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。

Self-Questioning Language Models

Beyond risk: A proto-framework for assessing the societal impact of AI systems

Supervised Dynamic Dimension Reduction with Deep Neural Network

EmoSteer-TTS: Fine-Grained and Training-Free Emotion-Controllable Text-to-Speech via Activation Steering

LLMs Have a Heart of Stone: Demystifying the Soft Thinking Ability of Large Reasoning Models

Industrial LLM-based Code Optimization under Regulation: A Mixture-of-Agents Approach

Reliable Evaluation Protocol for Low-Precision Retrieval

Landsat30-AU: A Vision-Language Dataset for Australian Landsat Imagery

Tool-integrated Reinforcement Learning for Repo Deep Search

CauKer: classification time series foundation models can be pretrained on synthetic data only

Context-Adaptive Multi-Prompt Embedding with Large Language Models for Vision-Language Alignment

DMSC: Dynamic Multi-Scale Coordination Framework for Time Series Forecasting

HyCodePolicy: Hybrid Language Controllers for Multimodal Monitoring and Decision in Embodied Agents

Entity Representation Learning Through Onsite-Offsite Graph for Pinterest Ads

Evaluating User Experience in Conversational Recommender Systems: A Systematic Review Across Classical and LLM-Powered Approaches

Spatial-Frequency Aware for Object Detection in RAW Image

Learning Pivoting Manipulation with Force and Vision Feedback Using Optimization-based Demonstrations

NCCR: to Evaluate the Robustness of Neural Networks and Adversarial Examples

ChartM$^3$: Benchmarking Chart Editing with Multimodal Instructions

From Entanglement to Alignment: Representation Space Decomposition for Unsupervised Time Series Domain Adaptation

EcoTransformer: Attention without Multiplication

Bob's Confetti: Phonetic Memorization Attacks in Music and Video Generation

SDBench: A Comprehensive Benchmark Suite for Speaker Diarization

True Multimodal In-Context Learning Needs Attention to the Visual Context

Gauge Flow Models

Zero-Shot Neural Architecture Search with Weighted Response Correlation

The Dark Side of LLMs: Agent-based Attacks for Complete Computer Takeover

CAVGAN: Unifying Jailbreak and Defense of LLMs via Generative Adversarial Attacks on their Internal Representations

VOTE: Vision-Language-Action Optimization with Trajectory Ensemble Voting

A Comparative Study of Specialized LLMs as Dense Retrievers

Sign Spotting Disambiguation using Large Language Models

UnMix-NeRF: Spectral Unmixing Meets Neural Radiance Fields

Thought Anchors: Which LLM Reasoning Steps Matter?

UITron-Speech: Towards Automated GUI Agents Based on Speech Instructions

15,500 Seconds: Lean UAV Classification Using EfficientNet and Lightweight Fine-Tuning

AtmosMJ: Revisiting Gating Mechanism for AI Weather Forecasting Beyond the Year Scale

On the Fundamental Impossibility of Hallucination Control in Large Language Models

Multi-Modal Multi-Task Federated Foundation Models for Next-Generation Extended Reality Systems: Towards Privacy-Preserving Distributed Intelligence in AR/VR/MR

Text-Only Reasoning Unleashes Zero-Shot Multimodal Evaluators

CAIN: Hijacking LLM-Humans Conversations via Malicious System Prompts

Explain Less, Understand More: Jargon Detection via Personalized Parameter-Efficient Fine-tuning

What Lives? A meta-analysis of diverse opinions on the definition of life

A Generative Neural Annealer for Black-Box Combinatorial Optimization

GRILL: Gradient Signal Restoration in Ill-Conditioned Layers to Enhance Adversarial Attacks on Autoencoders

CostFilter-AD: Enhancing Anomaly Detection through Matching Cost Filtering

Mj\"olnir: A Deep Learning Parametrization Framework for Global Lightning Flash Density

RGB-Event based Pedestrian Attribute Recognition: A Benchmark Dataset and An Asymmetric RWKV Fusion Framework

ProtoECGNet: Case-Based Interpretable Deep Learning for Multi-Label ECG Classification with Contrastive Learning

Beyond Wide-Angle Images: Structure-to-Detail Video Portrait Correction via Unsupervised Spatiotemporal Adaptation

CITRAS: Covariate-Informed Transformer for Time Series Forecasting

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

Empirical Analysis of Sim-and-Real Cotraining of Diffusion Policies for Planar Pushing from Pixels

SimpleRL-Zoo: Investigating and Taming Zero Reinforcement Learning for Open Base Models in the Wild

NuPlanQA: A Large-Scale Dataset and Benchmark for Multi-View Driving Scene Understanding in Multi-Modal Large Language Models

The Impact of Item-Writing Flaws on Difficulty and Discrimination in Item Response Theory

Through the Magnifying Glass: Adaptive Perception Magnification for Hallucination-Free VLM Decoding

Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning

Pull-Based Query Scheduling for Goal-Oriented Semantic Communication

Accelerating Focal Search in Multi-Agent Path Finding with Tighter Lower Bounds

RAILGUN: A Unified Convolutional Policy for Multi-Agent Path Finding Across Different Environments and Tasks

UltraSTF: Ultra-Compact Model for Large-Scale Spatio-Temporal Forecasting

PTQ1.61: Push the Real Limit of Extremely Low-Bit Post-Training Quantization Methods for Large Language Models

