Daily Arxiv

This page organizes papers related to artificial intelligence published around the world.
This page is summarized using Google Gemini and is operated on a non-profit basis.
The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.

A Convex Route to Thermomechanics: Learning Internal Energy and Dissipation

ResAdapt: Adaptive Resolution for Efficient Multimodal Reasoning

Detection of Adversarial Attacks in Robotic Perception

RAD-LAD: Rule and Language Grounded Autonomous Driving in Real-Time

LG-HCC: Local Geometry-Aware Hierarchical Context Compression for 3D Gaussian Splatting

Building evidence-based knowledge graphs from full-text literature for disease-specific biomedical reasoning

JaWildText: A Benchmark for Vision-Language Models on Japanese Scene Text Understanding

ITQ3_S: High-Fidelity 3-bit LLM Inference via Interleaved Ternary Quantization with Rotation-Domain Smoothing

Heracles: Bridging Precise Tracking and Generative Synthesis for General Humanoid Control

Cross-attentive Cohesive Subgraph Embedding to Mitigate Oversquashing in GNNs

Magic Words or Methodical Work? Challenging Conventional Wisdom in LLM-Based Political Text Annotation

SleepVLM: Explainable and Rule-Grounded Sleep Staging via a Vision-Language Model

Pseudo Label NCF for Sparse OHC Recommendation: Dual Representation Learning and the Separability Accuracy Trade off

Enes Causal Discovery

Cost-Sensitive Neighborhood Aggregation for Heterophilous Graphs: When Does Per-Edge Routing Help?

Robust Safety Monitoring of Language Models via Activation Watermarking

KARMA: Knowledge-Action Regularized Multimodal Alignment for Personalized Search at Taobao

LLMON: An LLM-native Markup Language to Leverage Structure and Semantics at the LLM Interface

Beyond Hard Constraints: Budget-Conditioned Reachability For Safe Offline Reinforcement Learning

LPNSR: Prior-Enhanced Diffusion Image Super-Resolution via LR-Guided Noise Prediction

ContractSkill: Repairable Contract-Based Skills for Multimodal Web Agents

X-World: Controllable Ego-Centric Multi-Camera World Models for Scalable End-to-End Driving

FedRG: Unleashing the Representation Geometry for Federated Learning with Noisy Clients

Inducing Sustained Creativity and Diversity in Large Language Models

Points-to-3D: Structure-Aware 3D Generation with Point Cloud Priors

How to do LLMs Compute Verbal Confidence

InCoder-32B: Code Foundation Model for Industrial Scenarios

100x Cost & Latency Reduction: Performance Analysis of AI Query Approximation using Lightweight Proxy Models

Sample-Efficient Hypergradient Estimation for Decentralized Bi-Level Reinforcement Learning

Real-Time Driver Safety Scoring Through Inverse Crash Probability Modeling

AgentDrift: Unsafe Recommendation Drift Under Tool Corruption Hidden by Ranking Metrics in LLM Agents

Multi-Agent Memory from a Computer Architecture Perspective: Visions and Challenges Ahead

Not All News Is Equal: Topic- and Event-Conditional Sentiment from Finetuned LLMs for Aluminum Price Forecasting

When Rubrics Fail: Error Enumeration as Reward in Reference-Free RL Post-Training for Virtual Try-On

Training for Technology: Adoption and Productive Use of Generative AI in Legal Analysis

When Metrics Disagree: Automatic Similarity vs. LLM-as-a-Judge for Clinical Dialogue Evaluation

Evidential Neural Radiance Fields

Mitigating “Epistemic Debt” in Generative AI-Scaffolded Novice Programming using Metacognitive Scripts

DGPO: RL-Steered Graph Diffusion for Neural Architecture Generation

Understanding vs. Generation: Navigating Optimization Dilemma in Multimodal Models

How to Train Your Long-Context Visual Document Model

When Test-Time Guidance Is Enough: Fast Image and Video Editing with Diffusion Guidance

Semantic Labeling for Third-Party Cybersecurity Risk Assessment: A Semi-Supervised Approach to Intent-Aware Question Retrieval

$V_0$: A Generalist Value Model for Any Policy at State Zero

PAIR-Former: Budgeted Relational MIL for miRNA Target Prediction

Temporal Sepsis Modeling: a Relational and Explainable-by-Design Framework

Dynamic Cogeneration of Bug Reproduction Test in Agentic Program Repair

The Mouth is Not the Brain: Bridging Energy-Based World Models and Language Generation

Hellinger Multimodal Variational Autoencoders

Merging Triggers, Breaking Backdoors: Defensive Poisoning for Instruction-Tuned Language Models

LeLaR: The First In-Orbit Demonstration of an AI-Based Satellite Attitude Controller

Provably Extracting the Features from a General Superposition

Stronger Normalization-Free Transformers

InfiniteVL: Synergizing Linear and Sparse Attention for Highly-Efficient, Unlimited-Input Vision-Language Models

A Systematic Framework for Enterprise Knowledge Retrieval: Leveraging LLM - Generated Metadata to Enhance RAG Systems

VLA Models Are More Generalizable Than You Think: Revisiting Physical and Spatial Modeling

ReAG: Reasoning-Augmented Generation for Knowledge-based Visual Question Answering

