牛哥精选 · 一年

1

📝 深度技术 arXiv 机器学习 2026-05-20

Divergence-Suppressing Couplings for Rectified Flow

提出抑制散度的耦合方法用于整流流，改善生成模型的稳定性和效率

arXiv:2605.17733v1 Announce Type: cross Abstract: The promise of Rectified Flow rests on producing self-generated couplings whose trajectories are str…

rectified divergence couplings 生成模型扩散模型

2

📝 深度技术 arXiv 机器学习 2026-05-20

Uncertainty-Calibrated Recommendations for Low-Active Users

针对低活跃用户的推荐不确定性校准方法，有效提升冷启动场景可靠性，已被KDD 2026接收。

arXiv:2605.17788v1 Announce Type: cross Abstract: A fundamental challenge in recommender systems is balancing reliability for Low-Active Users (LAUs) …

不确定性校准推荐系统低活跃用户冷启动 kdd

3

🤖 AI·大模型 arXiv 机器学习 2026-05-20

What is Holding Back Latent Visual Reasoning?

探究视觉语言模型中潜在视觉推理的瓶颈，揭示人类式中间视觉步骤的模拟障碍

arXiv:2605.18445v1 Announce Type: cross Abstract: Humans can approach complex visual problems by mentally simulating intermediate visual steps, rather…

潜在视觉推理视觉语言模型链式推理连续潜在标记视觉想象

4

📝 深度技术 arXiv 机器学习 2026-05-20

Predictable Confabulations: Factual Recall by LLMs Scales with Model Size and Topic Frequency

研究发现LLM的幻觉并非随机，而是与模型规模和主题频率呈可预测的比例关系。

arXiv:2605.18732v1 Announce Type: cross Abstract: While scaling laws govern aggregate large language model performance, no scaling law has linked fact…

llm 幻觉模型规模主题频率可预测性

5

📝 深度技术 arXiv 机器学习 2026-05-20

DashAttention: Differentiable and Adaptive Sparse Hierarchical Attention

提出可微分自适应稀疏层次注意力机制，显著提升长序列建模效率与计算可扩展性

arXiv:2605.18753v1 Announce Type: cross Abstract: Current hierarchical attention methods, such as NSA and InfLLMv2, select the top-k relevant key-valu…

注意力机制稀疏注意力可微分自适应层次注意力 dashattent

6

📝 深度技术 arXiv 机器学习 2026-05-20

Transformers for Learning on Noisy and Task-Level Manifolds: Approximation and Generalization Insights

从理论层面揭示Transformer在噪声与任务级流形上的学习能力，近似与泛化分析带来新洞察

arXiv:2505.03205v3 Announce Type: replace Abstract: Transformers serve as the foundational architecture for large language and video generation models…

transforme 噪声流形任务级流形泛化理论近似理论

7

📝 深度技术 arXiv 机器学习 2026-05-20

LLM Agents Are the Antidote to Walled Gardens

提出LLM智能体如何打破平台壁垒，重塑开放互联网生态。

arXiv:2506.23978v3 Announce Type: replace Abstract: While the Internet's core infrastructure was designed to be open and universal, today's applicatio…

llm agents 围墙花园开放api 互操作性平台垄断

8

📝 深度技术 arXiv 机器学习 2026-05-20

CoUn: Empowering Machine Unlearning via Contrastive Learning

用对比学习提升机器遗忘效果，解决现有方法遗忘不彻底的问题，为模型数据移除提供新思路。

arXiv:2509.16391v3 Announce Type: replace Abstract: Machine unlearning (MU) aims to remove the influence of specific "forget" data from a trained mode…

machine un contrastiv 模型遗忘数据移除深度学习

9

📝 深度技术 arXiv 机器学习 2026-05-20

The Bayesian Geometry of Transformer Attention

从贝叶斯几何视角重新阐释Transformer注意力机制，揭示其内在概率结构。

arXiv:2512.22471v5 Announce Type: replace Abstract: Transformers often appear to perform Bayesian reasoning in context, but verifying this rigorously …

transforme 注意力机制贝叶斯几何理论分析深度学习

10

📝 深度技术 arXiv 机器学习 2026-05-20

KASER: Knowledge-Aligned Student Error Simulator for Open-Ended Coding Tasks

ACL 2026 论文：知识对齐的学生错误模拟器，精准生成开放式编程任务中的学生典型错误，助力智能教育评估

arXiv:2601.06633v2 Announce Type: replace Abstract: Open-ended tasks, such as coding problems that are common in computer science education, provide d…

kaser 知识对齐学生错误模拟编程任务错误生成

11

📝 深度技术 arXiv 机器学习 2026-05-20

NanoQuant: Efficient Sub-1-Bit Quantization of Large Language Models

提出子1比特量化方法，大幅降低大语言模型存储与计算开销，兼顾效率与性能。

arXiv:2602.06694v2 Announce Type: replace Abstract: Weight-only quantization has become a standard approach for efficiently serving large language mod…

nanoquant 大模型量化子1比特高效量化 llm

12

📝 深度技术 arXiv 机器学习 2026-05-20

Meltdown: Circuits and Bifurcations in Point-Cloud-Conditioned 3D Diffusion Transformers

稀疏点云条件3D扩散模型中的电路与分岔现象，揭示安全关键重建的底层机制

arXiv:2602.11130v2 Announce Type: replace Abstract: Sparse point clouds are a common input modality for 3D surface reconstruction, including in safety…

