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.

Surgical Agent Orchestration Platform for Voice-directed Patient Data Interaction

FinRpt: Dataset, Evaluation System and LLM-based Multi-agent Framework for Equity Research Report Generation

Hard vs. Noise: Resolving Hard-Noisy Sample Confusion in Recommender Systems via Large Language Models

Learning Quantized Continuous Controllers for Integer Hardware

Differentiated Directional Intervention A Framework for Evading LLM Safety Alignment

TiS-TSL: Image-Label Supervised Surgical Video Stereo Matching via Time-Switchable Teacher-Student Learning

Active Learning for Animal Re-Identification with Ambiguity-Aware Sampling

Transolver is a Linear Transformer: Revisiting Physics-Attention through the Lens of Linear Attention

Exploiting Inter-Session Information with Frequency-enhanced Dual-Path Networks for Sequential Recommendation

SWE-fficiency: Can Language Models Optimize Real-World Repositories on Real Workloads?

A Remarkably Efficient Paradigm to Multimodal Large Language Models for Sequential Recommendation

EndoIR: Degradation-Agnostic All-in-One Endoscopic Image Restoration via Noise-Aware Routing Diffusion

Enhancing Diffusion Model Guidance through Calibration and Regularization

DRAGON: Guard LLM Unlearning in Context via Negative Detection and Reasoning

Report from Workshop on Dialogue alongside Artificial Intelligence

Token Is All You Need: Cognitive Planning through Belief-Intent Co-Evolution

Selective Diabetic Retinopathy Screening with Accuracy-Weighted Deep Ensembles and Entropy-Guided Abstention

SWE-Compass: Towards Unified Evaluation of Agentic Coding Abilities for Large Language Models

Alternative Fairness and Accuracy Optimization in Criminal Justice

I Detect What I Don't Know: Incremental Anomaly Learning with Stochastic Weight Averaging-Gaussian for Oracle-Free Medical Imaging

OMPILOT: Harnessing Transformer Models for Auto Parallelization to Shared Memory Computing Paradigms

Federated Learning with Gramian Angular Fields for Privacy-Preserving ECG Classification on Heterogeneous IoT Devices

How to Evaluate Speech Translation with Source-Aware Neural MT Metrics

GUI-AIMA: Aligning Intrinsic Multimodal Attention with a Context Anchor for GUI Grounding

Normalization in Attention Dynamics

Preference-based Reinforcement Learning beyond Pairwise Comparisons: Benefits of Multiple Options

Language over Content: Tracing Cultural Understanding in Multilingual Large Language Models

TACL: Threshold-Adaptive Curriculum Learning Strategy for Enhancing Medical Text Understanding

TraceCoder: Towards Traceable ICD Coding via Multi-Source Knowledge Integration

Bee: A High-Quality Corpus and Full-Stack Suite to Unlock Advanced Fully Open MLLMs

Comparative Analysis of Large Language Models for the Machine-Assisted Resolution of User Intentions

Evolutionary Profiles for Protein Fitness Prediction

Epistemic Diversity and Knowledge Collapse in Large Language Models

MENLO: From Preferences to Proficiency -- Evaluating and Modeling Native-like Quality Across 47 Languages

Towards Foundation Models for Zero-Shot Time Series Anomaly Detection: Leveraging Synthetic Data and Relative Context Discrepancy

CyberSOCEval: Benchmarking LLMs Capabilities for Malware Analysis and Threat Intelligence Reasoning

A Realistic Evaluation of Cross-Frequency Transfer Learning and Foundation Forecasting Models

TimeMosaic: Temporal Heterogeneity Guided Time Series Forecasting via Adaptive Granularity Patch and Segment-wise Decoding

RSVG-ZeroOV: Exploring a Training-Free Framework for Zero-Shot Open-Vocabulary Visual Grounding in Remote Sensing Images

Instance Generation for Meta-Black-Box Optimization through Latent Space Reverse Engineering

