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。

AC-DiT: Adaptive Coordination Diffusion Transformer for Mobile Manipulation

Self-Guided Process Reward Optimization with Redefined Step-wise Advantage for Process Reinforcement Learning

Crafting Hanzi as Narrative Bridges: An AI Co-Creation Workshop for Elderly Migrants

Distributional Soft Actor-Critic with Diffusion Policy

Skywork-Reward-V2: Scaling Preference Data Curation via Human-AI Synergy

Fast AI Model Splitting over Edge Networks

From Sentences to Sequences: Rethinking Languages in Biological System

MTCNet: Motion and Topology Consistency Guided Learning for Mitral Valve Segmentationin 4D Ultrasound

Horus: A Protocol for Trustless Delegation Under Uncertainty

Mixture of Reasonings: Teach Large Language Models to Reason with Adaptive Strategies

Benchmarking Generalizable Bimanual Manipulation: RoboTwin Dual-Arm Collaboration Challenge at CVPR 2025 MEIS Workshop

Red Teaming for Generative AI, Report on a Copyright-Focused Exercise Completed in an Academic Medical Center

AirV2X: Unified Air-Ground Vehicle-to-Everything Collaboration

Semantic Structure-Aware Generative Attacks for Enhanced Adversarial Transferability

Aligning Frozen LLMs by Reinforcement Learning: An Iterative Reweight-then-Optimize Approach

Distinguishing Predictive and Generative AI in Regulation

AIn't Nothing But a Survey? Using Large Language Models for Coding German Open-Ended Survey Responses on Survey Motivation

Text-Aware Image Restoration with Diffusion Models

How Good LLM-Generated Password Policies Are?

Towards an Explainable Comparison and Alignment of Feature Embeddings

Gradient-Based Model Fingerprinting for LLM Similarity Detection and Family Classification

Empowering Intelligent Low-altitude Economy with Large AI Model Deployment

Incorporating LLMs for Large-Scale Urban Complex Mobility Simulation

Generating Hypotheses of Dynamic Causal Graphs in Neuroscience: Leveraging Generative Factor Models of Observed Time Series

Traveling Across Languages: Benchmarking Cross-Lingual Consistency in Multimodal LLMs

Threat Modeling for AI: The Case for an Asset-Centric Approach

SoccerDiffusion: Toward Learning End-to-End Humanoid Robot Soccer from Gameplay Recordings

PAD: Phase-Amplitude Decoupling Fusion for Multi-Modal Land Cover Classification

Significativity Indices for Agreement Values

Transferrable Surrogates in Expressive Neural Architecture Search Spaces

Privacy-Preserving Operating Room Workflow Analysis using Digital Twins

Uncertainty-Guided Coarse-to-Fine Tumor Segmentation with Anatomy-Aware Post-Processing

CMD-HAR: Cross-Modal Disentanglement for Wearable Human Activity Recognition

Commander-GPT: Fully Unleashing the Sarcasm Detection Capability of Multi-Modal Large Language Models

Understanding-informed Bias Mitigation for Fair CMR Segmentation

HAPI: A Model for Learning Robot Facial Expressions from Human Preferences

MaizeField3D: A Curated 3D Point Cloud and Procedural Model Dataset of Field-Grown Maize from a Diversity Panel

Illuminant and light direction estimation using Wasserstein distance method

Fundamental Limits of Hierarchical Secure Aggregation with Cyclic User Association

LLM-Powered Prediction of Hyperglycemia and Discovery of Behavioral Treatment Pathways from Wearables and Diet

Interleaved Gibbs Diffusion: Generating Discrete-Continuous Data with Implicit Constraints

EquiTabPFN: A Target-Permutation Equivariant Prior Fitted Networks

Circuit-tuning: A Mechanistic Approach for Identifying Parameter Redundancy and Fine-tuning Neural Networks

EigenLoRAx: Recycling Adapters to Find Principal Subspaces for Resource-Efficient Adaptation and Inference

Learning Traffic Anomalies from Generative Models on Real-Time Observations

Enabling Population-Level Parallelism in Tree-Based Genetic Programming for Comprehensive GPU Acceleration

Parameters vs FLOPs: Scaling Laws for Optimal Sparsity for Mixture-of-Experts Language Models

Quantifying the Importance of Data Alignment in Downstream Model Performance

Quantum-enhanced causal discovery for a small number of samples

On Characterizations for Language Generation: Interplay of Hallucinations, Breadth, and Stability

Token Prepending: A Training-Free Approach for Eliciting Better Sentence Embeddings from LLMs

COEF-VQ: Cost-Efficient Video Quality Understanding through a Cascaded Multimodal LLM Framework

GeMID: Generalizable Models for IoT Device Identification

Next-Token Prediction Task Assumes Optimal Data Ordering for LLM Training in Proof Generation

Is Complex Query Answering Really Complex?

