牛哥精选 · 三个月

1

📝 深度技术 arXiv 机器学习 2026-05-23

On Robustness and Chain-of-Thought Consistency of RL-Finetuned VLMs

首份系统研究RL微调VLM的鲁棒性与思维链一致性，揭示模型脆弱性根源

arXiv:2602.12506v3 Announce Type: replace Abstract: Reinforcement learning (RL) finetuning has become a key technique for enhancing large language mod…

rl微调视觉语言模型鲁棒性思维链一致性模型脆弱性

2

🤖 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

3

📝 深度技术 arXiv AI 2026-05-20

ICED: Concept-level Machine Unlearning via Interpretable Concept Decomposition

提出ICED方法，通过可解释概念分解实现视觉语言模型中的概念级遗忘，精准移除目标知识而不影响无关语义。

arXiv:2605.14309v1 Announce Type: cross Abstract: Machine unlearning in Vision-Language Models (VLMs) is typically performed at the image or instance …

机器遗忘概念级遗忘可解释概念分解视觉语言模型计算机视觉

4

📝 深度技术 arXiv 机器学习 2026-05-20

Reasoning Portability: Guiding Continual Learning for MLLMs in the RLVR Era

提出“推理可移植性”新概念，为多模态大模型在强化学习时代的持续学习指明方向。

arXiv:2605.18903v1 Announce Type: new Abstract: Vision-Language Models in Continual Learning (VLM-CL) aim to continuously adapt to new multimodal task…

推理可移植性持续学习多模态大语言模型 rlvr 视觉语言模型

5

🔓 开源项目 arXiv 计算机视觉 2026-05-20

Single-Sample Black-Box Membership Inference Attack against Vision-Language Models via Cross-modal Semantic Alignment

针对视觉语言模型的单样本黑盒成员推理攻击，利用跨模态语义对齐检测模型记忆风险，为数据隐私安全提供新评估方法。

arXiv:2605.17341v1 Announce Type: new Abstract: Vision-Language Models (VLMs) have achieved remarkable success, yet their reliance on massive datasets…

视觉语言模型成员推理攻击跨模态对齐黑盒攻击隐私安全

6

📝 深度技术 arXiv 计算机视觉 2026-05-20

Test-Time Hinting for Black-Box Vision-Language Models

黑盒视觉语言模型无需微调，测试时通过提示提升性能，新方法突破传统限制

arXiv:2605.16410v1 Announce Type: new Abstract: Test-time scaling (TTS) methods have proven highly effective for LLMs, yet their application to vision…

视觉语言模型黑盒优化测试时提示 arxiv论文

7

📝 深度技术 arXiv 机器学习 2026-05-20

Continual Learning for VLMs: A Survey and Taxonomy Beyond Forgetting

梳理视觉-语言模型持续学习最新综述，超越遗忘视角解读多模态大模型演进挑战

arXiv:2508.04227v2 Announce Type: replace-cross Abstract: Vision-language models (VLMs) and the recent surge of Multimodal Large Language Models (MLLM…

持续学习视觉语言模型多模态大语言模型综述遗忘

8

📝 深度技术 arXiv AI 2026-05-20

Your CLIP has 164 dimensions of noise: Exploring the embeddings covariance eigenspectrum of contrastively pretrained vision-language transformers

发现CLIP模型嵌入空间存在164个噪声维度，协方差特征谱揭示视觉语言模型的结构异常

arXiv:2605.14893v1 Announce Type: cross Abstract: Contrastively pre-trained Vision-Language Models (VLMs) serve as powerful feature extractors. Yet, t…

clip 嵌入协方差特征谱视觉语言模型噪声维度

9

📝 深度技术 arXiv AI 2026-05-20

Descriptor: Distance-Annotated Traffic Perception Question Answering (DTPQA)

为自动驾驶场景设计带距离标注的交通感知问答数据集，评估VLM空间推理能力

arXiv:2511.13397v2 Announce Type: replace-cross Abstract: The remarkable progress of Vision-Language Models (VLMs) on a variety of tasks has raised in…

交通感知视觉语言模型自动驾驶问答数据集距离标注

10

📝 深度技术 arXiv 计算机视觉 2026-05-20

VLM-AutoDrive: Post-Training Vision-Language Models for Safety-Critical Autonomous Driving Events

