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.

Distillation Robustifies Unlearning

Cartridges: Lightweight and general-purpose long context representations via self-study

Text-to-LoRA: Instant Transformer Adaption

Information Bargaining: Bilateral Commitment in Bayesian Persuasion

Cross-lingual Collapse: How Language-Centric Foundation Models Shape Reasoning in Large Language Models

Heartcare Suite: Multi-dimensional Understanding of ECG with Raw Multi-lead Signal Modeling

Peer-Ranked Precision: Creating a Foundational Dataset for Fine-Tuning Vision Models from DataSeeds' Annotated Imagery

TissUnet: Improved Extracranial Tissue and Cranium Segmentation for Children through Adulthood

A Red Teaming Roadmap Towards System-Level Safety

Reason-to-Recommend: Using Interaction-of-Thought Reasoning to Enhance LLM Recommendation

Context Is Not Comprehension

Feature-Based Lie Group Transformer for Real-World Applications

Horizon Reduction Makes RL Scalable

SLAC: Simulation-Pretrained Latent Action Space for Whole-Body Real-World RL

A Diffusion-Driven Temporal Super-Resolution and Spatial Consistency Enhancement Framework for 4D MRI imaging

BiMa: Towards Biases Mitigation for Text-Video Retrieval via Scene Element Guidance

Retrieval-Augmented Generation as Noisy In-Context Learning: A Unified Theory and Risk Bounds

Deep Learning for Retinal Degeneration Assessment: A Comprehensive Analysis of the MARIO AMD Progression Challenge

CoT is Not True Reasoning, It Is Just a Tight Constraint to Imitate: A Theory Perspective

Rethinking the effects of data contamination in Code Intelligence

MINT: Multimodal Instruction Tuning with Multimodal Interaction Grouping

Protap: A Benchmark for Protein Modeling on Realistic Downstream Applications

NTPP: Generative Speech Language Modeling for Dual-Channel Spoken Dialogue via Next-Token-Pair Prediction

SATA-BENCH: Select All That Apply Benchmark for Multiple Choice Questions

Learning from Double Positive and Unlabeled Data for Potential-Customer Identification

Aligned but Blind: Alignment Increases Implicit Bias by Reducing Awareness of Race

Diversity of Transformer Layers: One Aspect of Parameter Scaling Laws

Noise-Robustness Through Noise: Asymmetric LoRA Adaption with Poisoning Expert

Large Language Models Often Know When They Are Being Evaluated

WorkForceAgent-R1: Incentivizing Reasoning Capability in LLM-based Web Agents via Reinforcement Learning

CAST: Contrastive Adaptation and Distillation for Semi-Supervised Instance Segmentation

PartInstruct: Part-level Instruction Following for Fine-grained Robot Manipulation

RainFusion: Adaptive Video Generation Acceleration via Multi-Dimensional Visual Redundancy

VeriThoughts: Enabling Automated Verilog Code Generation using Reasoning and Formal Verification

Enigmata: Scaling Logical Reasoning in Large Language Models with Synthetic Verifiable Puzzles

APE: Selective Fine-tuning with Acceptance Criteria for Language Model Adaptation

When Two LLMs Debate, Both Think They'll Win

Turb-L1: Achieving Long-term Turbulence Tracing By Tackling Spectral Bias

GRE Suite: Geo-localization Inference via Fine-Tuned Vision-Language Models and Enhanced Reasoning Chains

Can MLLMs Guide Me Home? A Benchmark Study on Fine-Grained Visual Reasoning from Transit Maps

Sample Complexity of Diffusion Model Training Without Empirical Risk Minimizer Access

EVADE: Multimodal Benchmark for Evasive Content Detection in E-Commerce Applications

Simulating Macroeconomic Expectations using LLM Agents

Mixture of Decoding: An Attention-Inspired Adaptive Decoding Strategy to Mitigate Hallucinations in Large Vision-Language Models

Mechanistic evaluation of Transformers and state space models

Toward Reliable Scientific Hypothesis Generation: Evaluating Truthfulness and Hallucination in Large Language Models

Pel, A Programming Language for Orchestrating AI Agents

MARVEL: Multi-Agent RTL Vulnerability Extraction using Large Language Models

Learning Pareto-Optimal Rewards from Noisy Preferences: A Framework for Multi-Objective Inverse Reinforcement Learning

Q-Policy: Quantum-Enhanced Policy Evaluation for Scalable Reinforcement Learning

BLEUBERI: BLEU is a surprisingly effective reward for instruction following

Position: We Need Responsible, Application-Driven (RAD) AI Research

Restoring Calibration for Aligned Large Language Models: A Calibration-Aware Fine-Tuning Approach

LookAlike: Consistent Distractor Generation in Math MCQs

Tree-Sliced Wasserstein Distance with Nonlinear Projection

Test-time Correlation Alignment

Pushing the Limits of Low-Bit Optimizers: A Focus on EMA Dynamics

OpenTCM: A GraphRAG-Empowered LLM-based System for Traditional Chinese Medicine Knowledge Retrieval and Diagnosis

A Comprehensive Survey in LLM(-Agent) Full Stack Safety: Data, Training and Deployment

MIB: A Mechanistic Interpretability Benchmark

Bipartite Ranking From Multiple Labels: On Loss Versus Label Aggregation

LLM-SRBench: A New Benchmark for Scientific Equation Discovery with Large Language Models

