牛哥精选 · 半年

1

📝 深度技术 arXiv 机器学习 2026-06-26

RolloutPipe: Overlapping Pipelined Rollout and Training in Disaggregated On-Policy LLM Reinforcement Learning

解耦策略学习瓶颈：论文提出RolloutPipe，通过流水线重叠技术并行化LLM强化学习的推理与训练，提升效率。

arXiv:2606.26997v1 Announce Type: cross Abstract: Large language model (LLM) post-training for reasoning increasingly relies on reinforcement learning…

llm强化学习流水线化推理训练重叠解耦架构 on-policy

2

📝 深度技术 Dev.to 2026-06-22

Astro Islands Architecture: When to Use Partial Hydration

一个真实案例揭示过度的React水合让页面加载了180kb无用JS，Astro Islands架构教你精准优化前端性能。

Astro Islands Architecture: When to Use Partial Hydration I've shipped enough full SPAs to know their dirty secret: they're overly expensive for most …

astro islands ar partial hy 前端性能 javascript

3

📝 深度技术 arXiv AI 2026-06-17

Correct When Paired, Wrong When Split: Decoupling and Editing Modality-Specific Neurons in MLLMs

多模态大语言模型在配对时表现准确、拆分时出错，研究通过解耦编辑特定模态神经元来修复这一现象。

arXiv:2606.17057v1 Announce Type: cross Abstract: Although Knowledge Editing provides an efficient mechanism for updating the knowledge of Multimodal …

多模态大模型神经元解耦模型编辑特定模态 mllms

4

🤖 AI·大模型 arXiv AI 2026-06-16

RollArt: Disaggregated Multi-Task Agentic RL Training at Scale

提出RollArt解耦架构，实现大规模多任务Agentic RL训练，显著提升效率与扩展性。

arXiv:2512.22560v2 Announce Type: replace-cross Abstract: Agentic Reinforcement Learning (RL) trains LLMs through multi-turn interactions with environ…

强化学习智能体多任务学习大规模训练解耦架构

5

📝 深度技术 arXiv 机器学习 2026-06-15

Natively Unlearnable Large Language Models

提出原生不可遗忘的LLM设计，通过解耦参数解决训练数据来源纠缠难题，为模型安全提供新思路。

arXiv:2606.13873v1 Announce Type: new Abstract: Unlearning aims to remove the influence of specific training data sources, but this has proved challen…

llm 遗忘学习参数解耦模型安全机器学习

6

📝 深度技术 arXiv 计算机视觉 2026-06-05

Dual Feature Decoupling for Fine-Grained OOD Detection

提出双特征解耦方法，专攻细粒度分布外检测难题，让模型精准识别细微异常。

arXiv:2606.05536v1 Announce Type: new Abstract: Out-of-distribution detection (OOD) is an indispensable technique when applying machine learning model…

ood检测细粒度特征解耦深度学习异常检测

7

📝 深度技术 arXiv NLP 2026-06-04

SparDA: Sparse Decoupled Attention for Efficient Long-Context LLM Inference

稀疏注意力新突破：解耦选择与计算，解决KV缓存瓶颈与PCIe传输困境。

arXiv:2606.04511v1 Announce Type: new Abstract: Sparse attention reduces compute and memory bandwidth for long-context LLM inference. However, two key…

稀疏注意力长上下文 llm推理 kv缓存效率优化

8

📝 深度技术 arXiv AI 2026-06-02

AlphaToken: Decoupling Adaptation and Stability for Path-Aware Response Token Valuation in LLM Post-Training

LLM后训练新突破：AlphaToken通过路径感知机制对响应token进行估值，巧妙解耦适应性与稳定性，提升模型微调效率与鲁棒性。

arXiv:2606.01635v1 Announce Type: cross Abstract: Token selection is pivotal for effective LLM post-training. However, existing methods mostly rely on…

alphatoken llm后训练路径感知 token估值适应性

9

📝 深度技术 arXiv 计算机视觉 2026-05-28

Decoupled Training with Local Reinforcement Fine-Tuning in Federated Learning

联邦学习新范式：解耦训练结合本地强化微调，有效平衡通信效率与个性化模型性能，被ICML 2026收录。

arXiv:2605.27900v1 Announce Type: new Abstract: Federated Learning (FL) with pre-trained Vision-Language Models (VLMs) has emerged as a promising para…

联邦学习解耦训练强化微调个性化 icml 2026

10

📝 深度技术 arXiv 计算机视觉 2026-05-20

HiDe: Rethinking The Zoom-IN method in High Resolution MLLMs via Hierarchical Decoupling

重新思考高分辨率多模态大模型中的Zoom-IN方法，提出Hierarchical Decoupling框架，显著提升视觉理解性能。

arXiv:2510.00054v2 Announce Type: replace Abstract: Multimodal Large Language Models (MLLMs) have made significant strides in visual understanding tas…

mllms 高分辨率图像层次解耦视觉理解 zoon-in方法

11

📝 深度技术 arXiv 计算机视觉 2026-05-20

Semantics Disentanglement and Composition for Universal Image Coding with Efficiently LLM Reasoning and Generative Diffusion