Foundation Model of Electronic Medical Records for Adaptive Risk Estimation

Tool Unlearning for Tool-Augmented LLMs

Vision without Images: End-to-End Computer Vision from Single Compressive Measurements

How Do Generative Models Draw a Software Engineer? A Case Study on Stable Diffusion Bias

3DTTNet: Multimodal Fusion-Based 3D Traversable Terrain Modeling for Off-Road Environments

DOGR: Towards Versatile Visual Document Grounding and Referring

Real-World Offline Reinforcement Learning from Vision Language Model フィードバック

Causality-Driven Audits of Model Robustness

AUTALIC: A Dataset for Anti-AUTistic Ableist Language In Context

Beyond Adapter Retrieval: Latent Geometry-Preserving Composition via Sparse Task Projection

Pyhgf: A neural network library for predictive coding

Human Bias in the Face of AI: Examining Human Judgment Against Text Labeled as AI Generated

AVG-LLaVA: An Efficient Large Multimodal Model with Adaptive Visual Granularity

Parse Trees Guided LLM Prompt Compression

One Model, Any Conjunctive Query: Graph Neural Networks for Answering Queries over Incomplete Knowledge Graphs

A Value Based Parallel Update MCTS Method for Multi-Agent Cooperative Decision Making of Connected and Automated Vehicles

Fairness Definitions in Language Models Explained

CityLight: A Neighborhood-inclusive Universal Model for Coordinated City-scale Traffic Signal Control

Time Evidence Fusion Network: Multi-source View in Long-Term Time Series Forecasting

Long-Term Visual Object Tracking with Event Cameras: An Associative Memory Augmented Tracker and A Benchmark Dataset

Hulk: A Universal Knowledge Translator for Human-Centric Tasks

From Cluster Assumption to Graph Convolution: Graph-based Semi-Supervised Learning Revisited

Environmental Sound Classification on An Embedded Hardware Platform

Data Dependency Inference for Industrial Code Generation Based on UML Sequence Diagrams

InqEduAgent: Adaptive AI Learning Partners with Gaussian Process Augmentation

SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with LLM-Based Agents

RL-PLUS: Countering Capability Boundary Collapse of LLMs in Reinforcement Learning with Hybrid-policy Optimization

Higher Gauge Flow Models

Think How to Think: Mitigating Overthinking with Autonomous Difficulty Cognition in Large Reasoning Models

IS-Bench: Evaluating Interactive Safety of VLM-Driven Embodied Agents in Daily Household Tasks

SLR: Automated Synthesis for Scalable Logical Reasoning

The SWE-Bench Illusion: When State-of-the-Art LLMs Remember Instead of Reason

APOLLO: Automated LLM and Lean Collaboration for Advanced Formal Reasoning

Adversarial Cooperative Rationalization: The Risk of Spurious Correlations in Even Clean Datasets

Learning to Inference Adaptively for Multimodal Large Language Models

Efficient rule induction by ignoring pointless rules

Why the Agent Made that Decision: Contrastive Explanation Learning for Reinforcement Learning

Evaluating Detection Thresholds: The Impact of False Positives and Negatives on Super-Resolution Ultrasound Localization Microscopy

SimpleRL-Zoo: Investigating and Taming Zero Reinforcement Learning for Open Base Models in the Wild

Created by

Haebom

作者

Weihao Zeng, Yuzhen Huang, Qian Liu, Wei Liu, Keqing He, Zejun Ma, Junxian He

概要

この論文は、DeepSeek-R1のゼロRL学習方法をさまざまな基盤モデルに適用した研究で、ルールベースの補償を使用した単純強化学習（RL）フレームワークを通じて長距離思考連鎖（CoT）推論が自然に現れることを示しています。従来の研究が主にQwen2.5モデルに焦点を当てたのとは異なり、LLaMa3-8B、Mistral-7B / 24B、DeepSeek-Math-7B、Qwen2.5-math-7Bなど、10の異なるベースモデルを使用してゼロRL学習を行いました。フォーマット補償調整やクエリ難易度制御などの戦略により、ほとんどの設定で推論精度と応答長が大幅に向上しました。しかし、学習ダイナミクスモニタリングにより、異なるベースモデルが独自の学習パターンを示すことが確認されました。たとえば、応答長の増加は、常に検証などの特定の認知行動の出現と相関しているわけではありません。具体的には、Qwenファミリーではなく、小規模モデルで初めて「aha moment」を観察しました。成功したゼロRL学習を可能にするコア設計と研究結果、そして実務経験を共有し、コード、モデル、分析ツールをオープンソースで公開します。

Takeaways、Limitations

•

Takeaways：

◦

さまざまな基盤モデルでゼロRL学習の有効性を検証し、成功した学習のためのコア設計戦略を提案しました。

◦

フォーマット補償の調整とクエリの難易度制御により、推論精度と応答長の改善がもたらされました。

◦

Qwen系以外の小規模モデルでも「aha moment」を観察し,モデルアーキテクチャの多様性とゼロRL学習の適用可能性を確認した。

◦

コード、モデル、分析ツールをオープンソースとして公開し、その後の研究を支援します。

•

Limitations：

◦

応答の長さの増加と認知行動の出現との間の相関関係が常に一致していないことを示すなど、モデル学習プロセスの深い理解が欠けている可能性があります。

◦

使用されているベースモデルの多様性にもかかわらず、特定のモデルシリーズに偏りがある可能性があります。

◦

「Aha moment」の定義と測定方法の明確な基準が必要な場合があります。

Made with Slashpage