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。

HoPE: Hyperbolic Rotary Positional Encoding for Stable Long-Range Dependency Modeling in Large Language Models

Comparative Analysis of Transformer Models in Disaster Tweet Classification for Public Safety

Emergent Social Dynamics of LLM Agents in the El Farol Bar Problem

The Good, the Bad and the Constructive: Automatically Measuring Peer Review's Utility for Authors

Energy Landscapes Enable Reliable Abstention in Retrieval-Augmented Large Language Models for Healthcare

DEXOP: A Device for Robotic Transfer of Dexterous Human Manipulation

Reinforcement Learning for Robust Ageing-Aware Control of Li-ion Battery Systems with Data-Driven Formal Verification

RepoDebug: Repository-Level Multi-Task and Multi-Language Debugging Evaluation of Large Language Models

Gravity Well Echo Chamber Modeling With An LLM-Based Confirmation Bias Model

Insights from Gradient Dynamics: Gradient Autoscaled Normalization

Efficient Virtuoso: A Latent Diffusion Transformer Model for Goal-Conditioned Trajectory Planning

MoSEs: Uncertainty-Aware AI-Generated Text Detection via Mixture of Stylistics Experts with Conditional Thresholds

DCPO: Dynamic Clipping Policy Optimization

DSDE: Dynamic Speculative Decoding with KLD Stability for Real-World Serving

Can AI be Auditable?

Robotic Fire Risk Detection based on Dynamic Knowledge Graph Reasoning: An LLM-Driven Approach with Graph Chain-of-Thought

Navigating the EU AI Act: Foreseeable Challenges in Qualifying Deep Learning-Based Automated Inspections of Class III Medical Devices

Complementary Learning System Empowers Online Continual Learning of Vehicle Motion Forecasting in Smart Cities

MultiPL-MoE: Multi-Programming-Lingual Extension of Large Language Models through Hybrid Mixture-of-Experts

QuadKAN: KAN-Enhanced Quadruped Motion Control via End-to-End Reinforcement Learning

MovieCORE: COgnitive REasoning in Movies

Automatic Prompt Optimization with Prompt Distillation

Membership Inference Attacks on LLM-based Recommender Systems

Leveraging Large Language Models for Accurate Sign Language Translation in Low-Resource Scenarios

Group Expectation Policy Optimization for Heterogeneous Reinforcement Learning

Convergence and Generalization of Anti-Regularization for Parametric Models

Jet-Nemotron: Efficient Language Model with Post Neural Architecture Search

CARFT: Boosting LLM Reasoning via Contrastive Learning with Annotated Chain-of-Thought-based Reinforced Fine-Tuning

Bridging Generalization and Personalization in Human Activity Recognition via On-Device Few-Shot Learning

FinAgentBench: A Benchmark Dataset for Agentic Retrieval in Financial Question Answering

Using Artificial Intuition in Distinct, Minimalist Classification of Scientific Abstracts for Management of Technology Portfolios

Semantic Discrepancy-aware Detector for Image Forgery Identification

Quantum-Efficient Reinforcement Learning Solutions for Last-Mile On-Demand Delivery

BadPromptFL: A Novel Backdoor Threat to Prompt-based Federated Learning in Multimodal Models

Uncertainty-Driven Reliability: Selective Prediction and Trustworthy Deployment in Modern Machine Learning

Real-Time Analysis of Unstructured Data with Machine Learning on Heterogeneous Architectures

VSI: Visual Subtitle Integration for Keyframe Selection to enhance Long Video Understanding

SGDFuse: SAM-Guided Diffusion for High-Fidelity Infrared and Visible Image Fusion

An Efficient Continuous-Time MILP for Integrated Aircraft Hangar Scheduling and Layout

DIRF: A Framework for Digital Identity Protection and Clone Governance in Agentic AI Systems

COLLAGE: Adaptive Fusion-based Retrieval for Augmented Policy Learning

Dynamically Adaptive Reasoning via LLM-Guided MCTS for Efficient and Context-Aware KGQA

Nested Graph Pseudo-Label Refinement for Noisy Label Domain Adaptation Learning

LanternNet: A Hub-and-Spoke System to Seek and Suppress Spotted Lanternfly Populations

RecPS: Privacy Risk Scoring for Recommender Systems

Supervised Fine Tuning on Curated Data is Reinforcement Learning (and can be improved)

Role-Playing LLM-Based Multi-Agent Support Framework for Detecting and Addressing Family Communication Bias

PLAME: Lightweight MSA Design Advances Protein Folding From Evolutionary Embeddings

Driver-Net: Multi-Camera Fusion for Assessing Driver Take-Over Readiness in Automated Vehicles

Leveraging Out-of-Distribution Unlabeled Images: Semi-Supervised Semantic Segmentation with an Open-Vocabulary Model

Visual Structures Helps Visual Reasoning: Addressing the Binding Problem in VLMs

Precise Bayesian Neural Networks

Transit for All: Mapping Equitable Bike2Subway Connection using Region Representation Learning

Scaling Intelligence: Designing Data Centers for Next-Gen Language Models

Image Segmentation with Large Language Models: A Survey with Perspectives for Intelligent Transportation Systems

SAIL: Faster-than-Demonstration Execution of Imitation Learning Policies

Persona-driven Simulation of Voting Behavior in the European Parliament with Large Language Models

Bipedal Balance Control with Whole-body Musculoskeletal Standing and Falling Simulations

