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.

GLM-4.1V-Thinking: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning

Horus: A Protocol for Trustless Delegation Under Uncertainty

Reducing Variability of Multiple Instance Learning Methods for Digital Pathology

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

DICE-BENCH: Evaluating the Tool-Use Capabilities of Large Language Models in Multi-Round, Multi-Party Dialogues

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

BioPars: A Pretrained Biomedical Large Language Model for Persian Biomedical Text Mining

TRACED: Transition-aware Regret Approximation with Co-learnability for Environment Design

Towards Safety Evaluations of Theory of Mind in Large Language Models

Fair Algorithms with Probing for Multi-Agent Multi-Armed Bandits

GraphGSOcc: Semantic-Geometric Graph Transformer with Dynamic-Static Decoupling for 3D Gaussian Splatting-based Occupancy Prediction

Screen Hijack: Visual Poisoning of VLM Agents in Mobile Environments

15,500 Seconds: Lean UAV Classification Leveraging PEFT and Pre-Trained Networks

Tightly-Coupled LiDAR-IMU-Leg Odometry with Online Learned Leg Kinematics Incorporating Foot Tactile Information

Time Series Representations for Classification Lie Hidden in Pretrained Vision Transformers

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

On the Fundamental Impossibility of Hallucination Control in Large Language Models

Adapting Rule Representation With Four-Parameter Beta Distribution for Learning Classifier Systems

Real-Time Blind Defocus Deblurring for Earth Observation: The IMAGIN-e Mission Approach

Scaling Up Liquid-Resistance Liquid-Capacitance Networks for Efficient Sequence Modeling

Self-reflective Uncertainties: Do LLMs Know Their Internal Answer Distribution?

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

Pre-training Large Memory Language Models with Internal and External Knowledge

Towards Universal Semantics With Large Language Models

Aitomia: Your Intelligent Assistant for AI-Driven Atomistic and Quantum Chemical Simulations

Enhancing Robustness to Missing Modalities through Clustered Federated Learning

Perceiving Beyond Language Priors: Enhancing Visual Comprehension and Attention in Multimodal Models

LZ Penalty: An information-theoretic repetition penalty for autoregressive language models

Towards Cardiac MRI Foundation Models: Comprehensive Visual-Tabular Representations for Whole-Heart Assessment and Beyond

TARO: Timestep-Adaptive Representation Alignment with Onset-Aware Conditioning for Synchronized Video-to-Audio Synthesis

Recursive Training Loops in LLMs: How training data properties modulate distribution shift in generated data?

Real-is-Sim: Bridging the Sim-to-Real Gap with a Dynamic Digital Twin

Concat-ID: Towards Universal Identity-Preserving Video Synthesis

How Metacognitive Architectures Remember Their Own Thoughts: A Systematic Review

SFO: Piloting VLM Feedback for Offline RL

Towards Efficient Educational Chatbots: Benchmarking RAG Frameworks

KatFishNet: Detecting LLM-Generated Korean Text through Linguistic Feature Analysis

Distribution Matching for Self-Supervised Transfer Learning

A Baseline Method for Removing Invisible Image Watermarks using Deep Image Prior

SKIL: Semantic Keypoint Imitation Learning for Generalizable Data-efficient Manipulation

AirRadar: Inferring Nationwide Air Quality in China with Deep Neural Networks

A Framework for Mining Collectively-Behaving Bots in MMORPGs

Continual Learning with Strategic Selection and Forgetting for Network Intrusion Detection

A Survey on Uncertainty Quantification of Large Language Models: Taxonomy, Open Research Challenges, and Future Directions

A Review of Bayesian Uncertainty Quantification in Deep Probabilistic Image Segmentation

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

There and Back Again: On the relation between Noise and Image Inversions in Diffusion Models

Unified Triplet-Level Hallucination Evaluation for Large Vision-Language Models

Contrastive Learning and Adversarial Disentanglement for Privacy-Aware Task-Oriented Semantic Communication

