness metric fusion: Improving the evaluation of LLM trustworthiness across domains

Retrieving Semantically Similar Decisions under Noisy Institutional Labels: Robust Comparison of Embedding Methods

InverseCrafter: Efficient Video ReCapture as a Latent Domain Inverse Problem

Feasibility of AI-Assisted Programming for End-User Development

Grounded Multilingual Medical Reasoning for Question Answering with Large Language Models

Modular Jets for Supervised Pipelines: Diagnosing Mirage vs Identifiability

A Comprehensive Framework for Automated Quality Control in the Automotive Industry

"Moralized" Multi-Step Jailbreak Prompts: Black-Box Testing of Guardrails in Large Language Models for Verbal Attacks

Created by

Haebom

Category

Empty

저자

Libo Wang

개요

본 연구는 다양한 분야에서 대규모 언어 모델(LLM)의 활용이 증가함에 따라 유해 콘텐츠 생성 및 안전장치 효과성에 대한 어려움이 커지고 있음을 지적합니다. GPT-4o, Grok-2 Beta, Llama 3.1 (405B), Gemini 1.5, 그리고 Claude 3.5 Sonnet 등의 LLM에 대해, 윤리적인 다단계 우회 프롬프트를 사용한 블랙박스 테스트를 통해 안전장치의 효과성을 평가했습니다. "승진 경쟁을 벌이는 기업 중간 관리자" 시나리오를 모방한 동일한 다단계 프롬프트를 설계하여 윤리적 공격을 수행했습니다. 그 결과, 모든 LLM의 안전장치가 우회되었고, 언어적 공격 내용이 생성되었음을 보였습니다. 다만, Claude 3.5 Sonnet은 다단계 우회 프롬프트에 대한 저항성이 더 높게 나타났습니다. 객관성을 위해 실험 과정, 블랙박스 테스트 코드, 그리고 개선된 안전장치 코드를 GitHub 저장소(https://github.com/brucewang123456789/GeniusTrail.git)에 공개했습니다.

시사점, 한계점

•

시사점: 다양한 LLM의 안전장치가 다단계 우회 프롬프트에 취약함을 보여줌. Claude 3.5 Sonnet이 상대적으로 우수한 저항성을 보임. LLM의 안전장치 개선 및 윤리적 사용에 대한 지속적인 연구 필요성 제기. 블랙박스 테스트 코드 공개를 통한 연구의 투명성 확보.

•

한계점: 특정 유형의 다단계 프롬프트에 국한된 테스트. 더욱 다양하고 정교한 우회 기법에 대한 추가 연구 필요. "기업 중간 관리자의 승진 경쟁" 시나리오에 국한된 상황 설정으로 일반화에 대한 제한. 실제 세계의 다양한 상황을 반영하지 못할 가능성.

Made with Slashpage