Toward Total Recall: Enhancing FAIRness through AI-Driven Metadata Standardization

Finding Interest Needle in Popularity Haystack: Improving Retrieval by Modeling Item Exposure

Geometrical Properties of Text Token Embeddings for Strong Semantic Binding in Text-to-Image Generation

sudo rm -rf agentic_security

Imagine to Hear: Auditory Knowledge Generation can be an Effective Assistant for Language Models

Towards Achieving Perfect Multimodal Alignment

nvBench 2.0: Resolving Ambiguity in Text-to-Visualization through Stepwise Reasoning

FedALT: Federated Fine-Tuning through Adaptive Local Training with Rest-of-World LoRA

RONA: Pragmatically Diverse Image Captioning with Coherence Relations

Unifying 2D and 3D Vision-Language Understanding

Revisiting semi-supervised learning in the era of foundation models

AI-based Framework for Robust Model-Based Connector Mating in Robotic Wire Harness Installation

Generalized Interpolating Discrete Diffusion

Towards Autonomous Reinforcement Learning for Real-World Robotic Manipulation with Large Language Models

Straight-Line Diffusion Model for Efficient 3D Molecular Generation

Examining the Mental Health Impact of Misinformation on Social Media Using a Hybrid Transformer-Based Approach

Dynamic spillovers and investment strategies across artificial intelligence ETFs, artificial intelligence tokens, and green markets

Dialogue Without Limits: Constant-Sized KV Caches for Extended Responses in LLMs

LLaSE-G1: Incentivizing Generalization Capability for LLaMA-based Speech Enhancement

from Benign import Toxic: Jailbreaking the Language Model via Adversarial Metaphors

EgoNormia: Benchmarking Physical Social Norm Understanding

PhantomWiki: On-Demand Datasets for Reasoning and Retrieval Evaluation

NeoBERT: A Next-Generation BERT

From Offline to Online Memory-Free and Task-Free Continual Learning via Fine-Grained Hypergradients

AMPO: Active Multi-Preference Optimization for Self-play Preference Selection

SYNTHIA: Novel Concept Design with Affordance Composition

DBudgetKV: Dynamic Budget in KV Cache Compression for Ensuring Optimal Performance

AlphaAgent: LLM-Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay

DISC: DISC: Dynamic Decomposition Improves LLM Inference Scaling

Predicting Bad Goods Risk Scores with ARIMA Time Series: A Novel Risk Assessment Approach

Space-O-RAN: Enabling Intelligent, Open, and Interoperable Non Terrestrial Networks in 6G

MindLLM: A Subject-Agnostic and Versatile Model for fMRI-to-Text Decoding

Improving the Diffusability of Autoencoders

Tree-of-Debate: Multi-Persona Debate Trees Elicit Critical Thinking for Scientific Comparative Analysis

MMTEB: Massive Multilingual Text Embedding Benchmark

From Sub-Ability Diagnosis to Human-Aligned Generation: Bridging the Gap for Text Length Control via MARKERGEN

How Expressive are Knowledge Graph Foundation Models?

Machine Learning Should Maximize Welfare, but Not by (Only) Maximizing Accuracy

A Red Teaming Roadmap Towards System-Level Safety

Created by

Haebom

저자

Zifan Wang, Christina Q. Knight, Jeremy Kritz, Willow E. Primack, Julian Michael

개요

본 논문은 대규모 언어 모델(LLM)의 안전장치, 특히 요청 거부 기능에 대한 적대적 머신러닝 및 AI 안전 관점에서의 레드팀 테스트 연구의 현황과 문제점을 분석합니다. 기존의 많은 LLM 레드팀 테스트 연구가 추상적인 사회적 편향이나 윤리적 원칙보다는 명확한 제품 안전 규격에 대한 테스트를 우선시해야 하며, 현실적인 위협 모델을 기반으로 해야 한다고 주장합니다. 또한, 배포 환경에서의 시스템 수준 안전성 확보를 레드팀 연구의 다음 단계로 제시하며, AI 모델이 새로운 위협뿐 아니라 악의적 사용자 탐지 및 차단과 같은 위협 완화 기능도 제공할 수 있음을 강조합니다. 즉, 빠르게 발전하는 AI 기술이 제기하는 새로운 위협에 효과적으로 대응하기 위해서는 레드팀 연구의 우선순위를 재정립해야 한다는 주장을 펼칩니다.

시사점, 한계점

•

시사점:

◦

LLM 안전장치의 효과적인 평가를 위해 명확한 제품 안전 규격 기반의 테스트가 중요함을 강조.

◦

현실적인 위협 모델을 고려한 레드팀 테스트의 필요성 제시.

◦

시스템 수준의 안전성 확보를 위한 연구 방향 제시 (악의적 사용자 탐지 및 차단 등).

◦

AI 기술 발전에 따른 새로운 위협에 대응하기 위한 레드팀 연구의 우선순위 재정립 필요성 제기.

•

한계점:

◦

기존 연구들의 구체적인 문제점과 사례 제시 부족.

◦

제시된 시스템 수준 안전성 확보 방안에 대한 구체적인 설명 부족.

◦

현실적인 위협 모델에 대한 구체적인 정의 및 예시 부족.

Made with Slashpage