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.

Dually Hierarchical Drift Adaptation for Online Configuration Performance Learning

Token-based Audio Inpainting via Discrete Diffusion

Mechanistic Indicators of Understanding in Large Language Models

PyVision: Agentic Vision with Dynamic Tooling

Low Resource Reconstruction Attacks Through Benign Prompts

DTECT: Dynamic Topic Explorer & Context Tracker

Your Absorbing Discrete Diffusion Secretly Models the Bayesian Posterior

Bias-Aware Mislabeling Detection via Decoupled Confident Learning

Bridging the Last Mile of Prediction: Enhancing Time Series Forecasting with Conditional Guided Flow Matching

Planted in Pretraining, Swayed by Finetuning: A Case Study on the Origins of Cognitive Biases in LLMs

FlexOlmo: Open Language Models for Flexible Data Use

Democratizing High-Fidelity Co-Speech Gesture Video Generation

GR-LLMs: Recent Advances in Generative Recommendation Based on Large Language Models

LIRA: Inferring Segmentation in Large Multi-modal Models with Local Interleaved Region Assistance

BayesSDF: Surface-Based Laplacian Uncertainty Estimation for 3D Geometry with Neural Signed Distance Fields

Fast Bilateral Teleoperation and Imitation Learning Using Sensorless Force Control via Accurate Dynamics Model

DESIGN: Encrypted GNN Inference via Server-Side Input Graph Pruning

Beyond classical and contemporary models: a transformative AI framework for student dropout prediction in distance learning using RAG, Prompt engineering, and Cross-modal fusion

Hear-Your-Click: Interactive Object-Specific Video-to-Audio Generation

PRIME: Large Language Model Personalization with Cognitive Memory and Thought Processes

LearnLens: LLM-Enabled Personalised, Curriculum-Grounded Feedback with Educators in the Loop

Zero-Shot Cyclic Peptide Design via Composable Geometric Constraints

Model Collapse Is Not a Bug but a Feature in Machine Unlearning for LLMs

SymbolicThought: Integrating Language Models and Symbolic Reasoning for Consistent and Interpretable Human Relationship Understanding

From Video to EEG: Adapting Joint Embedding Predictive Architecture to Uncover Visual Concepts in Brain Signal Analysis

An Efficient Deep Learning Framework for Brain Stroke Diagnosis Using Computed Tomography (CT) Images

De-Fake: Style based Anomaly Deepfake Detection

ForgeHLS: A Large-Scale, Open-Source Dataset for High-Level Synthesis

Expert-level validation of AI-generated medical text with scalable language models

DriveMRP: Enhancing Vision-Language Models with Synthetic Motion Data for Motion Risk Prediction

FlowSpec: Continuous Pipelined Speculative Decoding for Efficient Distributed LLM Inference

DeltaSHAP: Explaining Prediction Evolutions in Online Patient Monitoring with Shapley Values

Following the Clues: Experiments on Person Re-ID using Cross-Modal Intelligence

Description of the Training Process of Neural Networks via Ergodic Theorem : Ghost nodes

Quality over Quantity: An Effective Large-Scale Data Reduction Strategy Based on Pointwise V-Information

Imagine for Me: Creative Conceptual Blending of Real Images and Text via Blended Attention

When Small Guides Large: Cross-Model Co-Learning for Test-Time Adaptation

Perspective Dial: Measuring Perspective of Text and Guiding LLM Outputs

CRISP-SAM2: SAM2 with Cross-Modal Interaction and Semantic Prompting for Multi-Organ Segmentation

Positioning AI Tools to Support Online Harm Reduction Practice: Applications and Design Directions

Interpretable Time Series Autoregression for Periodicity Quantification

Teaching Models to Verbalize Reward Hacking in Chain-of-Thought Reasoning

Exploration Behavior of Untrained Policies

Ark: An Open-source Python-based Framework for Robot Learning

Hyperspherical Variational Autoencoders Using Efficient Spherical Cauchy Distribution