Unsupervised Panoptic Interpretation of Latent Spaces in GANs Using Space-Filling Vector Quantization

NegMerge: Sign-Consensual Weight Merging for Machine Unlearning

Drug Discovery SMILES-to-Pharmacokinetics Diffusion Models with Deep Molecular Understanding

Sublinear Regret for a Class of Continuous-Time Linear-Quadratic Reinforcement Learning Problems

Backdooring Bias (B^2) into Stable Diffusion Models

Embodied Instruction Following in Unknown Environments

Improving Consistency Models with Generator-Augmented Flows

OralBBNet: Spatially Guided Dental Segmentation of Panoramic X-Rays with Bounding Box Priors

Divergent Creativity in Humans and Large Language Models

SpikeNAS: A Fast Memory-Aware Neural Architecture Search Framework for Spiking Neural Network-based Embedded AI Systems

Squat: Quant Small Language Models on the Edge

Dataset Distillation via the Wasserstein Metric

The Boolean Solution Problem from the Perspective of Predicate Logic -- Extended Version

Can Large Language Models Develop Strategic Reasoning? Post-training Insights from Learning Chess

World-aware Planning Narratives Enhance Large Vision-Language Model Planner

Reasoning about Uncertainty: Do Reasoning Models Know When They Don't Know?

MMLU-Reason: Benchmarking Multi-Task Multi-modal Language Understanding and Reasoning

Adapting Probabilistic Risk Assessment for AI

Beating Transformers using Synthetic Cognition

MedAgent-Pro: Towards Evidence-based Multi-modal Medical Diagnosis via Reasoning Agentic Workflow

Using Large Language Models to Categorize Strategic Situations and Decipher Motivations Behind Human Behaviors

ChartCoder: Advancing Multimodal Large Language Model for Chart-to-Code Generation

MCCoder: Streamlining Motion Control with LLM-Assisted Code Generation and Rigorous Verification

DREAMS: A python framework for Training Deep Learning Models on EEG Data with Model Card Reporting for Medical Applications

Human Mobility Modeling with Household Coordination Activities under Limited Information via Retrieval-Augmented LLMs

Combating Confirmation Bias: A Unified Pseudo-Labeling Framework for Entity Alignment

AC-DiT: Adaptive Coordination Diffusion Transformer for Mobile Manipulation

Locality-aware Parallel Decoding for Efficient Autoregressive Image Generation

How Well Does GPT-4o Understand Vision? Evaluating Multimodal Foundation Models on Standard Computer Vision Tasks

SpecCLIP: Aligning and Translating Spectroscopic Measurements for Stars

Adaptability of ASR Models on Low-Resource Language: A Comparative Study of Whisper and Wav2Vec-BERT on Bangla

Exploring a Hybrid Deep Learning Approach for Anomaly Detection in Mental Healthcare Provider Billing: Addressing Label Scarcity through Semi-Supervised Anomaly Detection

End-to-End Large Portfolio Optimization for Variance Minimization with Neural Networks through Covariance Cleaning

Gradient-Adaptive Policy Optimization: Towards Multi-Objective Alignment of Large Language Models

AI4Research: A Survey of Artificial Intelligence for Scientific Research

Towards Foundation Auto-Encoders for Time-Series Anomaly Detection

Bridging UI Design and chatbot Interactions: Applying Form-Based Principles to Conversational Agents

mGRADE: Minimal Recurrent Gating Meets Delay Convolutions for Lightweight Sequence Modeling

MILP-SAT-GNN: Yet Another Neural SAT Solver

Empowering Manufacturers with Privacy-Preserving AI Tools: A Case Study in Privacy-Preserving Machine Learning to Solve Real-World Problems

LoRA Fine-Tuning Without GPUs: A CPU-Efficient Meta-Generation Framework for LLMs

How Do Vision-Language Models Process Conflicting Information Across Modalities?