Decoding Latent Attack Surfaces in LLMs: Prompt Injection via HTML in Web Summarization

Towards Methane Detection Onboard Satellites

OPERA: A Reinforcement Learning--Enhanced Orchestrated Planner-Executor Architecture for Reasoning-Oriented Multi-Hop Retrieval

Can LLM-Generated Textual Explanations Enhance Model Classification Performance? An Empirical Study

Towards Embodied Agentic AI: Review and Classification of LLM- and VLM-Driven Robot Autonomy and Interaction

CoCoLIT: ControlNet-Conditioned Latent Image Translation for MRI to Amyloid PET Synthesis

Beyond Algorethics: Addressing the Ethical and Anthropological Challenges of AI Recommender Systems

Imbalance in Balance: Online Concept Balancing in Generation Models

ReCode: Updating Code API Knowledge with Reinforcement Learning

Rethinking Losses for Diffusion Bridge Samplers

Zeroth-Order Optimization Finds Flat Minima

UniSite: The First Cross-Structure Dataset and Learning Framework for End-to-End Ligand Binding Site Detection

A Unified and Fast-Sampling Diffusion Bridge Framework via Stochastic Optimal Control

BroadGen: A Framework for Generating Effective and Efficient Advertiser Broad Match Keyphrase Recommendations

FB-RAG: Improving RAG with Forward and Backward Lookup

FedSEA-LLaMA: A Secure, Efficient and Adaptive Federated Splitting Framework for Large Language Models

Policy-Driven World Model Adaptation for Robust Offline Model-based Reinforcement Learning

RL in Name Only? Analyzing the Structural Assumptions in RL post-training for LLMs

Internal Causal Mechanisms Robustly Predict Language Model Out-of-Distribution Behaviors

Tool-Aided Evolutionary LLM for Generative Policy Toward Efficient Resource Management in Wireless Federated Learning

FALCON: False-Negative Aware Learning of Contrastive Negatives in Vision-Language Alignment

How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference

FaSDiff: Balancing Perception and Semantics in Face Compression via Stable Diffusion Priors

On the generalization of language models from in-context learning and finetuning: a controlled study

Enhancing Speech-to-Speech Dialogue Modeling with End-to-End Retrieval-Augmented Generation

A Multimodal Recaptioning Framework to Account for Perceptual Diversity Across Languages in Vision-Language Modeling

WildFireCan-MMD: A Multimodal Dataset for Classification of User-Generated Content During Wildfires in Canada

MULTI-LF: A Continuous Learning Framework for Real-Time Malicious Traffic Detection in Multi-Environment Networks

STAR-1: Safer Alignment of Reasoning LLMs with 1K Data

COPA: Comparing the incomparable in multi-objective model evaluation

CLEV: LLM-Based Evaluation Through Lightweight Efficient Voting for Free-Form Question-Answering

Towards Synthesizing High-Dimensional Tabular Data with Limited Samples

Explaining the Unexplainable: A Systematic Review of Explainable AI in Finance

Learning Vision-Based Neural Network Controllers with Semi-Probabilistic Safety Guarantees

MA-GTS: A Multi-Agent Framework for Solving Complex Graph Problems in Real-World Applications

On the Convergence and Stability of Upside-Down Reinforcement Learning, Goal-Conditioned Supervised Learning, and Online Decision Transformers

A Multi-Agent Conversational Bandit Approach to Online Evaluation and Selection of User-Aligned LLM Responses

Generalizing Weisfeiler-Lehman Kernels to Subgraphs

SCoTT: Strategic Chain-of-Thought Tasking for Wireless-Aware Robot Navigation in Digital Twins

GeMID: Generalizable Models for IoT Device Identification

Flash Inference: Near Linear Time Inference for Long Convolution Sequence Models and Beyond

Bayes Adaptive Monte Carlo Tree Search for Offline Model-based Reinforcement Learning

