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, 요약본 공유 시 출처만 명기하면 됩니다.
This service is supported by Google Gemini.

Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation

Stochastic Penalty-Barrier Methods for Constrained Machine Learning

MINTEval: Evaluating Memory under Multi-Target Interference in Long-Horizon Agent Systems

Prompt2Fingerprint: Plug-and-Play LLM Fingerprinting via Text-to-Weight Generation

What's Holding Back Latent Visual Reasoning?

A Simplex Witness Certificate for Constant Collapse in Variational Autoencoders

Balancing Knowledge Distillation for Imbalance Learning with Bilevel Optimization

DiagEval: Trajectory-Conditioned Diagnosis for Reliable Software Evaluation with GUI Agents

1GC-7RC: One Graphic Card -- Seven Research Challenges! How Good Are AI Agents at Doing Your Job?

Learning-Zone Energy: Online Data Selection for Efficient RL Post-Training

EfficientTDMPC: Improved MPC Objectives for Sample-Efficient Continuous Control

CHI-Bench: Can AI Agents Automate End-to-End, Long-Horizon, Policy-Rich Healthcare Workflows?

PrivScope: Task-scoped Disclosure Control for Hybrid Agentic Systems

SLEIGHT-Bench: A Benchmark of Evasion Attacks Against Agent Monitors

Nested Spatio-Temporal Time Series Forecasting

Membership Inference Attacks on Discrete Diffusion Language Models

StrLoRA: Towards Streaming Continual Visual Instruction Tuning for MLLMs

Argus: Evidence Assembly for Scalable Deep Research Agents

RoadmapBench: Evaluating Long-Horizon Agentic Software Development Across Version Upgrades

Pretraining Objective Matters in Extreme Low-Data FGVC: A Backbone-Controlled Study

DeltaPrompts: Escaping the Zero-Delta Trap in Multimodal Distillation

HoloMotion-1 Technical Report

VGGT-Edit: Feed-forward Native 3D Scene Editing with Residual Field Prediction

Dywave: Event-Aligned Dynamic Tokenization for Heterogeneous IoT Sensing Signals

Beyond Perplexity: A Geometric and Spectral Study of Low-Rank Pre-Training

Causality-Aware End-to-End Autonomous Driving via Ego-Centric Joint Scene Modeling

Protocol-Driven Development: Governing Generated Software Through Invariants and Continuous Evidence

VECTOR-Drive: Tightly Coupled Vision-Language and Trajectory Expert Routing for End-to-End Autonomous Driving

VT-Bench: A Unified Benchmark for Visual-Tabular Multi-Modal Learning

HoReN: Normalized Hopfield Retrieval for Large-Scale Sequential Model Editing

Flow-OPD: On-Policy Distillation for Flow Matching Models

Memory-Efficient Looped Transformer: Decoupling Compute from Memory in Looped Language Models

RELO: Reinforcement Learning to Localize for Visual Object Tracking

R$^3$L: Reasoning 3D Layouts from Relative Spatial Relations

PragLocker: Protecting Agent Intellectual Property in Untrusted Deployments via Non-Portable Prompts

GRALIS: A Unified Canonical Framework for Linear Attribution Methods via Riesz Representation

Skill Neologisms: Towards Skill-based Continual Learning

GEASS: Gated Evidence-Adaptive Selective Caption Trust for Vision-Language Models

One-Block Transformer (1BT) for EEG-Based Cognitive Workload Assessment

Addressing the Reality Gap: A Three-Tension Framework for Agentic AI Adoption

Designing escalation criteria for international AI incident response: criteria, triggers, and thresholds

MoBayes: A Modular Bayesian Framework for Separating Reasoning from Language in Conversational Clinical Decision Support

Is SAM3 ready for pathology segmentation?

HeadRank: Decoding-Free Passage Reranking via Preference-Aligned Attention Heads

A Lightweight Transformer for Pain Recognition from Brain Activity

Class Unlearning via Depth-Aware Removal of Forget-Specific Directions

C-ReD: A Comprehensive Chinese Benchmark for AI-Generated Text Detection Derived from Real-World Prompts

Efficient Emotion-Aware Iconic Gesture Prediction for Robot Co-Speech

DSPR: Dual-Stream Physics-Residual Networks for Trustworthy Industrial Time Series Forecasting

Learning Stable Predictors from Weak Supervision under Distribution Shift

Target-Aligned Reinforcement Learning

IMPACT: Influence Modeling for Open-Set Time Series Anomaly Detection

How do LLMs Compute Verbal Confidence

Synthetic Data Generation for Brain-Computer Interfaces: Overview, Benchmarking, and Future Directions

Does AI See like Art Historians? Interpreting How Vision Language Models Recognize Artistic Style

ViroGym: Realistic Large-Scale Benchmarks for Evaluating Viral Proteins

How to Model AI Agents as Personas?: Applying the Persona Ecosystem Playground to 41,300 Posts on Moltbook for Behavioral Insights

3D aperture-engineered diffractive neural networks for super-resolution electromagnetic wave computing

DLEBench: Evaluating Small-scale Object Editing Ability for Instruction-based Image Editing Model

A Geometric Analysis of Small-sized Language Model Hallucinations

Language Model Memory and Memory Models for Language

How does longer temporal context enhance multimodal narrative video processing in the brain?

