https://xuquant.com/categories/foundation-models/Recent content in Foundation Models on Xu'Bloghttps://xuquant.com/og-default.pnghttps://xuquant.com/og-default.pngHugo -- 0.152.2zhThu, 28 May 2026 22:30:00 +0800https://xuquant.com/posts/foundation-models/qwen-vla/Thu, 28 May 2026 22:30:00 +0800https://xuquant.com/posts/foundation-models/qwen-vla/Qwen Team 2026-05-28 放出的 Qwen-VLA (arXiv:2605.30280) 把 Qwen3.5-4B 多模态骨干和 1.15B 单流 DiT 流匹配动作专家拼成统一具身策略，最有意思的不是数字而是 T2A——冻住 VLM、屏蔽图像，只用文本和 embodiment prompt 把动作先验学出来，再分别灌图像、专门化、RL。本文照 paper 走一遍架构、四阶段 recipe、五维 T2A 消融、流匹配 PPO 的 log-prob 技巧、DOMINO 零样本 26.6% 这个数字背后的含义，以及几条保留的质疑。https://xuquant.com/posts/foundation-models/vla-geometric-fusion-three-architectures/Thu, 28 May 2026 20:00:00 +0800https://xuquant.com/posts/foundation-models/vla-geometric-fusion-three-architectures/NVIDIA + MIT + UT Austin 团队（arXiv:2605.24642）把 GR00T-N1.5 (manipulation VLA) 跟 VGGT (geometric foundation model) 拼起来，做了 Early Fusion / Late Fusion / Spatial Forcing 三种几何注入架构的 controlled 对照。主结果是一个负结果——standard finetune 下三种几何 VLA 都不显著超过 GR00T baseline。但 ablation 链里的几条判断（don't unfreeze LLM、probe 改进不等于 task 改进、mid-training 比架构选择影响更大、gate 近零起步）跟 production AD VLA 的工程决策直接相关。https://xuquant.com/posts/foundation-models/hif-vla-codec-motion-temporal-memory/Wed, 27 May 2026 22:00:00 +0800https://xuquant.com/posts/foundation-models/hif-vla-codec-motion-temporal-memory/CVPR 2026 的 HiF-VLA (arXiv:2512.09928) 在 OpenVLA 基础上加了一组从 MPEG-4 编码副产物里抠出来的 motion vectors，前向预测未来 motion，反向用历史 motion 通过 AdaLN 调制动作流。本文照着 paper 和 motion_layers/ 代码走一遍，覆盖表征选择、Hindsight Encoder 的真实代码维度、Joint Expert 的 AdaLN 调制、Table 3 延迟分解，以及几个 paper 没讲透的点。https://xuquant.com/posts/foundation-models/atlas-one-word-visual-reasoning/Thu, 21 May 2026 20:00:00 +0800https://xuquant.com/posts/foundation-models/atlas-one-word-visual-reasoning/解读 ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both：把画辅助线、框选区域、箭头指示、文本标注等中间视觉操作压缩成可训练的 functional tokens。https://xuquant.com/posts/foundation-models/codepercept-perception-bottleneck/Sat, 02 May 2026 10:00:00 +0800https://xuquant.com/posts/foundation-models/codepercept-perception-bottleneck/深度解读 CVPR 2026 论文 CodePercept：通过系统性缩放实验论证感知（而非推理）才是 STEM 视觉推理的真正瓶颈，提出以可执行代码为感知媒介的双通道范式，8B 模型超越 72B 基线 6.2%，32B 模型在 STEM2Code-Eval 上超越 GPT5-Thinking。https://xuquant.com/posts/foundation-models/kaiming-he-cvpr2026-five-papers-flow-matching-breakthrough/Sat, 18 Apr 2026 18:00:00 +0800https://xuquant.com/posts/foundation-models/kaiming-he-cvpr2026-five-papers-flow-matching-breakthrough/以 iMF（Improved Mean Flow，arXiv:2512.02012）为主线深读何恺明 2026 CVPR 工作，并把它放回 ResNet / MoCo / MAE / SiT 十年脉络中，抓四条贯穿性的方法论 DNA：朴素到极致、改变问题假设、强先验少假设、方法与任务解耦。强链 mathematics/diffusion 系列。https://xuquant.com/posts/foundation-models/deepseek-thinking-with-visual-primitives/Sat, 04 Apr 2026 20:00:00 +0800https://xuquant.com/posts/foundation-models/deepseek-thinking-with-visual-primitives/解读 DeepSeek 联合北大/清华提出的「以视觉原语思考」技术报告：将坐标和边界框作为思维链原语穿插在 CoT 中，尝试用结构化的空间符号缓解推理过程中的指代漂移。本文整理其方法机制并对其「modality 即 ontology」的本体论提案做批判性审视。https://xuquant.com/posts/foundation-models/sceneverse-plus-data-engine-for-3d-scene-understanding/Sat, 21 Mar 2026 18:00:00 +0800https://xuquant.com/posts/foundation-models/sceneverse-plus-data-engine-for-3d-scene-understanding/Deep analysis of CVPR 2026 SceneVerse++: how to build the largest-scale real-world 3D scene dataset from unlabeled internet videos, covering detection, segmentation, spatial VQA, and vision-language navigation.https://xuquant.com/posts/foundation-models/qwen3-vs-qwen3-5-architecture/Sat, 07 Mar 2026 14:00:00 +0800https://xuquant.com/posts/foundation-models/qwen3-vs-qwen3-5-architecture/深入对比 Qwen3.5 与 Qwen3 的架构差异：混合注意力机制、联合多模态训练策略、高稀疏 MoE、部分 RoPE 在注意力、视觉与 MoE 三个维度的演进https://xuquant.com/posts/foundation-models/coral-autonomous-multi-agent-evolution/Sat, 22 Nov 2025 10:00:00 +0800https://xuquant.com/posts/foundation-models/coral-autonomous-multi-agent-evolution/将进化搜索的关键决策委托给自主Agent而非固定启发式规则，如何在数学优化和系统优化任务上实现更快的收敛和更强的结果。https://xuquant.com/posts/foundation-models/reconvla-gaze-crop-implicit-grounding/Mon, 27 Oct 2025 22:00:00 +0800https://xuquant.com/posts/foundation-models/reconvla-gaze-crop-implicit-grounding/OpenHelix 的 ReconVLA (arXiv:2508.10333) 在 OpenVLA 风格的 backbone 后挂一个 3 层 DiT，用 gaze-crop 的 VAE-latent 重建当辅助监督，把 VLA 的注意力锚到目标物体上。本文对照 paper 与开源 code 读一遍，包含 paper 没强调的工程细节，以及几个 paper 没回答的问题——recon-on/off ablation 缺位，'隐式接地' 在训练 supervision 上其实依赖 offline YOLO bbox。https://xuquant.com/posts/foundation-models/inspatio-world-4d-simulator/Sat, 25 Oct 2025 10:00:00 +0800https://xuquant.com/posts/foundation-models/inspatio-world-4d-simulator/InSpatio-World 深度技术分析：一个 13 亿参数的实时 4D 世界模拟器，通过隐式时空缓存与显式几何约束的结合，实现从单目视频以 24 FPS 进行新视角合成。https://xuquant.com/posts/foundation-models/deepseek_series1_mla/Sat, 13 Sep 2025 10:00:00 +0800https://xuquant.com/posts/foundation-models/deepseek_series1_mla/从 DeepSeek V2/V3 的实际部署视角分析 MLA：KV cache 压缩比、推理 throughput、与 GQA/MQA 的工程对比、长 context 下的真实收益。MLA 的数学推导见配套文章。