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.

Segment Anything for Satellite Imagery: A Strong Baseline and a Regional Dataset for Automatic Field Delineation

VRAIL: Vectorized Reward-based Attribution for Interpretable Learning

LLM Web Dynamics: Tracing Model Collapse in a Network of LLMs

Accurate and scalable exchange-correlation with deep learning

Working Document -- Formalising Software Requirements with Large Language Models

AlphaDecay: Module-wise Weight Decay for Heavy-Tailed Balancing in LLMs

Bures-Wasserstein Flow Matching for Graph Generation

Distributional Training Data Attribution

AFBS:Buffer Gradient Selection in Semi-asynchronous Federated Learning

Robust LLM Unlearning with MUDMAN: Meta-Unlearning with Disruption Masking And Normalization

Exploring the Secondary Risks of Large Language Models

Supernova Event Dataset: Interpreting Large Language Models' Personality through Critical Event Analysis

Agent-RLVR: Training Software Engineering Agents via Guidance and Environment Rewards

C-SEO Bench: Does Conversational SEO Work?

xInv: Explainable Optimization of Inverse Problems

Effective Red-Teaming of Policy-Adherent Agents

SLED: A Speculative LLM Decoding Framework for Efficient Edge Serving

PlantDeBERTa: An Open Source Language Model for Plant Science

Reinforcement Learning Teachers of Test Time Scaling

DriveSuprim: Towards Precise Trajectory Selection for End-to-End Planning

OWMM-Agent: Open World Mobile Manipulation With Multi-modal Agentic Data Synthesis

Optimizing Sensory Neurons: Nonlinear Attention Mechanisms for Accelerated Convergence in Permutation-Invariant Neural Networks for Reinforcement Learning

Dual Debiasing for Noisy In-Context Learning for Text Generation

Eye of Judgement: Dissecting the Evaluation of Russian-speaking LLMs with POLLUX

Maximizing Confidence Alone Improves Reasoning

Analysis and Evaluation of Synthetic Data Generation in Speech Dysfluency Detection

FRAMES-VQA: Benchmarking Fine-Tuning Robustness across Multi-Modal Shifts in Visual Question Answering

Position is Power: System Prompts as a Mechanism of Bias in Large Language Models (LLMs)

Cross from Left to Right Brain: Adaptive Text Dreamer for Vision-and-Language Navigation

Pretraining Language Models to Ponder in Continuous Space

SIPDO: Closed-Loop Prompt Optimization via Synthetic Data Feedback

Learning to Insert for Constructive Neural Vehicle Routing Solver

LightRetriever: A LLM-based Hybrid Retrieval Architecture with 1000x Faster Query Inference

MIRAGE: A Multi-modal Benchmark for Spatial Perception, Reasoning, and Intelligence

Introducing voice timbre attribute detection

The Voice Timbre Attribute Detection 2025 Challenge Evaluation Plan

TrumorGPT: Graph-Based Retrieval-Augmented Large Language Model for Fact-Checking

Democracy of AI Numerical Weather Models: An Example of Global Forecasting with FourCastNetv2 Made by a University Research Lab Using GPU

Learning to Reason under Off-Policy Guidance

Protecting Your Voice: Temporal-aware Robust Watermarking

Personalized News Recommendation with Multi-granularity Candidate-aware User Modeling

Learning from Reference Answers: Versatile Language Model Alignment without Binary Human Preference Data

AutoPDL: Automatic Prompt Optimization for LLM Agents

PiCo: Jailbreaking Multimodal Large Language Models via $\textbf{Pi}$ctorial $\textbf{Co}$de Contextualization

From Easy to Hard: Building a Shortcut for Differentially Private Image Synthesis

Context-Aware Human Behavior Prediction Using Multimodal Large Language Models: Challenges and Insights

Rubric Is All You Need: Enhancing LLM-based Code Evaluation With Question-Specific Rubrics

Shapley Revisited: Tractable Responsibility Measures for Query Answers

Simple and Critical Iterative Denoising: A Recasting of Discrete Diffusion in Graph Generation

Large Language Models powered Malicious Traffic Detection: Architecture, Opportunities and Case Study

TreeSynth: Synthesizing Diverse Data from Scratch via Tree-Guided Subspace Partitioning

A Dual-Directional Context-Aware Test-Time Learning for Text Classification

LED: LLM Enhanced Open-Vocabulary Object Detection without Human Curated Data Generation

HiRAG: Retrieval-Augmented Generation with Hierarchical Knowledge

Trajectory Prediction for Autonomous Driving: Progress, Limitations, and Future Directions

POPGym Arcade: Parallel Pixelated POMDPs

BAnG: Bidirectional Anchored Generation for Conditional RNA Design

Directional Gradient Projection for Robust Fine-Tuning of Foundation Models

ParamMute: Suppressing Knowledge-Critical FFNs for Faithful Retrieval-Augmented Generation

Steering LLMs for Formal Theorem Proving

Exploring the Potential of Encoder-free Architectures in 3D LMMs

Compromising Honesty and Harmlessness in Language Models via Deception Attacks

Large Language Model Guided Self-Debugging Code Generation

ASCenD-BDS: Adaptable, Stochastic and Context-aware framework for Detection of Bias, Discrimination and Stereotyping