3d扩散模型点云条件表面重建电路分析分岔理论

13

📝 深度技术 arXiv 机器学习 2026-05-20

Self-Play Only Evolves When Self-Synthetic Pipeline Ensures Learnable Information Gain

揭示自我对弈仅在自合成数据提供可学习信息增益时才有效演化，为AI训练策略提供关键理论指导。

arXiv:2603.02218v2 Announce Type: replace Abstract: Large language models (LLMs) make it plausible to build systems that improve through self-evolving…

self-play 自合成管线可学习信息增益 ai训练理论分析

14

📝 深度技术 arXiv 机器学习 2026-05-20

SlimQwen: Exploring the Pruning and Distillation in Large MoE Model Pre-training

揭秘MoE大模型预训练中剪枝与蒸馏技术，SlimQwen优化效率与性能。

arXiv:2605.08738v2 Announce Type: replace Abstract: Structured pruning and knowledge distillation (KD) are typical techniques for compressing large la…

moe 剪枝蒸馏预训练大模型

15

📝 深度技术 arXiv 机器学习 2026-05-20

Unlearning Isn't Deletion: Investigating Reversibility of Machine Unlearning in LLMs

LLM的机器遗忘并非真正删除，最新研究揭示了其可逆性，挑战现有安全假设。

arXiv:2505.16831v3 Announce Type: replace-cross Abstract: Unlearning in large language models (LLMs) aims to remove specified data, but its efficacy i…

llm 机器遗忘可逆性深度学习

16

📝 深度技术 arXiv 机器学习 2026-05-20

Prompt reinforcing for long-term planning of large language models

新方法Prompt Reinforcing提升大模型长期规划能力，精准解决复杂任务。

arXiv:2510.05921v3 Announce Type: replace-cross Abstract: Large language models (LLMs) have achieved remarkable success in a wide range of natural lan…

大语言模型长期规划提示增强推理方法

17

📝 深度技术 arXiv 机器学习 2026-05-20

Enhancing Cloud Network Resilience via a Robust LLM-Empowered Multi-Agent Reinforcement Learning Framework

融合LLM与多智能体强化学习，打造弹性的云网络防御框架，有效应对虚拟化带来的安全挑战。

arXiv:2601.07122v2 Announce Type: replace-cross Abstract: While virtualization and resource pooling empower cloud networks with structural flexibility…

llm 多智能体强化学习云网络弹性网络安全资源隔离

18

📝 深度技术 arXiv 机器学习 2026-05-20

The Expert Strikes Back: Interpreting Mixture-of-Experts Language Models at Expert Level

ICML 2026 顶会论文：深入 Mixture-of-Experts 语言模型的专家级别内部机制，揭示专家如何协同与对抗。

arXiv:2604.02178v2 Announce Type: replace-cross Abstract: Mixture-of-Experts (MoE) architectures have become the dominant choice for scaling Large Lan…

mixture-of 语言模型模型可解释性深度学习 icml 2026

19

📝 深度技术 arXiv 机器学习 2026-05-20

SMART Fine-tuning Factor Augmented Neural Lasso

提出SMART框架，将预训练模型融入高维非参数变量选择，为微调提供理论基础。

arXiv:2604.12288v2 Announce Type: replace-cross Abstract: Fine-tuning is a widely used strategy for adapting pre-trained models to new tasks, yet its …

微调非参数高维统计变量选择 smart框架

20

📝 深度技术 arXiv 机器学习 2026-05-20

Causal Bias Detection in Generative Artificial Intelligence

提出新方法检测生成式AI中的因果偏差，为提升模型公平性提供理论支撑。

arXiv:2605.11365v2 Announce Type: replace-cross Abstract: Automated systems built on artificial intelligence (AI) are increasingly deployed across hig…

因果偏差检测生成式人工智能模型公平性因果推断深度学习

🐂 牛哥精选

Divergence-Suppressing Couplings for Rectified Flow

Uncertainty-Calibrated Recommendations for Low-Active Users

What is Holding Back Latent Visual Reasoning?

Predictable Confabulations: Factual Recall by LLMs Scales with Model Size and Topic Frequency

DashAttention: Differentiable and Adaptive Sparse Hierarchical Attention

Transformers for Learning on Noisy and Task-Level Manifolds: Approximation and Generalization Insights

LLM Agents Are the Antidote to Walled Gardens

CoUn: Empowering Machine Unlearning via Contrastive Learning

The Bayesian Geometry of Transformer Attention

KASER: Knowledge-Aligned Student Error Simulator for Open-Ended Coding Tasks

NanoQuant: Efficient Sub-1-Bit Quantization of Large Language Models

Meltdown: Circuits and Bifurcations in Point-Cloud-Conditioned 3D Diffusion Transformers

Self-Play Only Evolves When Self-Synthetic Pipeline Ensures Learnable Information Gain

SlimQwen: Exploring the Pruning and Distillation in Large MoE Model Pre-training

Unlearning Isn't Deletion: Investigating Reversibility of Machine Unlearning in LLMs

Prompt reinforcing for long-term planning of large language models

Enhancing Cloud Network Resilience via a Robust LLM-Empowered Multi-Agent Reinforcement Learning Framework

The Expert Strikes Back: Interpreting Mixture-of-Experts Language Models at Expert Level

SMART Fine-tuning Factor Augmented Neural Lasso

Causal Bias Detection in Generative Artificial Intelligence

📅 日期