Scaling Laws of Motion Forecasting and Planning - Technical Report

Efficient $Q$-Learning and Actor-Critic Methods for Robust Average Reward Reinforcement Learning

Who Gets Credit or Blame? Attributing Accountability in Modern AI Systems

Unsupervised Evolutionary Cell Type Matching via Entropy-Minimized Optimal Transport

Multi-output Classification using a Cross-talk Architecture for Compound Fault Diagnosis of Motors in Partially Labeled Condition

SoloSpeech: Enhancing Intelligibility and Quality in Target Speech Extraction through a Cascaded Generative Pipeline

Steering LLM Reasoning Through Bias-Only Adaptation

MetaSTH-Sleep: Towards Effective Few-Shot Sleep Stage Classification for Health Management with Spatial-Temporal Hypergraph Enhanced Meta-Learning

InterFeat: A Pipeline for Finding Interesting Scientific Features

HumaniBench: A Human-Centric Framework for Large Multimodal Models Evaluation

Advancing Scientific Text Classification: Fine-Tuned Models with Dataset Expansion and Hard-Voting

Test It Before You Trust It: Applying Software Testing for Trustworthy In-context Learning

Action Flow Matching for Continual Robot Learning

Addressing Concept Mislabeling in Concept Bottleneck Models Through Preference Optimization

Nemotron-H: A Family of Accurate and Efficient Hybrid Mamba-Transformer Models

Byzantine-Robust Federated Learning Using Generative Adversarial Networks

Beyond SHAP and Anchors: A large-scale experiment on how developers struggle to design meaningful end-user explanations

VIPER: Visual Perception and Explainable Reasoning for Sequential Decision-Making

DistJoin: A Decoupled Join Cardinality Estimator based on Adaptive Neural Predicate Modulation

Low-Confidence Gold: Refining Low-Confidence Samples for Efficient Instruction Tuning

Assistance or Disruption? Exploring and Evaluating the Design and Trade-offs of Proactive AI Programming Support

Soft Token Attacks Cannot Reliably Audit Unlearning in Large Language Models

CHIRLA: Comprehensive High-resolution Identification and Re-identification for Large-scale Analysis

Kolmogorov-Arnold Fourier Networks

Position: LLMs Can be Good Tutors in English Education

Predicting Steady-State Behavior in Complex Networks with Graph Neural Networks

Separate Motion from Appearance: Customizing Motion via Customizing Text-to-Video Diffusion Models

Motion-enhanced Cardiac Anatomy Segmentation via an Insertable Temporal Attention Module

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

OmniThink: Expanding Knowledge Boundaries in Machine Writing through Thinking

DispFormer: A Pretrained Transformer Incorporating Physical Constraints for Dispersion Curve Inversion

Integrating Evidence into the Design of XAI and AI-based Decision Support Systems: A Means-End Framework for End-users in Construction

Revealing the impact of synthetic native samples and multi-tasking strategies in Hindi-English code-mixed humour and sarcasm detection

Neural Port-Hamiltonian Differential Algebraic Equations for Compositional Learning of Electrical Networks

Sequential Controlled Langevin Diffusions

Privacy-Preserving Federated Learning via Homomorphic Adversarial Networks

CAREL: Instruction-guided reinforcement learning with cross-modal auxiliary objectives

Lessons from Studying Two-Hop Latent Reasoning

HierTOD: A Task-Oriented Dialogue System Driven by Hierarchical Goals

Flexible Coded Distributed Convolution Computing for Enhanced Straggler Resilience and Numerical Stability in Distributed CNNs

FACEGroup: Feasible and Actionable Counterfactual Explanations for Group Fairness

ETF: An Entity Tracing Framework for Hallucination Detection in Code Summaries

DCPO: Dynamic Clipping Policy Optimization

Created by

Haebom

作者

Shihui Yang, Chengfeng Dou, Peidong Guo, Kai Lu, Qiang Ju, Fei Deng, Rihui Xin

概要

本稿では、強化学習による大規模言語モデルの推論能力を向上させる新しいフレームワークである動的クリッピングポリシー最適化（DCPO）を提案します。既存の方法であるGRPOのゼロ勾配問題を解決するために、トークン別の事前確率に基づく動的クリッピング戦略と累積訓練段階にわたる滑らかな利点標準化技術を導入しました。 DCPOは、4つのモデルに基づいた4つのベンチマークで最先端のパフォーマンスを達成し、特にAIME24およびAIME25ベンチマークで従来の方法であるGRPO、DAPO、GSPOを上回る結果を示しました。また、GRPOに比べて非営利勾配比率を平均28%向上させ、DAPOより訓練効率を2倍高め、トークンクリッピング率を大幅に減少させました。

Takeaways、Limitations

•

Takeaways：

◦

大規模言語モデルの強化学習におけるゼロ勾配問題を効果的に解決する新しい方法を提示

◦

動的クリッピング戦略と滑らかな利点標準化手法によって生成されたデータをより効率的に活用

◦

さまざまなベンチマークで既存の方法を上回る優れたパフォーマンスを実現

◦

トレーニング効率とトークンクリッピング率の向上

•

Limitations：

◦

提示された方法の一般化性能に関するさらなる研究の必要性

◦

さまざまなモデルとベンチマークの追加実験が必要

◦

動的クリッピング戦略のパラメータ調整に関する追加研究の必要性

Made with Slashpage