牛哥精选 · 本周

1

📝 深度技术 arXiv 机器学习 2026-05-20

When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search

教你利用激活几何诊断转向瓶颈，GRACE框架让概念工程优化更高效。

arXiv:2605.16362v1 Announce Type: new Abstract: Activation steering offers a lightweight way to control LLMs without retraining, but its effectiveness…

激活几何概念工程 grace 转向优化深度技术

2

📝 深度技术 arXiv 机器学习 2026-05-20

Distinguishable Deletion: Unifying Knowledge Erasure and Refusal for Large Language Model Unlearning

提出可区分删除方法，统一知识擦除与拒绝机制，为LLM去学习难题提供新思路。

arXiv:2605.16776v1 Announce Type: new Abstract: Mitigating sensitive and harmful outputs is fundamental to ensuring safe deployment of LLMs. Existing …

llm去学习知识擦除模型安全可区分删除 icml2026

3

🤖 AI·大模型 arXiv 机器学习 2026-05-20

Confidence Geometry Reveals Trace-Level Correctness in Large Language Model Reasoning

置信度几何方法揭示大语言模型推理中痕迹级正确性的全新框架，附开源代码。

arXiv:2605.16824v1 Announce Type: new Abstract: Large language models (LLMs) generate not only reasoning text, but also token-level confidence traject…

大语言模型推理置信度几何痕迹级正确性开源代码

4

📝 深度技术 arXiv 机器学习 2026-05-20

PhysioSeq2Seq: A Hybrid Physiological Digital Twin and Sequence-to-Sequence LSTM for Long-Horizon Glucose Forecasting in Type 1 Diabetes

用混合数字孪生+LSTM实现1型糖尿病长期血糖预测，精准度突破现有方法。

arXiv:2605.16860v1 Announce Type: new Abstract: Accurate long-horizon glucose forecasting is critical for automated insulin delivery systems, which he…

糖尿病预测数字孪生 lstm 序列到序列医疗ai

5

📝 深度技术 arXiv 机器学习 2026-05-20

Step-wise Rubric Rewards for LLM Reasoning

提出逐步评分奖励机制，优化LLM推理的中间步骤监督，突破传统仅奖励最终答案的局限。

arXiv:2605.17291v1 Announce Type: new Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) is widely used to improve reasoning in large lan…

llm推理强化学习分步奖励 rlvr 研究论文

6

🤖 AI·大模型 arXiv 机器学习 2026-05-20

Post-Trained MoE Can Skip Half Experts via Self-Distillation

最新研究：后训练MoE模型通过自蒸馏跳过一半专家，无需从头预训练，显著降低计算量。

arXiv:2605.18643v1 Announce Type: new Abstract: Mixture-of-Experts (MoE) scales language models efficiently through sparse expert activation, and its …

moe 混合专家模型自蒸馏稀疏激活大模型效率优化

7

📝 深度技术 arXiv 机器学习 2026-05-20

VolTA-3D: Self-Supervised Learning for Brain MRI using 3D Volumetric Token Alignment

脑部MRI自监督学习新框架，利用3D体令牌对齐实现跨数据集通用表征

arXiv:2605.16775v1 Announce Type: cross Abstract: Self-supervised learning (SSL) has advanced medical image analysis be enabling learning form large u…

自监督学习脑部mri 3d医学影像 volta-3d

8

📝 深度技术 arXiv 机器学习 2026-05-20

HyDRA: Hybrid Dynamic Routing Architecture for Heterogeneous LLM Pools

针对异构LLM池的混合动态路由架构，无需因模型目录变更而重训练，解决现有二元路由的耦合问题。

arXiv:2605.17106v1 Announce Type: cross Abstract: Production LLM deployments increasingly maintain heterogeneous model pools spanning order-of-magnitu…

hydra llm 路由架构异构模型池混合动态路由

9

📝 深度技术 arXiv 机器学习 2026-05-20

Why Do Safety Guardrails Degrade Across Languages?

揭秘多语言下AI安全护栏失效的根本原因，从数据分布到tokenization的深度剖析。

arXiv:2605.17173v1 Announce Type: cross Abstract: Large language models exhibit safety degradation in non-English languages. Standard evaluation relie…

安全护栏多语言 ai安全模型退化语言偏差

10

📝 深度技术 arXiv 机器学习 2026-05-20

SURGE: Approximation-free Training Free Particle Filter for Diffusion Surrogate

提出无近似、无需训练的粒子滤波方法用于扩散代理模型，被ICML 2026接收，开创性极强。

arXiv:2605.18745v1 Announce Type: cross Abstract: Diffusion-based generative models increasingly rely on inference-time guidance, adding a drift term …

surge 粒子滤波扩散模型无训练机器学习

11

📝 深度技术 arXiv 机器学习 2026-05-20

Towards Migrating Neural Network Implementations

如何自动将神经网络实现从一种深度学习框架迁移到另一种？这篇论文提出新方法，解决框架间兼容性难题

arXiv:2511.02610v2 Announce Type: replace Abstract: The development of smart systems (i.e., systems enhanced with AI components) has thrived thanks to…

神经网络深度学习框架代码迁移自动化 aiware会议

12

📝 深度技术 arXiv 机器学习 2026-05-20

Beyond Superficial Unlearning: Sharpness-Aware Robust Erasure of Hallucinations in Multimodal LLMs

多模态大模型消除幻觉新方法：锐度感知鲁棒擦除，超越浅层遗忘，提升模型可靠性

arXiv:2601.16527v2 Announce Type: replace Abstract: Multimodal LLMs are powerful but prone to object hallucinations, which describe non-existent entit…