Aerial Vision-and-Language Navigation via Semantic-Topo-Metric Representation Guided LLM Reasoning

Offline Reinforcement Learning for Learning to Dispatch for Job Shop Scheduling

Reconsidering the energy efficiency of spiking neural networks

Exploring the Integration of Large Language Models in Industrial Test Maintenance Processes

Sequence-aware Pre-training for Echocardiography Probe Movement Guidance

Anatomical Foundation Models for Brain MRIs

Learning From Crowdsourced Noisy Labels: A Signal Processing Perspective

Quantifying the Cross-sectoral Intersecting Discrepancies within Multiple Groups Using Latent Class Analysis Towards Fairness

Delving into LLM-assisted writing in biomedical publications through excess vocabulary

Towards a Novel Measure of User Trust in XAI Systems

Avoiding Catastrophe in Online Learning by Asking for Help

Improving the Robustness of Distantly-Supervised Named Entity Recognition via Uncertainty-Aware Teacher Learning and Student-Student Collaborative Learning

Beyond Scale: The Diversity Coefficient as a Data Quality Metric for Variability in Natural Language Data

Kernel Density Bayesian Inverse Reinforcement Learning

Embodied AI Agents: Modeling the World

Mind2Web 2: Evaluating Agentic Search with Agent-as-a-Judge

AI Flow: Perspectives, Scenarios, and Approaches

A framework for Conditional Reasoning in Answer Set Programming

Autoformalization in the Era of Large Language Models: A Survey

Agentic AI Process Observability: Discovering Behavioral Variability

Artificial Intelligence Index Report 2025

MAPS: Advancing Multi-Modal Reasoning in Expert-Level Physical Science

XGeM: A Multi-Prompt Foundation Model for Multimodal Medical Data Generation

Direct Preference Optimization Using Sparse Feature-Level Constraints

Unsupervised Cognition

Urban Region Pre-training and Prompting: A Graph-based Approach

Road Graph Generator: Mapping roads at construction sites from GPS データ

Point3R: Streaming 3D Reconstruction with Explicit Spatial Pointer Memory

LiteReality: Graphics-Ready 3D Scene Reconstruction from RGB-D Scans

Answer Matching Outperforms Multiple Choice for Language Model Evaluation

Subtyping in DHOL - Extended preprint

MOTIF: Modular Thinking via Reinforcement Fine-tuning in LLMs

USAD: An Unsupervised Data Augmentation Spatio-Temporal Attention Diffusion Network

DNN-Based Precoding in RIS-Aided mmWave MIMO Systems With Practical Phase Shift

SynapseRoute: An Auto-Route Switching Framework on Dual-State Large Language Model

Self-Correction Bench: Revealing and Addressing the Self-Correction Blind Spot in LLMs

Multi-agent Auditory Scene Analysis

Fast and Simplex: 2-Simplicial Attention in Triton

Synthesizable by Design: A Retrosynthesis-Guided Framework for Molecular Analog Generation

Linear Attention with Global Context: A Multipole Attention Mechanism for Vision and Physics

Early Signs of Steganographic Capabilities in Frontier LLMs

Meta SecAlign: A Secure Foundation LLM Against Prompt Injection Attacks

FairHuman: Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models

APT: Adaptive Personalized Training for Diffusion Models with Limited Data

ASDA: Audio Spectrogram Differential Attention Mechanism for Self-Supervised Representation Learning

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

Created by

Haebom

作者

Soumya Rani Samineni, Durgesh Kalwar, Karthik Valmeekam, Kaya Stechly, Subbarao Kambhampati

概要

この論文は、強化学習ベースの大規模言語モデル（LLM）事後学習、特にDeepSeek R1でGRPOを適用した方法について批判的に検討します。 LLM学習をマルコフ意思決定過程(MDP)でモデル化する際の一般的な構造的仮定(1. MDP状態を行動の連結として定義し、状態をコンテキストウィンドウとし、行動をトークンとみなし、2.状態-行動経路の補償を均等に分配)がMDPを退化させて強化学習/GRPO装置が不要にするという点をこの単純化された仮定は、強化学習ベースのアプローチが結果主導の指導学習と効果的に同等になることを示す包括的な分析を提供します。 GSM8KとCountdownのベンチマークを使用したQwen-2.5基本モデル実験では、ポジティブサンプルとネガティブサンプルの両方を統合した反復マップ学習の微調整は、GRPOベースの学習と比較してパフォーマンスを達成することを示しています。さらに、この構造的仮定は、間接的にＲＬがより長い中間トークンシーケンスを生成するように誘導し、「ＲＬがより長い事故の兆候を生成する」という記述を強化することを論じる。強化学習はLLMの推論能力を向上させるのに非常に有用な技術であるかもしれませんが、基本的なMDPをモデル化するときの単純化された構造的仮定は、広く使用されているLLM RLフレームワークと解釈に疑問を抱くことを示しています。

Takeaways、Limitations

•

Takeaways： LLM事後学習における強化学習の効果に関する既存の研究の構造的仮定の批判的レビューを提供します。単純な指導学習微調整でも強化学習ベースの方法と同様の性能を達成できることを実験的に示す。強化学習はLLM推論能力の向上に貢献するかもしれませんが、現在のMDPモデリングスキームの限界を指摘しています。

•

Limitations：特定のLLMとベンチマークの実験結果のみが提示され、一般化の可能性に関するさらなる研究が必要です。強化学習の他の構造的仮定や設計の分析が不足している。より複雑な推論課題に対する性能比較が必要である。

Made with Slashpage