Daily Arxiv

世界中で発行される人工知能関連の論文をまとめるページです。
このページはGoogle Geminiを活用して要約し、非営利で運営しています。
論文の著作権は著者および関連機関にあり、共有する際は出典を明記してください。

Dehazing Light Microscopy Images with Guided Conditional Flow Matching: finding a sweet spot between fidelity and realism

EFRame: Deeper Reasoning via Exploration-Filter-Replay Reinforcement Learning Framework

Refine-POI: Reinforcement Fine-Tuned Large Language Models for Next Point-of-Interest Recommendation

HalluSegBench: Counterfactual Visual Reasoning for Segmentation Hallucination Evaluation

Potemkin Understanding in Large Language Models

OmniEval: A Benchmark for Evaluating Omni-modal Models with Visual, Auditory, and Textual Inputs

How to Retrieve Examples in In-context Learning to Improve Conversational Emotion Recognition using Large Language Models?

Position: Machine Learning Conferences Should Establish a "Refutations and Critiques" Track

Arabic Dialect Classification using RNNs, Transformers, and Large Language Models: A Comparative Analysis

Improving Student-AI Interaction Through Pedagogical Prompting: An Example in Computer Science Education

GLIMPSE: Gradient-Layer Importance Mapping for Prompted Visual Saliency Explanation for Generative LVLMs

Automatic Depression Assessment using Machine Learning: A Comprehensive Survey

Generalizing vision-language models to novel domains: A comprehensive survey

Comparative Evaluation of ChatGPT and DeepSeek Across Key NLP Tasks: Strengths, Weaknesses, and Domain-Specific Performance

AI-Generated Song Detection via Lyrics Transcripts

KAG-Thinker: Interactive Thinking and Deep Reasoning in LLMs via Knowledge-Augmented Generation

Data Quality Issues in Multilingual Speech Datasets: The Need for Sociolinguistic Awareness and Proactive Language Planning

Double Entendre: Robust Audio-Based AI-Generated Lyrics Detection via Multi-View Fusion

Aligning Evaluation with Clinical Priorities: Calibration, Label Shift, and Error Costs

Value-Free Policy Optimization via Reward Partitioning

VFEFL: Privacy-Preserving Federated Learning against Malicious Clients via Verifiable Functional Encryption

Enabling Precise Topic Alignment in Large Language Models Via Sparse Autoencoders

Robust LLM Unlearning with MUDMAN: Meta-Unlearning with Disruption Masking And Normalization

CMI-Bench: A Comprehensive Benchmark for Evaluating Music Instruction Following

StepProof: Step-by-step verification of natural language mathematical proofs

Scalable Non-Equivariant 3D Molecule Generation via Rotational Alignment

Improved Supervised Fine-Tuning for Large Language Models to Mitigate Catastrophic Forgetting

SLED: A Speculative LLM Decoding Framework for Efficient Edge Serving

FZOO: Fast Zeroth-Order Optimizer for Fine-Tuning Large Language Models towards Adam-Scale Speed

VeriLoC: Line-of-Code Level Prediction of Hardware Design Quality from Verilog Code

Multi Layered Autonomy and AI Ecologies in Robotic Art Installations

Bridging Subjective and Objective QoE: Operator-Level Aggregation Using LLM-Based Comment Analysis and Network MOS Comparison

Quantum computing and artificial intelligence: status and perspectives

Fine-Tuning Next-Scale Visual Autoregressive Models with Group Relative Policy Optimization

A Large Language Model-Enabled Control Architecture for Dynamic Resource Capability Exploration in Multi-Agent Manufacturing Systems

Spotlight-TTS: Spotlighting the Style via Voiced-Aware Style Extraction and Style Direction Adjustment for Expressive Text-to-Speech

WeatherEdit: Controllable Weather Editing with 4D Gaussian Field

From Alignment to Advancement: Bootstrapping Audio-Language Alignment with Synthetic Data