将语义解耦与LLM推理、扩散生成融合，实现通用图像编码新范式。

arXiv:2412.18158v2 Announce Type: replace Abstract: Learned image compression methods have shown impressive performance but are often highly specializ…

通用图像编码语义解耦 llm推理扩散模型图像压缩

12

📝 深度技术 arXiv NLP 2026-05-20

Implicit Hierarchical GRPO: Decoupling Tool Invocation from Execution for Tool-Integrated Mathematical Reasoning

将工具调用与执行解耦，提出隐式层次化GRPO框架，显著提升数学推理中的工具集成效率与泛化能力。

arXiv:2605.18500v1 Announce Type: new Abstract: Large language models (LLMs) have increasingly leveraged tool invocation to enhance their reasoning ca…

grpo 工具集成数学推理层次化强化学习解耦设计

13

📝 深度技术 arXiv 机器学习 2026-05-20

Matrix-Decoupled Concentration for Autoregressive Sequences: Dimension-Free Guarantees for Sparse Long-Context Rewards

自回归序列的矩阵解耦集中不等式，为稀疏长上下文奖励提供无维度保证，理论创新突破。

arXiv:2605.06017v2 Announce Type: replace Abstract: Sequence-level evaluations in autoregressive Large Language Models (LLMs) rely on highly dependent…

自回归序列集中不等式矩阵解耦无维度保证长上下文奖励

14

📝 深度技术 arXiv AI 2026-05-19

Decouple Searching from Training: Scaling Data Mixing via Model Merging for Large Language Model Pre-training

提出通过模型合并解耦数据混合搜索与训练，高效扩展LLM预训练的数据配比策略。

arXiv:2602.00747v2 Announce Type: replace-cross Abstract: Determining an effective data mixture is a key factor in Large Language Model (LLM) pre-trai…

llm预训练数据混合模型合并解耦搜索

15

📝 深度技术 arXiv AI 2026-05-19

DiLA: Disentangled Latent Action World Models

DiLA将潜在动作解耦为几何与纹理流，实现高保真视频预测，突破LAMs的抽象-保真权衡。

arXiv:2605.15725v1 Announce Type: cross Abstract: Latent Action Models (LAMs) enable the learning of world models from unlabeled video by inferring ab…

dila 解耦潜在动作世界模型自监督学习计算机视觉

16

🤖 AI·大模型 arXiv AI 2026-05-19

Structure-BiEval: A Self-Supervised, Dual-Track Framework for Decoupling Structure and Content in LLM Evaluation for Web Information Systems

自监督双轨框架首次解耦结构与内容，精准评估LLM在Web系统中的表现。

arXiv:2601.19923v2 Announce Type: replace-cross Abstract: As Large Language Models (LLMs) evolve into the core of Web-based autonomous agents and comp…

llm评估自监督学习双轨框架 web信息系统结构内容解耦

17

📝 深度技术 arXiv AI 2026-05-19

Mechanical Enforcement for LLM Governance:Evidence of Governance-Task Decoupling in Financial Decision Systems

一篇揭示LLM金融治理中“治理-任务解耦”的突破性论文。作者提出机械执行四原语，将合规决策从模型视角剥离，使任务准确率从0.43飙升至0.88，同时大幅提升可审计性——治理质量不再随任务性能崩溃，从根本上重塑了可信AI治理范式。

arXiv:2605.14744v1 Announce Type: cross Abstract: Large language models in regulated financial workflows are governed by natural-language policies tha…

llm 治理任务解耦机械执行

🐂 牛哥精选

RolloutPipe: Overlapping Pipelined Rollout and Training in Disaggregated On-Policy LLM Reinforcement Learning

Astro Islands Architecture: When to Use Partial Hydration

Correct When Paired, Wrong When Split: Decoupling and Editing Modality-Specific Neurons in MLLMs

RollArt: Disaggregated Multi-Task Agentic RL Training at Scale

Natively Unlearnable Large Language Models

Dual Feature Decoupling for Fine-Grained OOD Detection

SparDA: Sparse Decoupled Attention for Efficient Long-Context LLM Inference

AlphaToken: Decoupling Adaptation and Stability for Path-Aware Response Token Valuation in LLM Post-Training

Decoupled Training with Local Reinforcement Fine-Tuning in Federated Learning

HiDe: Rethinking The Zoom-IN method in High Resolution MLLMs via Hierarchical Decoupling

Semantics Disentanglement and Composition for Universal Image Coding with Efficiently LLM Reasoning and Generative Diffusion

Implicit Hierarchical GRPO: Decoupling Tool Invocation from Execution for Tool-Integrated Mathematical Reasoning

Matrix-Decoupled Concentration for Autoregressive Sequences: Dimension-Free Guarantees for Sparse Long-Context Rewards

Decouple Searching from Training: Scaling Data Mixing via Model Merging for Large Language Model Pre-training

DiLA: Disentangled Latent Action World Models

Structure-BiEval: A Self-Supervised, Dual-Track Framework for Decoupling Structure and Content in LLM Evaluation for Web Information Systems

Mechanical Enforcement for LLM Governance:Evidence of Governance-Task Decoupling in Financial Decision Systems

📅 日期