Learning Rate Matters: Vanilla LoRA May Suffice for LLM Fine-tuning

Protein Autoregressive Modeling via Multiscale Structure Generation

Rethinking the Design Space of Reinforcement Learning for Diffusion Models: On the Importance of Likelihood Estimation Beyond Loss Design

Enabling Real-Time Colonoscopic Polyp Segmentation on Commodity CPUs via Ultra-Lightweight Architecture

WIND: Weather Inverse Diffusion for Zero-Shot Atmospheric Modeling

Whispers of Wealth: Red-Teaming Google's Agent Payments Protocol via Prompt Injection

SuperInfer: SLO-Aware Rotary Scheduling and Memory Management for LLM Inference on Superchips

FreeOrbit4D: Training-Free Arbitrary Camera Redirection for Monocular Videos via Foreground-Complete 4D Reconstruction

Disentangling generalization and memorization in large language models using chess

Multimodal system for skin cancer detection

Q-learning with Adjoint Matching

STAGE: A Full-Screenplay Benchmark for Reasoning over Evolving Storie

Tracing Moral Foundations in Large Language Models

Eliminating Inductive Bias in Reward Models with Information-Theoretic Guidance

Toward Training Superintelligent Software Agents through Self-Play SWE-RL

MaxShapley: Towards Incentive-compatible Generative Search with Fair Context Attribution

NORi: An ML-Augmented Ocean Boundary Layer Parameterization

Open-Set Domain Adaptation Under Background Distribution Shift: Challenges and A Provably Efficient Solution

HarmonicAttack: An Adaptive Cross-Domain Audio Watermark Removal

Data-driven Acceleration of MPC with Guarantees

Adaptive Residual-Update Steering for Low-Overhead Hallucination Mitigation in Large Vision Language Models

PlantTraitNet: An Uncertainty-Aware Multimodal Framework for Global-Scale Plant Trait Inference from Citizen Science Data

Noise-corrected GRPO: From Noisy Rewards to Unbiased Gradients

Needles in the Landscape: Semi-Supervised Pseudolabeling for Archaeological Site Discovery under Label Scarcity

Dr.LLM: Dynamic Layer Routing in LLMs

WARC-Bench: Web Archive Based Benchmark for GUI Subtask Executions

Proximal Diffusion Neural Sampler

Beyond Majority Voting: LLM Aggregation by Leveraging Higher-Order Information

HiDe: Rethinking The Zoom-IN method in High Resolution MLLMs via Hierarchical Decoupling

Jailbreaking on Text-to-Video Models via Scene Splitting Strategy

YAC: Bridging Natural Language and Interactive Visual Exploration with Generative AI for Biomedical Data Discovery

ORCA: An Agentic Reasoning Framework for Hallucination and Adversarial Robustness in Vision-Language Models

Digital Voices of Survival: From Social Media Disclosures to Support Provisions for Domestic Violence Victims

PiKV: KV Cache Management System for Mixture of Experts

CoLD: Counterfactually-Guided Length Debiasing for Process Reward Models in Mathematical Reasoning

Fine-tuning Large Language Model for Automated Algorithm Design

Mechanistic Interpretability Needs Philosophy

Recursive Entropic Risk Optimization in Discounted MDPs: Sample Complexity Bounds with a Generative Model

Lookahead Q-Cache: Achieving More Consistent KV Cache Eviction via Pseudo Query

Created by

Haebom

저자

Yixuan Wang, Shiyu Ji, Yijun Liu, Yuzhuang Xu, Yang Xu, Qingfu Zhu, Wanxiang Che

개요

본 논문은 대규모 언어 모델(LLM)의 디코딩 속도를 높이기 위해 사용되는 키-값 캐시(KV 캐시)의 메모리 사용량 문제를 해결하는 새로운 방법인 Lookahead Q-Cache (LAQ)를 제안합니다. 기존 방법들은 사전 채우기 단계의 어텐션 점수를 이용하여 토큰을 제거하지만, 특히 메모리 용량이 부족할 때 실제 추론 쿼리와의 불일치 문제가 발생합니다. LAQ는 저렴한 비용으로 미래의 추론 쿼리를 예측하는 의사 룩어헤드 쿼리를 생성하여, 실제 추론 시나리오와 더 일치하는 정확한 KV 캐시 제거를 수행합니다.