Tag
该论文提出一种免训练的持续学习方法，通过GEPA式提示进化在冻结的LLM预测器（Llama-3.3-70B）的提示中演化约900词的引导文本，用于多模态时间序列预测，并在严格防数据污染的基准上超越TabPFN-TS-3等时间序列基础模型（GM MASE 0.776 vs 0.898）。
This arXiv monograph defines physical AI and presents its technical landscape, covering multimodal perception, world models, embodied foundation models, motor control, multi-agent coordination, and hardware-software co-design, while distinguishing it from embodied AI and cyber-physical systems and outlining challenges toward Physical AGI.
该论文提出记忆中心持续学习（MCCL）框架，主张 agent 记忆应以是否改善未来行为来评判，而非仅看检索效果，并给出三轴分类法、双循环生命周期、评估协议与记忆扩展框架，以实现自主 agent 的持续自我改进。
初创公司 Electric Plant 融资 1120 万美元种子轮，推出 AI 大语言模型 Bombadil，可通过廉价传感器解读植物茎干中的微弱电信号，实时获知植物对水、养分和光照的需求。
文章比较了 DNA 语言模型（如 Evo 2）与 AlphaGenome 在预测 DNA 变异影响上的差异：前者基于序列似然评分，后者基于实验测量预测基因表达与剪接等读出，且仅限研究用途。
MemT-1.0 introduces a memory-native architecture that embeds continuous memory directly into foundation model computation, achieving ~1.86x inference speedup and 66x storage reduction at 128K history versus full-context KV cache, while improving agentic task success. The release includes the MemT-1.0-Nano and MemT-1.0-Flash model family.
A commenter analyzes the release of DeepSeek-V4.1-Flash-1017 as a positive for the AI industry's profitability: improved token cost-efficiency will drive up AI penetration and demand for cloud providers' infrastructure. Unlike the earlier DeepSeek shock, market conditions have since changed.
Wired/Quanta 报道指出，AI 基因组基础模型（如 Evo 2、Genos、AlphaGenome）仍无法解读人类基因组复杂的三维物理结构及其基因调控机制，说明 AI 目前尚不足以真正破解我们的基因组。
前 NVIDIA 机器人平台产品营销负责人、现 Techable VC 创始合伙人 Amulya Vishwanath 在 TechCrunch Equity 播客中表示，机器人仍处于「乐高时代」，尚未迎来自己的 ChatGPT 时刻，基础模块和标准仍在重建中。
首尔大学团队提出KBS（Korean Benchmark Suite），包含Ko-ARC、Ko-GSM8K等六个基准共9,396个条目，采用非纯翻译方式构建韩语评测集，并评估了韩国自研模型、多语言开源模型及闭源API模型，指出评测结果高度依赖任务构建、评分规则与提示方式。
QD-Rec 提出一种质量-多样性优化方法，结合上下文老虎机与基础模型，通过档案机制演化生成多样化、有理由支撑的高质量推荐，适用于并购、客户定向等高风险决策场景。
YerevaNN researchers show that for image self-distillation, sample repetition effects are dataset-dependent and can violate standard scaling law assumptions — a 1.5% random subset of Maxar even outperforms the full 50M-image pool under equal compute — and propose a modified Chinchilla-style law accounting for unique samples and repetition.
PathFM 是一个基于流匹配的轨迹式时间序列预测模型，能够快速生成完整未来路径并支持多路径采样，同时提出 fps-bench 评测基准以衡量路径级依赖结构的建模能力。
TimeRSI 提出一个围绕冻结的时间序列基础模型的自改进智能体部署系统，通过误差引导的代码迭代修订与自适应选择提升零样本预测性能，在 TIME 和 fev-bench 的 198 项任务上超越所有提交结果。
This edition of Quantized AI News covers three developments: OpenAI launches “Dots,” a persistent background agent (requires a Pro/Business Premium/Enterprise license); Google makes a strong return to the frontier with Gemini 4 Argon (benchmark performance approaching GPT-6 levels); and Pi, a minimalist, model-agnostic coding agent, releases version 1.0 (with native MCP support and just four tools).
Physical Intelligence announced the beta release of the Pi API, allowing teams to adapt its robot foundation models to diverse robots and applications, which observers see as a major step toward cross-embodiment robotics.
该论文提供了关于「智能体机器人」（agentic robots）的统一综述，将现有研究组织为模型、数据、环境和评测（Harness）四大功能领域的17个类别，并探讨了从机器人策略到智能体系统的演进及未来开放挑战。
Industrial Next, a startup founded by ex-Tesla engineers, unveiled AI-driven robotic stations that let factories retool production lines in minutes instead of weeks, and announced a partnership with BMW Group Technology Office at TechCrunch Disrupt 2026.
Physical Intelligence 发布 Pi API Beta，让开发者通过托管 API 上传数据、微调 π0/π0.5 机器人基础模型并执行云端推理，已在多个真实工业场景中验证模型的广泛适用性。
AI labs are increasingly buying code bases and workspace data (emails, chats, documents) from failing startups via firms like SimpleClosure and Sunset, as public and synthetic training datasets hit their limits. No clear benchmarks yet prove real-world archives outperform synthetic data, but deals are becoming routine, with valuations driven by speculative 'research theses' about data utility.