LoRA-One: One-Step Full Gradient Could Suffice for Fine-Tuning Large Language Models, Provably and Efficiently

Segmentation-Aware Generative Reinforcement Network (GRN) for Tissue Layer Segmentation in 3-D Ultrasound Images for Chronic Low-back Pain (cLBP) Assessment

AnyEnhance: A Unified Generative Model with Prompt-Guidance and Self-Critic for Voice Enhancement

SEAL: Scaling to Emphasize Attention for Long-Context Retrieval

CAD-GPT: Synthesising CAD Construction Sequence with Spatial Reasoning-Enhanced Multimodal LLMs

The Impact of Input Order Bias on Large Language Models for Software Fault Localization

GeAR: Graph-enhanced Agent for Retrieval-augmented Generation

Rethinking Cancer Gene Identification through Graph Anomaly Analysis

SurgSora: Object-Aware Diffusion Model for Controllable Surgical Video Generation

DSGram: Dynamic Weighting Sub-Metrics for Grammatical Error Correction in the Era of Large Language Models

Human Action CLIPs: Detecting AI-generated Human Motion

FLARE: Toward Universal Dataset Purification against Backdoor Attacks

G3Flow: Generative 3D Semantic Flow for Pose-aware and Generalizable Object Manipulation

Cross-Camera Distracted Driver Classification through Feature Disentanglement and Contrastive Learning

Song Form-aware Full-Song Text-to-Lyrics Generation with Multi-Level Granularity Syllable Count Control

Generating Energy-efficient code with LLMs

DeepMedcast: A Deep Learning Method for Generating Intermediate Weather Forecasts among Multiple NWP Models

How Far is Video Generation from World Model: A Physical Law Perspective

One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Models

The Hive Mind is a Single Reinforcement Learning Agent

How Numerical Precision Affects Arithmetical Reasoning Capabilities of LLMs

FutureFill: Fast Generation from Convolutional Sequence Models

Leveraging Model Guidance to Extract Training Data from Personalized Diffusion Models

Machine-learning based high-bandwidth magnetic sensing

MOST: MR reconstruction Optimization for multiple downStream Tasks via continual learning

Bridging Geometric Diffusion and Energy Minimization: A Unified Framework for Neural Message Passing

Large Language Models for Disease Diagnosis: A Scoping Review

RePST: Language Model Empowered Spatio-Temporal Forecasting via Semantic-Oriented Reprogramming

Smooth InfoMax -- Towards Easier Post-Hoc Interpretability

PREMAP: A Unifying PREiMage APproximation Framework for Neural Networks

Reasoning Circuits in Language Models: A Mechanistic Interpretation of Syllogistic Inference

UniMoT: Unified Molecule-Text Language Model with Discrete Token Representation

Handling Numeric Expressions in Automatic Speech Recognition

"I understand why I got this grade": Automatic Short Answer Grading with Feedback

DART: An Automated End-to-End Object Detection Pipeline with Data Diversification, Open-Vocabulary Bounding Box Annotation, Pseudo-Label Review, and Model Training

Rich Interoperable Metadata for Cultural Heritage Projects at Jagiellonian University

Step-Wise Formal Verification for LLM-Based Mathematical Problem Solving

Created by

Haebom

저자

Kuo Zhou, Lu Zhang

개요

본 논문은 대규모 언어 모델(LLM)이 수학 문제 해결 과정에서 논리적 추론 및 계산 오류를 범할 수 있다는 점을 지적하며, LLM이 생성한 솔루션의 정확성을 공식적으로 검증하는 프레임워크인 MATH-VF를 제안합니다. MATH-VF는 자연어 솔루션을 공식적 맥락으로 변환하는 Formalizer와 컴퓨터 대수 시스템 및 SMT 솔버와 같은 외부 도구를 통합하여 각 명제의 정확성을 평가하고 오류 발생 시 수정 피드백을 제공하는 Critic으로 구성됩니다. MATH-VF의 효과를 검증 및 수정 두 가지 시나리오에서 실험적으로 연구하며, MATH500 및 ProcessBench와 같은 널리 사용되는 수학 벤치마크를 통해 기존 방식보다 우수함을 입증합니다.

시사점, 한계점

•

시사점:

◦

LLM 기반 수학 문제 해결의 신뢰성 향상에 기여하는 새로운 검증 프레임워크 제시

◦

Formalizer와 Critic의 결합을 통한 효과적인 오류 검출 및 수정

◦

MATH500 및 ProcessBench 벤치마크를 통한 실험적 성능 검증 및 기존 방식 대비 우수성 입증

•

한계점:

◦

Formalizer의 LLM 의존성으로 인한 LLM의 한계가 MATH-VF의 성능에 영향을 미칠 수 있음.

◦

사용된 외부 도구(CAS, SMT solver)의 성능 및 한계가 MATH-VF의 성능을 제한할 수 있음.

◦

특정 유형의 수학 문제에 대한 성능이 다른 유형에 비해 상대적으로 낮을 수 있음. (벤치마크 데이터셋의 편향 가능성)

◦

매우 복잡한 문제에 대한 검증의 효율성 및 정확성에 대한 추가적인 연구가 필요할 수 있음.

Made with Slashpage