Are Vision Transformer Representations Semantically Meaningful? A Case Study in Medical Imaging

Probing Evaluation Awareness of Language Models

MuRating: A High Quality Data Selecting Approach to Multilingual Large Language Model Pretraining

BranchNet: A Neuro-Symbolic Learning Framework for Structured Multi-Class Classification

GPU-based complete search for nonlinear minimization subject to bounds

Enhanced Generative Model Evaluation with Clipped Density and Coverage

Tuning without Peeking: Provable Privacy and Generalization Bounds for LLM Post-Training

ECCV 2024 W-CODA: 1st Workshop on Multimodal Perception and Comprehension of Corner Cases in Autonomous Driving

Towards culturally-appropriate conversational AI for health in the majority world: An exploratory study with citizens and professionals in Latin America

Reasoning as an Adaptive Defense for Safety

Created by

Haebom

저자

Taeyoun Kim, Fahim Tajwar, Aditi Raghunathan, Aviral Kumar

개요

본 논문은 수학 및 코드와 같이 검증이 용이한 영역에서 LLM의 성능을 향상시키는 적응적 테스트 시간 컴퓨팅 할당 추론 방법을 안전 취약성에 대한 강건성을 갖는 모델을 훈련하는 데 활용하는 방법을 연구합니다. 연구진은 안전과 작업 완료 간의 균형을 맞추는 보상 신호와 사고 과정 추적을 사용하여 모델이 안전에 대해 추론하도록 훈련하는 강화 학습(RL) 접근 방식인 $\textit{TARS}$(Training Adaptive Reasoners for Safety)라는 레시피를 제시합니다. TARS는 (1) "경량" 웜스타트 SFT 단계, (2) 지나친 거부와 같은 지름길 행동을 방지하기 위한 유해, 무해, 모호한 프롬프트의 혼합, (3) 훈련 중 추론 능력의 퇴화를 방지하기 위한 보상 함수라는 세 가지 중요한 설계 선택을 포함합니다. TARS로 훈련된 모델은 모호한 쿼리에 더 많은 컴퓨팅 리소스를 사용하여 적응적 동작을 보이며, 안전 거부 간의 더 나은 절충안을 제공합니다. 또한 안전한 프롬프트와 안전하지 않은 프롬프트를 더 잘 구분하고 화이트박스(예: GCG) 및 블랙박스 공격(예: PAIR) 모두에 대한 더 큰 강건성을 달성합니다. 전반적으로 이 연구는 프롬프트별 추론을 통해 탈옥 및 유해한 요청으로부터 LLM을 훈련하는 효과적이고 공개적인 레시피를 제공합니다.

시사점, 한계점

•

시사점:

◦

적응적 컴퓨팅 할당 추론 방법을 활용하여 LLM의 안전 취약성에 대한 강건성을 향상시키는 효과적인 방법 제시.

◦

$\textit{TARS}$ 레시피를 통해 안전과 작업 완료 간의 균형을 맞추는 훈련 전략 제시.

◦

모호한 쿼리에 대한 적응적 컴퓨팅 리소스 할당을 통한 안전-거부 절충안 개선.

◦

화이트박스 및 블랙박스 공격에 대한 강건성 향상.

◦

안전한 프롬프트와 안전하지 않은 프롬프트 구분 능력 향상.

◦

LLM의 탈옥 및 유해한 요청에 대한 방어력 향상을 위한 공개적인 레시피 제공.

•

한계점:

◦

$\textit{TARS}$ 레시피의 일반화 성능 및 다양한 LLM 아키텍처에 대한 적용 가능성에 대한 추가 연구 필요.

◦

보상 함수 설계 및 튜닝의 어려움.

◦

다양한 유형의 안전 취약성에 대한 포괄적인 평가 부족.

◦

실제 환경 적용 시 발생 가능한 문제점 및 한계에 대한 추가 연구 필요.

Made with Slashpage