TReB: A Comprehensive Benchmark for Evaluating Table Reasoning Capabilities of Large Language Models

The Debugging Decay Index: Rethinking Debugging Strategies for Code LLMs

Towards Effective Complementary Security Analysis using Large Language Models

Vision-Guided Chunking Is All You Need: Enhancing RAG with Multimodal Document Understanding

Learning from M-Tuple Dominant Positive and Unlabeled Data

BreastDCEDL: A Comprehensive Breast Cancer DCE-MRI Dataset and Transformer Implementation for Treatment Response Prediction

Pisces: An Auto-regressive Foundation Model for Image Understanding and Generation

SLED: A Speculative LLM Decoding Framework for Efficient Edge Serving

BIS Reasoning 1.0: The First Large-Scale Japanese Benchmark for Belief-Inconsistent Syllogistic Reasoning

FuseUNet: A Multi-Scale Feature Fusion Method for U-like Networks

Subgroups Matter for Robust Bias Mitigation

Robust Stability Analysis of Positive Lure System with Neural Network Feedback

Next-token pretraining implies in-context learning

A modular framework for automated evaluation of procedural content generation in serious games with deep reinforcement learning agents

Learning Flexible Forward Trajectories for Masked Molecular Diffusion

LEXam: Benchmarking Legal Reasoning on 340 Law Exams

Large Language Model Psychometrics: A Systematic Review of Evaluation, Validation, and Enhancement

Quantum Observers: A NISQ Hardware Demonstration of Chaotic State Prediction Using Quantum Echo-state Networks

Visual Test-time Scaling for GUI Agent Grounding

Consistency in Language Models: Current Landscape, Challenges, and Future Directions

Avoiding Leakage Poisoning: Concept Interventions Under Distribution Shifts

Roll the dice & look before you leap: Going beyond the creative limits of next-token prediction

Adaptive Non-local Observable on Quantum Neural Networks

Leveraging Large Language Models for Multi-Class and Multi-Label Detection of Drug Use and Overdose Symptoms on Social Media

Bypassing LLM Guardrails: An Empirical Analysis of Evasion Attacks against Prompt Injection and Jailbreak Detection Systems

Divergence of Empirical Neural Tangent Kernel in Classification Problems

Rethinking the Foundations for Continual Reinforcement Learning

Predictive Modeling: BIM Command Recommendation Based on Large-scale Usage Logs

Unfair Learning: GenAI Exceptionalism and Copyright Law

OS-Kairos: Adaptive Interaction for MLLM-Powered GUI Agents

LLMs' Leaning in European Elections

COVID-19 Pneumonia Diagnosis Using Medical Images: Deep Learning-Based Transfer Learning Approach

Can A Society of Generative Agents Simulate Human Behavior and Inform Public Health Policy? A Case Study on Vaccine Hesitancy

Learning-Order Autoregressive Models with Application to Molecular Graph Generation

KodCode: A Diverse, Challenging, and Verifiable Synthetic Dataset for Coding

Class-Aware PillarMix: Can Mixed Sample Data Augmentation Enhance 3D Object Detection with Radar Point Clouds?

No, of Course I Can! Deeper Fine-Tuning Attacks That Bypass Token-Level Safety Mechanisms

Disambiguate First, Parse Later: Generating Interpretations for Ambiguity Resolution in Semantic Parsing

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

B-cos LM: Efficiently Transforming Pre-trained Language Models for Improved Explainability

Brain Latent Progression: Individual-based Spatiotemporal Disease Progression on 3D Brain MRIs via Latent Diffusion

Integrating Generative Artificial Intelligence in ADRD: A Roadmap for Streamlining Diagnosis and Care in Neurodegenerative Diseases