Selection of LLM Fine-Tuning Data based on Orthogonal Rules

Benchmarking Domain Generalization Algorithms in Computational Pathology

Reference-Guided Verdict: LLMs-as-Judges in Automatic Evaluation of Free-Form QA

Identifying treatment response subgroups in observational time-to-event data

Informed Correctors for Discrete Diffusion Models

Integrating Artificial Intelligence into Operating Systems: A Survey on Techniques, Applications, and Future Directions

Multimodal Adversarial Defense for Vision-Language Models by Leveraging One-To-Many Relationships

Spikingformer: A Key Foundation Model for Spiking Neural Networks

DeepPersona: A Generative Engine for Scaling Deep Synthetic Personas

Two Heads are Better than One: Distilling Large Language Model Features Into Small Models with Feature Decomposition and Mixture

A Theoretical Analysis of Detecting Large Model-Generated Time Series

Green AI: A systematic review and meta-analysis of its definitions, lifecycle models, hardware and measurement attempts

Spilling the Beans: Teaching LLMs to Self-Report Their Hidden Objectives

Reasoning with Confidence: Efficient Verification of LLM Reasoning Steps via Uncertainty Heads

When Object-Centric World Models Meet Policy Learning: From Pixels to Policies, and Where It Breaks

ScRPO: From Errors to Insights

Deep Value Benchmark: Measuring Whether Models Generalize Deep Values or Shallow Preferences

Glia: A Human-Inspired AI for Automated Systems Design and Optimization

Better Estimation of the KL Divergence Between Language Models

Created by

Haebom

저자

Afra Amini, Tim Vieira, Ryan Cotterell

개요

본 논문은 언어 모델 간의 Kullback-Leibler (KL) divergence를 추정하는 문제를 다룹니다. KL divergence의 정확한 계산은 어렵기 때문에, 실제로는 샘플링 기반 추정량을 사용합니다. 기존의 Monte Carlo (MC) 추정량은 높은 분산을 가지고 음수 값을 가질 수 있다는 문제점이 있습니다. 본 논문에서는 Rao-Blackwellized 추정량을 제시하여, MC 추정량보다 분산이 작거나 같은 불편 추정량을 제공합니다. 감정 제어 미세 조정 실험을 통해 제안된 추정량이 더 안정적인 KL 추정치를 제공하고 분산을 크게 줄이는 것을 보였습니다. 또한, KL divergence 기울기의 Rao-Blackwellized 추정량을 유도하여, 더 안정적인 훈련과 보상 대 KL의 Pareto frontier에 더 자주 나타나는 모델을 생성합니다.

시사점, 한계점

•

시사점:

◦

Rao-Blackwellized 추정량을 이용하여 언어 모델 간 KL divergence를 더욱 안정적이고 효율적으로 추정할 수 있는 방법을 제시했습니다.

◦

기존 MC 추정량의 높은 분산 및 음수 값 문제를 해결하여 KL divergence 추정의 신뢰성을 높였습니다.

◦

KL divergence 기울기의 Rao-Blackwellized 추정량을 통해 더 안정적인 모델 학습 및 향상된 성능을 달성했습니다.

◦

RLHF, 해석성, 지식 증류 등 KL divergence 추정이 필요한 다양한 응용 분야에 기여할 수 있습니다.

•

한계점:

◦

제안된 방법의 효과는 특정 실험 설정(감정 제어 미세 조정)에 국한되어 있으며, 다른 설정에서의 일반화 가능성은 추가 연구가 필요합니다.

◦

Rao-Blackwellized 추정량의 계산 비용이 MC 추정량보다 높을 수 있습니다. 비용과 효율성 간의 절충점을 고려해야 합니다.

◦

본 논문에서 다룬 것은 KL divergence 추정에 대한 것이며, KL divergence 자체의 해석 및 적용에 대한 심도있는 논의는 부족합니다.

Made with Slashpage