这篇论文提出了模块化后训练框架，将视觉语言模型适配到自动驾驶安全关键事件的异常检测，融合元数据、LLM描述与VQA及CoT推理，提升准确性。

arXiv:2603.18178v2 Announce Type: replace Abstract: The rapid growth of ego-centric dashcam footage presents a major challenge for detecting safety-cr…

视觉语言模型后训练自动驾驶异常检测模块化框架

11

📝 深度技术 arXiv 机器学习 2026-05-19

Neutral-Reference Prompting for Vision-Language Models

提出中性参考提示方法，直击视觉语言模型基类-新类权衡这一核心难题，为迁移学习提供全新解决思路。

arXiv:2605.15615v1 Announce Type: cross Abstract: Efficient transfer learning of vision-language models (VLMs) commonly suffers from a Base-New Trade-…

视觉语言模型提示学习迁移学习基类-新类权衡中性参考提示

12

📝 深度技术 arXiv 计算机视觉 2026-05-19

Unlocking Dense Metric Depth Estimation in VLMs

新型方法让视觉语言模型突破3D密集几何感知瓶颈，实现高效深度估计。

arXiv:2605.15876v1 Announce Type: new Abstract: Vision-Language Models (VLMs) excel at 2D tasks such as grounding and captioning, yet remain limited i…

vlm 深度估计 3d理解密集几何视觉语言模型

13

📝 深度技术 arXiv 计算机视觉 2026-05-19

Deep Pre-Alignment for VLMs

VLMs靠轻量投影器映射视觉特征，但早期层对齐不足浪费深度，论文提出深度预对齐方案解决此缺陷。

arXiv:2605.15300v1 Announce Type: new Abstract: Most Vision Language Models (VLMs) directly map outputs from ViT encoders to the LLM via a lightweight…

视觉语言模型 vlm 深度预对齐对齐挑战 vit

14

📝 深度技术 arXiv AI 2026-05-19

When Language Overwrites Vision: Over-Alignment and Geometric Debiasing in Vision-Language Models

揭示视觉语言模型语言压倒视觉导致幻觉的根源，提出几何去偏方法提升可靠性。

arXiv:2605.08245v3 Announce Type: replace-cross Abstract: Vision-Language Models (VLMs) increasingly power high-stakes applications, from medical imag…

视觉语言模型过度对齐几何去偏幻觉 vlm

15

📝 深度技术 arXiv AI 2026-05-19

On the Cultural Anachronism and Temporal Reasoning in Vision Language Models

视觉语言模型正被广泛应用于文化遗产领域，但一篇新研究揭示了其隐藏的硬伤：文化时代错位。模型难以准确理解历史文物中的时间语境，尤其在非西方视觉文化上表现糟糕。这项基准测试戳中了多模态AI无法跨越的文化盲区。

arXiv:2605.15071v1 Announce Type: cross Abstract: Vision-Language Models (VLMs) are increasingly applied to cultural heritage materials, from digital …

文化错位时间推理视觉语言模型基准测试历史文物

🐂 牛哥精选

On Robustness and Chain-of-Thought Consistency of RL-Finetuned VLMs

What is Holding Back Latent Visual Reasoning?

ICED: Concept-level Machine Unlearning via Interpretable Concept Decomposition

Reasoning Portability: Guiding Continual Learning for MLLMs in the RLVR Era

Single-Sample Black-Box Membership Inference Attack against Vision-Language Models via Cross-modal Semantic Alignment

Test-Time Hinting for Black-Box Vision-Language Models

Continual Learning for VLMs: A Survey and Taxonomy Beyond Forgetting

Your CLIP has 164 dimensions of noise: Exploring the embeddings covariance eigenspectrum of contrastively pretrained vision-language transformers

Descriptor: Distance-Annotated Traffic Perception Question Answering (DTPQA)

VLM-AutoDrive: Post-Training Vision-Language Models for Safety-Critical Autonomous Driving Events

Neutral-Reference Prompting for Vision-Language Models

Unlocking Dense Metric Depth Estimation in VLMs

Deep Pre-Alignment for VLMs

When Language Overwrites Vision: Over-Alignment and Geometric Debiasing in Vision-Language Models

On the Cultural Anachronism and Temporal Reasoning in Vision Language Models

📅 日期