Masked IRL: LLM-Guided Reward Disambiguation from Demonstrations and Language

Towards High-Consistency Embodied World Model with Multi-View Trajectory Videos

EchoMark: Perceptual Acoustic Environment Transfer with Watermark-Embedded Room Impulse Response

ZeroFlood: Flood Hazard Mapping from Single-Modality SAR Using Geo-Foundation Models

Automated Algorithm Design for Auto-Tuning Optimizers

MA-SAPO: Multi-Agent Reasoning for Score-Aware Prompt Optimization

Zero-Shot Coordination in Ad Hoc Teams with Generalized Policy Improvement and Difference Rewards

Local Causal Discovery for Statistically Efficient Causal Inference

ShishuLM: Achieving Optimal and Efficient Parameterization with Low Attention Transformer Models

A Semi-amortized Lifted Learning-to-Optimize Masked (SALLO-M) Transformer Model for Scalable and Generalizable Beamforming

ARROW: An Adaptive Rollout and Routing Method for Global Weather Forecasting

Past, Present, and Future of Bug Tracking in the Generative AI Era

TransFIRA: Transfer Learning for Face Image Recognizability Assessment

REN: Anatomically-Informed Mixture-of-Experts for Interstitial Lung Disease Diagnosis

Expressive Power of Implicit Models: Rich Equilibria and Test-Time Scaling

Align Your Query: Representation Alignment for Multimodality Medical Object Detection

Learning Inter-Atomic Potentials without Explicit Equivariance

MSG: Multi-Stream Generative Policies for Sample-Efficient Robotic Manipulation

Semantic Voting: A Self-Evaluation-Free Approach for Efficient LLM Self-Improvement on Unverifiable Open-ended Tasks

SecureVibeBench: Evaluating Secure Coding Capabilities of Code Agents with Realistic Vulnerability Scenarios

Incorporating LLM Embeddings for Variation Across the Human Genome

Generative AI on Wall Street -- Opportunities and Risk Controls

Improving Liver Disease Diagnosis with SNNDeep: A Custom Spiking Neural Network Using Diverse Learning Algorithms

Multi-Level Knowledge Distillation and Dynamic Self-Supervised Learning for Continual Learning

TTA-DAME: Test-Time Adaptation with Domain Augmentation and Model Ensemble for Dynamic Driving Conditions

Generative Logic: A New Computer Architecture for Deterministic Reasoning and Knowledge Generation

QuestA: Expanding Reasoning Capacity in LLMs via Question Augmentation

Streaming 4D Visual Geometry Transformer

LaSM: Layer-wise Scaling Mechanism for Defending Pop-up Attack on GUI Agents

Accelerating Diffusion Large Language Models with SlowFast Sampling: The Three Golden Principles

AVA-Bench: Atomic Visual Ability Benchmark for Vision Foundation Models

Denoising the Future: Top-p Distributions for Moving Through Time

FA-INR: Adaptive Implicit Neural Representations for Interpretable Exploration of Simulation Ensembles

AI-Generated Compromises for Coalition Formation

Balancing Efficiency and Empathy: Healthcare Providers' Perspectives on AI-Supported Workflows for Serious Illness Conversations in the Emergency Department

LLM-Meta-SR: In-Context Learning for Evolving Selection Operators in Symbolic Regression

ProFashion: Prototype-guided Fashion Video Generation with Multiple Reference Images

Aleph-Alpha-GermanWeb: Improving German-language LLM pre-training with model-based data curation and synthetic data generation

We'll Fix it in Post: Improving Text-to-Video Generation with Neuro-Symbolic Feedback

EventChat: Implementation and user-centric evaluation of a large language model-driven conversational recommender system for exploring leisure events in an SME context

SemioLLM: Evaluating Large Language Models for Diagnostic Reasoning from Unstructured Clinical Narratives in Epilepsy

GenOL: Generating Diverse Examples for Name-only Online Learning

Early Exiting Predictive Coding Neural Networks for Edge AI

LoRA Users Beware: A Few Spurious Tokens Can Manipulate Your Finetuned Model

Created by

Haebom

저자

Marcel Mateos Salles, Praney Goyal, Pradyut Sekhsaria, Hai Huang, Randall Balestriero

개요

LoRA를 활용한 LLM 미세 조정을 할 때, LoRA의 자원 효율성이 높을수록 모델이 SSTI 공격에 취약해진다는 것을 발견했습니다. SSTI는 미세 조정 중 단일 토큰을 주입하여 테스트 시 모델 예측을 조작할 수 있게 합니다. 다양한 모델과 데이터셋을 사용하여 SSTI의 영향을 평가하고, 기존의 방어 방법으로는 이 공격을 방어할 수 없음을 확인했습니다.

시사점, 한계점

•

LoRA를 사용한 LLM은 SSTI 공격에 취약하며, LoRA의 자원 효율성이 높을수록 취약성이 증가합니다.

•

SSTI 공격은 미세 조정 중 단일 토큰 주입만으로 모델의 예측을 조작할 수 있습니다.

•

기존의 데이터 검증 도구나 전처리 방법으로는 SSTI 공격을 방어할 수 없습니다.

•

데이터 품질 및 AI 안전성에 대한 새로운 우려를 제기합니다.

PDF 보기

Made with Slashpage