多模态llm 消除幻觉锐度感知鲁棒擦除 ai安全

13

📝 深度技术 arXiv 机器学习 2026-05-20

Concordia: Self-Improving Synthetic Tables for Federated LLMs

联邦学习下用合成表格数据自我优化大模型，Concordia框架有望提升隐私与效率。

arXiv:2605.09855v2 Announce Type: replace Abstract: Federated learning (FL) enables training large language models (LLMs) without sharing raw data, bu…

concordia 联邦学习合成数据大语言模型自我改进

14

📝 深度技术 arXiv 机器学习 2026-05-20

LightTransfer: Your Long-Context LLM is Secretly a Hybrid Model with Effortless Adaptation

长上下文LLM无需重训练即可轻松转为混合模型，LightTransfer方法实现高效适配，论文被TMLR 2025收录。

arXiv:2410.13846v3 Announce Type: replace-cross Abstract: Scaling language models to handle longer contexts introduces substantial memory challenges d…

长上下文大模型混合模型轻量适配 tmlr

15

🤖 AI·大模型 arXiv 机器学习 2026-05-20

Difficulty-Based Preference Data Selection by DPO Implicit Reward Gap

新方法用DPO隐式奖励差距衡量样本难度，自动筛选高质量偏好数据，提升模型训练效率。

arXiv:2508.04149v2 Announce Type: replace-cross Abstract: Aligning large language models (LLMs) with human preferences is a critical challenge in AI r…

偏好数据选择 dpo 隐式奖励数据筛选难度评估

16

📝 深度技术 arXiv 机器学习 2026-05-20

Self-Supervised Bootstrapping of Action-Predictive Embodied Reasoning

RSS 2026 论文提出自监督方法，让机器人通过预测行动自主学习推理，无需大量标注数据。

arXiv:2602.08167v2 Announce Type: replace-cross Abstract: Embodied Chain-of-Thought (CoT) reasoning has significantly enhanced Vision-Language-Action …

自监督学习行动预测具身推理机器人强化学习

17

📝 深度技术 arXiv 机器学习 2026-05-20

Anomaly-Preference Image Generation

ICML 2026前沿成果，提出异常偏好图像生成新范式，为可控生成开辟新方向。

arXiv:2605.02439v2 Announce Type: replace-cross Abstract: Synthesizing realistic and diverse anomalous samples from limited data is vital for robust m…

异常偏好图像生成 icml 2026 大模型偏好学习

18

📝 深度技术 arXiv NLP 2026-05-20

Multilingual jailbreaking of LLMs using low-resource languages

用非洲低资源语言玩多轮对话，成功绕过ChatGPT、Gemini等主流大模型的安全护栏，安全漏洞新发现。

arXiv:2605.18239v1 Announce Type: new Abstract: Large Language Models (LLMs) remain vulnerable to jailbreak attempts that circumvent safety guardrails…

多语言越狱大语言模型低资源语言非洲语言安全机制

19

📝 深度技术 arXiv NLP 2026-05-20

Responsible Agentic AI Requires Explicit Provenance

自主AI可信度关键：显式溯源机制保障责任伦理，前沿论文剖析Agentic AI治理新范式

arXiv:2605.17169v1 Announce Type: cross Abstract: Agentic AI is rapidly proliferating across diverse real-world domains such as software engineering, …

负责任ai 自主智能体数据溯源可信ai 人工智能伦理

20

📝 深度技术 arXiv NLP 2026-05-20

Mixture-of-Experts Can Surpass Dense LLMs Under Strictly Equal Resource

MoE架构在严格等资源条件下首次证明超越稠密大模型，ICLR 2026最新研究。

arXiv:2506.12119v2 Announce Type: replace Abstract: Mixture-of-Experts (MoE) language models dramatically expand model capacity and achieve remarkable…

moe 大模型资源效率 iclr 2026 深度学习

🐂 牛哥精选

When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search

Distinguishable Deletion: Unifying Knowledge Erasure and Refusal for Large Language Model Unlearning

Confidence Geometry Reveals Trace-Level Correctness in Large Language Model Reasoning

PhysioSeq2Seq: A Hybrid Physiological Digital Twin and Sequence-to-Sequence LSTM for Long-Horizon Glucose Forecasting in Type 1 Diabetes

Step-wise Rubric Rewards for LLM Reasoning

Post-Trained MoE Can Skip Half Experts via Self-Distillation

VolTA-3D: Self-Supervised Learning for Brain MRI using 3D Volumetric Token Alignment

HyDRA: Hybrid Dynamic Routing Architecture for Heterogeneous LLM Pools

Why Do Safety Guardrails Degrade Across Languages?

SURGE: Approximation-free Training Free Particle Filter for Diffusion Surrogate

Towards Migrating Neural Network Implementations

Beyond Superficial Unlearning: Sharpness-Aware Robust Erasure of Hallucinations in Multimodal LLMs

Concordia: Self-Improving Synthetic Tables for Federated LLMs

LightTransfer: Your Long-Context LLM is Secretly a Hybrid Model with Effortless Adaptation

Difficulty-Based Preference Data Selection by DPO Implicit Reward Gap

Self-Supervised Bootstrapping of Action-Predictive Embodied Reasoning

Anomaly-Preference Image Generation

Multilingual jailbreaking of LLMs using low-resource languages

Responsible Agentic AI Requires Explicit Provenance

Mixture-of-Experts Can Surpass Dense LLMs Under Strictly Equal Resource

📅 日期