[공지사항]을 빙자한 안부와 근황

Show more

Daily Arxiv

世界中で発行される人工知能関連の論文をまとめるページです。
このページはGoogle Geminiを活用して要約し、非営利で運営しています。
論文の著作権は著者および関連機関にあり、共有する際は出典を明記してください。

Merge Kernel for Bayesian Optimization on Permutation Space

Demographic-aware fine-grained classification of pediatric wrist fractures

Generative Multi-Target Cross-Domain Recommendation

ParaStudent: Generating and Evaluating Realistic Student Code by Teaching LLMs to Struggle

Modeling Open-World Cognition as On-Demand Synthesis of Probabilistic Models

EgoVLA: Learning Vision-Language-Action Models from Egocentric Human Videos

Inversion-DPO: Precise and Efficient Post-Training for Diffusion Models

A Simple Baseline for Stable and Plastic Neural Networks

WildFX: A DAW-Powered Pipeline for In-the-Wild Audio FX Graph Modeling

From KMMLU-Redux to KMMLU-Pro: A Professional Korean Benchmark Suite for LLM Evaluation

Agent KB: Leveraging Cross-Domain Experience for Agentic Problem Solving

How Not to Detect Prompt Injections with an LLM

Critiques of World Models

The role of large language models in UI/UX design: A systematic literature review

LearnLens: LLM-Enabled Personalised, Curriculum-Grounded Feedback with Educators in the Loop

STACK: Adversarial Attacks on LLM Safeguard Pipelines

ZonUI-3B: A Lightweight Vision-Language Model for Cross-Resolution GUI Grounding

Understanding Reasoning in Thinking Language Models via Steering Vectors

Agentic Neural Networks: Self-Evolving Multi-Agent Systems via Textual Backpropagation

EvolveNav: Self-Improving Embodied Reasoning for LLM-Based Vision-Language Navigation

TextDiffuser-RL: Efficient and Robust Text Layout Optimization for High-Fidelity Text-to-Image Synthesis

SpecMaskFoley: Steering Pretrained Spectral Masked Generative Transformer Toward Synchronized Video-to-audio Synthesis via ControlNet

Exploring Graph Representations of Logical Forms for Language Modeling

DeepSeek-Prover-V2: Advancing Formal Mathematical Reasoning via Reinforcement Learning for Subgoal Decomposition

ParaPO: Aligning Language Models to Reduce Verbatim Reproduction of Pre-training Data

DP2Unlearning: An Efficient and Guaranteed Unlearning Framework for LLMs

CDUPatch: Color-Driven Universal Adversarial Patch Attack for Dual-Modal Visible-Infrared Detectors

Hands-On: Segmenting Individual Signs from Continuous Sequences

Can we ease the Injectivity Bottleneck on Lorentzian Manifolds for Graph Neural Networks?

Align Your Rhythm: Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation

HoH: A Dynamic Benchmark for Evaluating the Impact of Outdated Information on Retrieval-Augmented Generation

AIvaluateXR: An Evaluation Framework for on-Device AI in XR with Benchmarking Results

An Empirical Risk Minimization Approach for Offline Inverse RL and Dynamic Discrete Choice Model

Evaluating link prediction: New perspectives and recommendations

Learning to Reason at the Frontier of Learnability

Stonefish: Supporting Machine Learning Research in Marine Robotics

Harmony in Divergence: Towards Fast, Accurate, and Memory-efficient Zeroth-order LLM Fine-tuning

On the Transfer of Knowledge in Quantum Algorithms

Code Readability in the Age of Large Language Models: An Industrial Case Study from Atlassian

Bias in Decision-Making for AI's Ethical Dilemmas: A Comparative Study of ChatGPT and Claude

ASTRID - An Automated and Scalable TRIaD for the Evaluation of RAG-based Clinical Question Answering Systems

Consistency of Responses and Continuations Generated by Large Language Models on Social Media

From Code to Compliance: Assessing ChatGPT's Utility in Designing an Accessible Webpage -- A Case Study

Temporal reasoning for timeline summarisation in social media

Invisible Textual Backdoor Attacks based on Dual-Trigger

Towards scientific discovery with dictionary learning: Extracting biological concepts from microscopy foundation models

Two-Stage Pretraining for Molecular Property Prediction in the Wild

Towards Practical Operation of Deep Reinforcement Learning Agents in Real-World Network Management at Open RAN Edges

An Approach for Auto Generation of Labeling Functions for Software Engineering Chatbots

Bridging Local and Global Knowledge via Transformer in Board Games

Entropy Loss: An Interpretability Amplifier of 3D Object Detection Network for Intelligent Driving

FBSDiff: Plug-and-Play Frequency Band Substitution of Diffusion Features for Highly Controllable Text-Driven Image Translation

On Pre-training of Multimodal Language Models Customized for Chart Understanding

Visual Grounding Methods for Efficient Interaction with Desktop Graphical User Interfaces

Instruct-MusicGen: Unlocking Text-to-Music Editing for Music Language Models via Instruction Tuning

Meta4XNLI: A Crosslingual Parallel Corpus for Metaphor Detection and Interpretation

SecurePose: Automated Face Blurring and Human Movement Kinematics Extraction from Videos Recorded in Clinical Settings

Improved DDIM Sampling with Moment Matching Gaussian Mixtures

Eye-tracked Virtual Reality: A Comprehensive Survey on Methods and Privacy Challenges

From Roots to Rewards: Dynamic Tree Reasoning with RL

Illuminating the Three Dogmas of Reinforcement Learning under Evolutionary Light

Instance space analysis of the capacitated vehicle routing problem

Multi-Agent LLMs as Ethics Advocates for AI-Based Systems

GATSim: Urban Mobility Simulation with Generative Agents

Reasoning about Uncertainty: Do Reasoning Models Know When They Don't Know?

