2026-08-12 科技日报

扫描 591 篇候选内容 · 覆盖 650 个信息源 · 纳入 576 篇

📌 今日要点

• FACT: Failure-Aware Causal Training for World-Action Models
• 突发，OpenAI 八年元老离职！
• OpenAI 搬空 Claude Code！用户家当一键进 Codex，唯独带不走 Claude
• Claude 和 GPT 思维链竟被两步蒸馏！116 页论文曝光 API 致命漏洞
• 代码榜刷满分，AI 科研却撞墙？140 道真实研究题撕开「自进化」真相
• 全球首款！Claude 引发的数据恐慌，被这家中国黑马补上了

📝 科技简讯

• FACT: Failure-Aware Causal Training for World-Action Models — ArXiv RO (cs.RO)
• 突发，OpenAI 八年元老离职！ — 新智元
• OpenAI 搬空 Claude Code！用户家当一键进 Codex，唯独带不走 Claude — 新智元
• Claude 和 GPT 思维链竟被两步蒸馏！116 页论文曝光 API 致命漏洞 — 新智元
• 代码榜刷满分，AI 科研却撞墙？140 道真实研究题撕开「自进化」真相 — 新智元
• 全球首款！Claude 引发的数据恐慌，被这家中国黑马补上了 — 新智元
• LEGO: Leveled Language Gaussian Splatting — ArXiv CV (cs.CV)
• Signpost Watermarking: Joint Optimization for Visual Watermark Coexistence — ArXiv CV (cs.CV)
• 4D-WAM: 4D Consistent World Modeling for Autonomous Driving — ArXiv CV (cs.CV)
• P3CA: Encoder-Agnostic Interpretation of Vision Foundation Model Embeddings via Spatial Probing — ArXiv CV (cs.CV)
• MAD-HOI: Masked Autoregressive Diffusion for Generating Articulated Hand Object Interactions from Text — ArXiv CV (cs.CV)
• Closed-Loop LLM Co-Pilots for Digital Agriculture — arXiv AI (cs.AI)
• SPOTting the Future: Lookahead Explanations for Deep Reinforcement Learning — arXiv AI (cs.AI)
• MIDAS: Mutual Information Disentanglement with Uncertainty-Aware Fusion for Incomplete Multimodal Sentiment Analysis — arXiv AI (cs.AI)
• Towards Sustainable Artificial Intelligence: A Comprehensive Review and Comparative Analysis of Deep Learning Models’ Carbon Footprint — arXiv AI (cs.AI)
• ReCBM: Uncertainty-Gated Relational Reasoning for Concept Bottleneck Models — arXiv AI (cs.AI)
• What Is Dreamforce? Everything About Salesforce’s Annual Event — Salesforce AI Blog
• The 5 Systems that Power the Headless 360 Architecture — Salesforce AI Blog
• How to Make AI a Trusted Business Partner With Anti-Hallucination Practices — Salesforce AI Blog
• 《置身谷内》！Jeff Dean 上顶会自曝离职现场：被 1500 人围堵 — 量子位
• Anthropic CEO 整天神神叨叨，投资人受不了了 — 量子位
• 国产具身智能创全球新纪录！以 30% 成本跑赢 Figure AI 45% 效率，聪明的具身大脑成关键 — 量子位
• 紫东太初推出 GMC 核心集剪枝方法，少 80%Token 仍满血保真多模态能力 — 量子位
• 2026 中国科创投资夏季峰会暨陕西科创产业生态大会圆满落幕 — 量子位
• 人形之外，擎羽把“身体”变成具身智能的新变量 — 量子位
• Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis — Hugging Face Blog
• LFM2.5-VL-3B for Better and Faster Vision Capabilities for the Edge — Hugging Face Blog
• Building an End-to-End Data Science Portfolio Project — KDnuggets
• 5 Easy Ways to Install Python on Windows — KDnuggets
• Putting sign language AI into users’ hands — Google DeepMind
• MindTopo reveals VLMs’ spatial reasoning abilities — Microsoft Research
• Transformer Geometry Observatory TGO-IV: Developmental Topology Observatory — ArXiv ML (cs.LG)
• Uncertainty-Aware Ensemble Deep Randomized Neural Networks for Classification — ArXiv ML (cs.LG)
• CurveFP: Rational-Radix Logarithmic Datatypes with Closed Products for Language Models — ArXiv ML (cs.LG)
• Sheaf-Based Federated Representation Learning — ArXiv ML (cs.LG)
• DOCSCHISEL: Adaptive Tool Documentation Optimization Framework for LLM Agents — ArXiv ML (cs.LG)
• FlowScout: From Execution Feedback to Reliable Tool-Using Agent Workflows — ArXiv ML (cs.LG)
• UserToolBench: A User-Profile-Hidden Benchmark for Personalized Decision Making in Tool-Use LLMs — ArXiv ML (cs.LG)
• Part 2: Amazon Bedrock cost attribution with Amazon Athena and CUDOS — AWS Machine Learning
• How OneAdvanced deployed over 50 AI agents on UK-sovereign AWS — AWS Machine Learning

📊 数据概览