Error Optimization: Overcoming Exponential Signal Decay in Deep Predictive Coding Networks

TinyAlign: Boosting Lightweight Vision-Language Models by Mitigating Modal Alignment Bottlenecks

Super-Resolution Generative Adversarial Networks based Video Enhancement

Object detection in adverse weather conditions for autonomous vehicles using Instruct Pix2Pix

INSIGHT: Bridging the Student-Teacher Gap in Times of Large Language Models

SConU: Selective Conformal Uncertainty in Large Language Models

MetaSynth: Meta-Prompting-Driven Agentic Scaffolds for Diverse Synthetic Data Generation

Sculpting Memory: Multi-Concept Forgetting in Diffusion Models via Dynamic Mask and Concept-Aware Optimization

Achieving binary weight and activation for LLMs using Post-Training Quantization

A Consequentialist Critique of Binary Classification Evaluation Practices

Redefining Evaluation Standards: A Unified Framework for Evaluating the Korean Capabilities of Language Models

Test-Time Reasoning Through Visual Human Preferences with VLMs and Soft Rewards

FedMM-X: A Trustworthy and Interpretable Framework for Federated Multi-Modal Learning in Dynamic Environments

Automating Adjudication of Cardiovascular Events Using Large Language Models

ATTENTION2D: Communication Efficient Distributed Self-Attention Mechanism

Visual Position Prompt for MLLM ベースの Visual Grounding

Time-R1: Post-Training Large Vision Language Model for Temporal Video Grounding

Privacy Ethics Alignment in AI: A Stakeholder-Centric Framework for Ethical AI

Characterizing GPU Resilience and Impact on AI/HPC Systems

Explainable Sentiment Analysis with DeepSeek-R1: Performance, Efficiency, and Few-Shot Learning

Neurons: Emulating the Human Visual Cortex Improves Fidelity and Interpretability in fMRI-to-Video Reconstruction

The Problem of the Priors, or Posteriors?

Gumiho: A Hybrid Architecture to Prioritize Early Tokens in Speculative Decoding

Disrupting Model Merging: A Parameter-Level Defense Without Sacrificing Accuracy

What can large language models do for sustainable food?

Enough Coin Flips Can Make LLMs Act Bayesian

How to Move Your Dragon: Text-to-Motion Synthesis for Large-Vocabulary Objects

Time-MQA: Time Series Multi-Task Question Answering with Context Enhancement

PipeOffload: Improving Scalability of Pipeline Parallelism with Memory Optimization

Space-Time Graphs of Convex Sets for Multi-Robot Motion Planning

HalCECE: A Framework for Explainable Hallucination Detection through Conceptual Counterfactuals in Image Captioning

LNUCB-TA: Linear-nonlinear Hybrid Bandit Learning with Temporal Attention

No, of course I can! Refusal Mechanisms Can Be Exploited Using Harmless Fine-Tuning Data

Investigating the Impact of Quantization Methods on the Safety and Reliability of Large Language Models

Retrieval Augmented Generation Based LLM Evaluation For Protocol State Machine Inference With Chain-of-Thought Reasoning

A general language model for peptide identification

Cluster and Predict Latent Patches for Improved Masked Image Modeling

Semantic-Aware Adaptive Video Streaming Using Latent Diffusion Models for Wireless Networks

KMI: A Dataset of Korean Motivational Interviewing Dialogues for Psychotherapy

Mechanistic Interpretability of Emotion Inference in Large Language Models

Multimodal Medical Code Tokenizer

Time to Rethink AI for Combinatorial Optimization: Classical Algorithms Remain Tough to Match

Simultaneous Multi-Robot Motion Planning with Projected Diffusion Models

Environment-Driven Online LiDAR-Camera Extrinsic Calibration

Riddle Me This! Stealthy Membership Inference for Retrieval-Augmented Generation

DReSS: Data-driven Regularized Structured Streamlining for Large Language Models

Towards Automated Self-Supervised Learning for Truly Unsupervised Graph Anomaly Detection

