AI资讯日报 - 2026/6/25

👨‍🔬 Shuyi Zhang, Yunfan Lou, Hongyang Cheng, Yichen Guo, Chuyao Fu, Yaoxu Lyu, Xiaojie Zhang, Haoran Li, Pengwei Wang, Zhongyuan Wang, Shanghang Zhang

On-Policy Self-Distillation with Sampled Demonstrations Reduces Output Diversity

学术论文 ArXiv 重要度: 8

发现同策略自蒸馏虽提高pass@1，但降低输出多样性，理论分析其偏差，并指出在分布外场景表现差。

👨‍🔬 Andrei Liviu Nicolicioiu, Mohammad Pezeshki, Aaron Courville

Can Trustless Agents Be Trusted? An Empirical Study of the ERC-8004 Decentralized AI Agent Ecosystem

学术论文 ArXiv 重要度: 8

实证分析ERC-8004协议，发现大多数注册为空壳，声誉系统可被操纵，大量Sybil行为，提出改进建议。

👨‍🔬 Xihan Xiong, Zelin Li, Wei Wei, Qin Wang, William Knottenbelt, Zhipeng Wang

Natural Ungrokking: Asymmetric Control of Which Rules Survive Pretraining

学术论文 ArXiv 重要度: 7

发现预训练中规则的自然遗忘现象，由训练数据中规则获胜频率决定，且规则一旦遗忘无法恢复。

👨‍🔬 Juliana Li, Diya Sreedhar

TriViewBench: Controlled Complexity Scaling for Multi-View Structural Reasoning in MLLMs

学术论文 ArXiv 重要度: 7

构建多视图推理基准，发现所有MLLM性能随复杂度单调下降，链式思维提示几乎无效，瓶颈在于跨视图空间表征。

👨‍🔬 Yu-Yang Chen, Lan-Zhe Guo

Privacy Vulnerabilities of Attention Layers in Tabular Foundation Models and Protection of High-Risk Queries

学术论文 ArXiv 重要度: 7

提出AMIA攻击利用注意力模式泄露成员信息，并设计基于k-匿名的推理时防御，降低泄漏同时保持效用。

👨‍🔬 Tânia Carvalho, Maxime Cordy

A cross-process welding penetration status prediction algorithm based on unsupervised domain adaptation in laser and TIG welding

学术论文 ArXiv 重要度: 5

提出无监督域适应框架，实现激光与TIG焊接穿透状态跨过程预测，准确率大幅提升，降低重标注成本。

👨‍🔬 Sen Li, Haichao Cui, Chendong Shao, Yaqi Wang, Xinhua Tang

A welding penetration prediction model for laser welding process based on self-supervised learning using physics-informed neural networks

学术论文 ArXiv 重要度: 5

提出SimPhysNet，利用物理信息对比学习实现仅200张标签图片即达96.06%激光焊接穿透预测精度。

👨‍🔬 Sen Li, Xiaoying Liu, Xiaojian Xu, Chendong Shao, Yaqi Wang, Ling Lan, Xinhua Tang, Haichao Cui

🤖 AI资讯日报

📊 今日趋势总结

MIT Non-AI License

Why Boring Businesses Outlast AI Hype Cycles

Ask HN: Anyone concerned about NYC Local Law 144?

Ask HN: What's the pain using current AI algorithms?

Ask HN: Is the rate of progress in AI exponential?

Ask HN: What would you read to learn about "artificial intelligence"?

The AI Crackpot Index

NLP, AI, ML, bots – a passing trend or much more? What's your take on this?

Show HN: Startup Raising capital through Book Sales

Common Lisp + Machine Learning Internship at Google (Mountain View, CA)

The Next Bill Gates or Albert Einstein in AI “Chris Clark” – Yourobot

Bioinformatician

The Unfireable Safety Kernel: Execution-Time AI Alignment for AI Agents and Other Escapable AI Systems

Learning Action Priors for Cross-embodiment Robot Manipulation

Neglected Free Lunch from Post-training: Progress Advantage for LLM Agents

Model Forensics: Investigating Whether Concerning Behavior Reflects Misalignment

FORCE: Efficient VLA Reinforcement Fine-Tuning via Value-Calibrated Warm-up and Self-Distillation

On-Policy Self-Distillation with Sampled Demonstrations Reduces Output Diversity

Can Trustless Agents Be Trusted? An Empirical Study of the ERC-8004 Decentralized AI Agent Ecosystem

Natural Ungrokking: Asymmetric Control of Which Rules Survive Pretraining

TriViewBench: Controlled Complexity Scaling for Multi-View Structural Reasoning in MLLMs

Privacy Vulnerabilities of Attention Layers in Tabular Foundation Models and Protection of High-Risk Queries

A cross-process welding penetration status prediction algorithm based on unsupervised domain adaptation in laser and TIG welding

A welding penetration prediction model for laser welding process based on self-supervised learning using physics-informed neural networks

📅 历史日报目录