Logits are All We Need to Adapt Closed Models

Oracular Programming: A Modular Foundation for Building LLM-Enabled Software

Learning Traffic Anomalies from Generative Models on Real-Time Observations

BiDepth: A Bidirectional-Depth Neural Network for Spatio-Temporal Prediction

Not all tokens are created equal: Perplexity Attention Weighted Networks for AI generated text detection

Prune 'n Predict: Optimizing LLM Decision-making with Conformal Prediction

Relation-aware Hierarchical Prompt for Open-vocabulary Scene Graph Generation

An Interoperable Machine Learning Pipeline for Pediatric Obesity Risk Estimation

BEExformer: A Fast Inferencing Binarized Transformer with Early Exits

VIVID-10M: A Dataset and Baseline for Versatile and Interactive Video Local Editing

Tiny-Align: Bridging Automatic Speech Recognition and Large Language Model on the Edge

DuSEGO: Dual Second-order Equivariant Graph Ordinary Differential Equation

Knowledge-Augmented Multimodal Clinical Rationale Generation for Disease Diagnosis with Small Language Models

No, of Course I Can! Deeper Fine-Tuning Attacks That Bypass Token-Level Safety Mechanisms

Created by

Haebom

저자

Joshua Kazdan, Abhay Puri, Rylan Schaeffer, Lisa Yu, Chris Cundy, Jason Stanley, Sanmi Koyejo, Krishnamurthy Dvijotham

개요

본 논문은 OpenAI와 Anthropic과 같은 주요 대규모 언어 모델(LM) 제공업체들이 특정 용도에 맞춰 최첨단 LM을 미세 조정할 수 있도록 허용하는 상황에서, 악용을 방지하기 위해 해당 제공업체들이 과도하게 유해한 데이터에 대한 미세 조정을 차단하는 필터를 적용하는 것에 대한 연구입니다. 기존의 안전 정렬이 "얕다"는 것을 보여주는 과거 연구와 마찬가지로, 기존의 미세 조정 공격 또한 "얕다"는 것을 보여줍니다. 즉, 공격은 모델 응답의 처음 몇 개 토큰만을 타겟으로 하며, 따라서 정렬된 모델로 처음 몇 개의 응답 토큰을 생성함으로써 차단될 수 있다는 것입니다. 본 논문에서는 먼저 유해한 요청을 거부한 다음 응답하는 새로운 미세 조정 공격을 제시하여 공격을 더욱 심화시키는 방법을 개념적으로 설명합니다. 이 "거부 후 준수" 전략은 얕은 방어를 우회하고 출력 필터를 회피하는 유해한 응답을 생성합니다. 마지막으로 방어 기능이 장착된 오픈소스 모델과 프로덕션 모델 모두에 대한 새로운 미세 조정 공격의 효능을 보여줍니다. GPT-4o와 Claude Haiku에 대해 각각 57%와 72%의 공격 성공률을 달성하였으며, OpenAI로부터 2000달러의 버그 바운티를 받았고 Anthropic에 의해 취약성으로 인정받았습니다.

시사점, 한계점

•

시사점: 모델이 처음에 유해한 요청을 거부하기 때문에 안전하다는 개념을 훼손하고, 프로덕션 미세 조정 API에 직면하는 공격의 범위에 대한 인식을 넓힙니다. 기존의 얕은 방어 메커니즘의 한계를 드러내고, 더욱 정교한 공격 기법의 필요성을 강조합니다. "거부 후 준수" 전략과 같은 새로운 공격 방식의 위험성을 제시합니다.

•

한계점: 제시된 공격 기법의 일반화 가능성 및 다른 모델에 대한 적용 가능성에 대한 추가 연구가 필요합니다. 현재 공격 성공률은 특정 모델에 국한될 수 있으며, 모델 업데이트나 방어 메커니즘 개선에 따라 성공률이 달라질 수 있습니다. 모든 유형의 유해한 요청과 응답에 대해 효과적이지 않을 수 있습니다.

Made with Slashpage