The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity

Strategic Reflectivism In Intelligent Systems

SafeAgent: Safeguarding LLM Agents via an Automated Risk Simulator

What the F*ck Is Artificial General Intelligence?

Towards Reasoning Era: A Survey of Long Chain-of-Thought for Reasoning Large Language Models

From Words to Collisions: LLM-Guided Evaluation and Adversarial Generation of Safety-Critical Driving Scenarios

To Code or not to Code? Adaptive Tool Integration for Math Language Models via Expectation-Maximization

BLAST: A Stealthy Backdoor Leverage Attack against Cooperative Multi-Agent Deep Reinforcement Learning based Systems

UniEmoX: Cross-modal Semantic-Guided Large-Scale Pretraining for Universal Scene Emotion Perception

CorMulT: A Semi-supervised Modality Correlation-aware Multimodal Transformer for Sentiment Analysis

Toward Temporal Causal Representation Learning with Tensor Decomposition

Kolmogorov Arnold Networks (KANs) for Imbalanced Data - An Empirical Perspective

NoHumansRequired: Autonomous High-Quality Image Editing Triplet Mining

Lessons from the TREC Plain Language Adaptation of Biomedical Abstracts (PLABA) トラック

Multi-Centre Validation of a Deep Learning Model for Scoliosis Assessment

The Emotion-Memory Link: Do Memorability Annotations Matter for Intelligent Systems?

DENSE: Longitudinal Progress Note Generation with Temporal Modeling of Heterogeneous Clinical Notes Across Hospital Visits

Edge Intelligence with Spiking Neural Networks

VLA-Mark: A cross modal watermark for large vision-language alignment model

Noradrenergic-inspired gain modulation attenuates the stability gap in joint training

A multi-strategy improved snake optimizer for 3-dimensional UAV path planning and engineering problems

Photonic Fabric Platform for AI Accelerators

OrthoInsight: Rib Fracture Diagnosis and Report Generation Based on Multi-Modal Large Models

CSD-VAR: Content-Style Decomposition in Visual Autoregressive Models

A segmented robot grasping perception neural network for edge AI

Bottom-up Domain-specific Superintelligence: A Reliable Knowledge Graph is What We Need

DUALRec: A Hybrid Sequential and Language Model Framework for Context-Aware Movie Recommendation

Exploiting Primacy Effect To Improve Large Language Models

Generalist Forecasting with Frozen Video Models via Latent Diffusion

Convergent transformations of visual representation in brains and models

Preprint: Did I Just Browse A Website Written by LLMs?

The Levers of Political Persuasion with Conversational AI

Political Leaning and Politicalness Classification of Texts

Self-supervised learning on gene expression data

Using LLMs to identify features of personal and professional skills in an open-response situational judgment test

Enhancing Spatial Reasoning in Vision-Language Models via Chain-of-Thought Prompting and Reinforcement Learning

Created by

Haebom

作者

Binbin Ji, Siddharth Agrawal, Qiance Tang, Yvonne Wu

概要

この研究では、Chain-of-Thought（CoT）プロンプトと強化学習を通じて、視覚言語モデル（VLM）の空間推論能力を調査します。単純なCoTフォーマットはパフォーマンスを向上させるのに役立たないか、むしろパフォーマンスを低下させますが、視覚グラフベースの多段階構造化プロンプト（SceneGraph CoT）は空間推論精度を大幅に向上させることがわかりました。 Group Relative Policy Optimization（GRPO）を使用してSATデータセットでモデルを微調整し、CVBenchでパフォーマンスを評価しました。マップ学習微調整（SFT）と比較して、GRPOはPass @ 1評価でより高い精度を達成し、分布外（OOD）条件で優れた堅牢性を示します。特に、SFTは表面的な言語パターンに過剰に適合し、テスト時に構文が変更されると（たとえば、「closer to」から「farther from」に）パフォーマンスが低下する可能性がありますが、GRPOはより安定して一般化し、そのような変化にも安定したパフォーマンスを維持します。この研究は、強化学習と構造化されたプロンプトが最新のVLMの空間推論能力と一般化性能を向上させる方法についての洞察を提供します。すべてのコードはhttps://github.com/Yvonne511/spatial-vlm-investigatorで公開されています。

GitHub - Yvonne511/spatial-vlm-investigator

Contribute to Yvonne511/spatial-vlm-investigator development by creating an account on GitHub.

Takeaways、Limitations

•

Takeaways：

◦

SceneGraph CoTプロンプトがVLMの空間推論性能を向上させることを明らかにした。

◦

GRPOベースの強化学習により、SFTよりも高い精度とOOD堅牢性を達成。

◦

SFTの過適合問題とGRPOの優れた一般化能力を提示

◦

VLMの空間推論能力向上のための新しい方法論の提示

•

Limitations：

◦

研究に使用されたデータセットとモデルの一般化の可能性に関するさらなる研究が必要です。

◦

異なるタイプの空間推論問題に対するGRPOの性能評価の必要性

◦

GRPOの計算コストと効率に関する追加分析の必要性

Made with Slashpage