Adaptive Rank Allocation for Federated Parameter-Efficient Fine-Tuning of Language Models

DisCoPatch: Taming Adversarially-driven Batch Statistics for Improved Out-of-Distribution Detection

An Investigation into Seasonal Variations in Energy Forecasting for Student Residences

Efficiently Serving Large Multimodal Models Using EPD Disaggregation

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

AlignGuard: Scalable Safety Alignment for Text-to-Image Generation

A Library for Learning Neural Operators

ZipAR: Parallel Auto-regressive Image Generation through Spatial Locality

Pretrained Reversible Generation as Unsupervised Visual Representation Learning

FLOAT: Generative Motion Latent Flow Matching for Audio-driven Talking Portrait

SEUF: Is Unlearning One Expert Enough for Mixture-of-Experts LLMs?

Recommender Systems for Good (RS4Good): Survey of Use Cases and a Call to Action for Research that Matters

Foundation Models for Wearable Movement Data in Mental Health Research

GenBFA: An Evolutionary Optimization Approach to Bit-Flip Attacks on LLMs

Enhancing Diffusion Posterior Sampling for Inverse Problems by Integrating Crafted Measurements

No, of course I can! Refusal Mechanisms Can Be Exploited Using Harmless Fine-Tuning Data

Created by

Haebom

作者

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

概要

この論文は、OpenAIやAnthropicなどの主要言語モデル（LM）プロバイダが特定の用途に合わせて最先端のLMを微調整できるようにする状況で、悪用を防ぐために、プロバイダが過度に有害なデータの微調整をブロックするフィルタを適用することに関する研究です。従来の研究では、安全アライメントが「浅い」ことを示したように、従来の微調整攻撃も浅いことを示しています。つまり、攻撃はモデル応答の最初の数個のトークンのみをターゲットにしているため、ソートされたモデルで最初の数個の応答トークンを生成してブロックできます。しかし、この論文では「拒否後のコンプライアンス」戦略を紹介し、攻撃をさらに深める方法を紹介します。この戦略は、最初に有害な要求を拒否し、次に応答することで、浅い防御を回避し、出力フィルタを回避する有害な応答を生成します。実験の結果、オープンソースモデルと商用モデルの両方で新しい微調整攻撃の効果が実証され、GPT-4oとClaude Haikuに対してそれぞれ57％と72％の攻撃成功率を達成しました。この研究は、OpenAIから2000ドルのバグバウンティを受け、アントロピックによって脆弱性として認められました。結論として、モデルが最初に有害な要求を拒否するという理由だけで安全だと考えるのは間違っていることを示し、運用中の微調整APIに直面する攻撃の範囲に対する認識を広げます。

Takeaways、Limitations

•

Takeaways:

◦

既存の安全整列および防御メカニズムの限界を明確に提示し、より深刻な攻撃技術の可能性を示すことによって、言語モデルの安全性を確保するための新しいアプローチの必要性を強調します。

◦

「拒否後の遵守」戦略による微調整攻撃の有効性を実証的に証明し、実際のサービス環境における安全脅威に対する警戒心を高めます。

◦

OpenAIやAnthropicなどの主要企業のモデルにも脆弱性が存在することを実証することで、より強力な安全装置の開発の重要性を強調しています。

•

Limitations：

◦

提示された「拒否後のコンプライアンス」攻撃戦略がすべてのタイプの言語モデルと防御メカニズムに有効であるかどうかをさらに調査する必要があります。

◦

攻撃成功率が100％ではないため、より洗練された攻撃技術が開発される可能性を排除できません。

◦

本研究は、特定の言語モデルと防御メカニズムの分析に限定されており、他のモデルやメカニズムの一般化の可能性は限られている可能性があります。

Made with Slashpage