arXivDaily arXiv每日学术速递 周一至周五更新

大厂专区

2026-07-28 至 2026-07-28 共收录 8
2607.15845 2026-07-28 cs.AI 版本更新

Knowledge-Centric Agents for Workflow Generation in ComfyUI

用于工作流生成的以知识为中心的智能体

Zhendong Li, Lei Sun, Ruibo Ming, He Zhang, Danda Pani Paudel, Luc Van Gool, Jinjin Gu

机构 * INSAIT(未知机构) Sofia University “St. Kliment Ohridski”(索非亚大学“圣克莱门特·奥赫里德斯基”分校) Adobe Research(Adobe研究院)

AI总结 研究视觉创作系统中工作流生成问题,提出以知识为中心的框架,通过知识反转、注入和可逆推理进行工作流生成,实验证明该方法生成的工作流在多样性、结构连贯性和执行成功率上优于现有系统。

Comments Accepted to ECCV 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2606.24621 2026-07-28 math.CT cs.AI math.ST stat.TH 版本更新

Infinitesimal Causality

无穷小因果性

Sridhar Mahadevan

机构 * Adobe Research and University of Massachusetts, Amherst(Adobe研究院和马萨诸塞大学阿姆赫斯特分校)

AI总结 本文在Frobenius Markov范畴中引入无穷小因果性的范畴化描述,通过切丛语义刻画干预作为复制/丢弃结构的切形变,并定义因果充分性为两种Frobenius结构的兼容性。

Comments Revised version with substantial corrections and expanded exposition. We clarify the categorical and tangent semantics of $\mathsf{Stat}_\infty$, and recast the three IC rules as tangent-level compatibility tests. We also add corrected examples, counterexamples delimiting the causal interpretation of Lie-bracket residuals, and new diagrammatic explanations

详情

展开后加载摘要…

URL PDF HTML 收藏
2606.19610 2026-07-28 cs.LG cs.AI 版本更新

Latent Confounded Causal Discovery via Lie Bracket Geometry

基于李括号几何的潜在混杂因果发现

Sridhar Mahadevan

机构 * Adobe Research(Adobe研究院) University of Massachusetts, Amherst(马萨诸塞大学阿默斯特分校)

AI总结 利用信息几何和范畴论,提出两种算法(BRIDGE和SKFM),通过干预诱导流的李括号非闭合性检测潜在混杂,大幅缩减因果图搜索空间。

Comments Revised to remove unsupported Kan-extension interpretations and clarify scope. Lie brackets are now nonspecific closure diagnostics, not certificates of latent confounding. Theorems are narrowed to conditional screen retention, residual-footprint rank, and order-dependent acyclicity. BRIDGE/SKFM algorithms and experiments remain intact

详情

展开后加载摘要…

URL PDF HTML 收藏
2606.00754 2026-07-28 stat.ME cs.AI cs.LG 版本更新

Causal Density Functions

因果密度函数

Sridhar Mahadevan

机构 * Adobe Research(Adobe研究院) University of Massachusetts(马萨诸塞大学) Amherst(阿默斯特)

AI总结 提出因果密度函数作为干预分布与观测分布的Radon-Nikodym导数,用于局部密度比衡量因果效应,并给出估计与检验方法。

Comments Substantially revised and narrowed the paper: removed the erroneous categorical theorems; corrected the intervention, domination, convergence, and density-ratio claims; repositioned the estimator as a baseline; corrected reported experimental values and citations; and documented the pairwise graph scorer as a negative result. A detailed errata/change log appears in the supplement

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.30694 2026-07-28 cs.LG 版本更新

Universal Decision Learners

通用决策学习器

Sridhar Mahadevan

机构 * Adobe Research(Adobe研究院) University of Massachusetts(马萨诸塞大学) Amherst(阿默斯特)

AI总结 本文提出通用决策学习器(UDL)的范畴论框架,通过左Kan扩展和右Kan扩展将局部决策行为规范地扩展到全局一致行为,统一了规划、强化学习、因果干预、在线学习和博弈均衡等多种决策形式。

Comments Revised the central construction to use type-correct staged left/right Kan extensions; corrected the semantic quotient theorem; added an explicit discounted-MDP/Bellman derivation; and expanded the non-RL decision examples and limitations. A detailed change log appears in the supplement

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.27259 2026-07-28 cs.LG 版本更新

Kan Extension Transformers: A Categorical Unification of Attention, Diffusion, and Predict-Detach Self-Conditioning

Kan扩展变换器:注意力、扩散和预测-分离自条件的范畴统一

Sridhar Mahadevan

机构 * Adobe Research(Adobe研究院) University of Massachusetts(马萨诸塞大学) Amherst(阿默斯特)

AI总结 提出Kan扩展变换器(KETs)作为多种Transformer实现的统一范畴框架,将Transformer层视为加权结构化扩展算子,并通过预测-分离机制实现有效的自条件化,实验表明预测-分离机制比改变邻域族带来更大性能提升。

Comments Clarified that predict-detach is strictly autoregressive only under a target-relative prefix-measurability condition; restricted exact Kan-extension claims to representable enriched cases; and clarified experimental coverage across depths 2, 8, 16 and widths 64, 256. A detailed change log appears in the supplement

详情

展开后加载摘要…

URL PDF HTML 收藏
2510.23942 2026-07-28 cs.AI 版本更新

Decentralized Causal Discovery using Judo Calculus

使用柔道演算的去中心化因果发现

Sridhar Mahadevan

机构 * Adobe Research and University of Massachusetts, Amherst(Adobe研究院和马萨诸塞大学)

AI总结 该研究提出用柔道演算进行因果发现的直觉主义去中心化框架,其将上下文依赖性形式化为局部真值,描述了算法和实现框架,结合多种因果发现方法,通过实验展示了层理论因果发现的计算效率及相对于经典方法的性能提升。

Comments Substantially revised and shortened version. We refocus the paper on decentralized multi-environment causal discovery and its experiments, and relegate the theory to the companion revised j-do-calculus paper. A detailed revision log is included in the paper

详情

展开后加载摘要…

URL PDF HTML 收藏
2510.17944 2026-07-28 cs.LO cs.AI 版本更新

Intuitionistic $j$-Do-Calculus in Topos Causal Models

拓扑斯因果模型中的直觉主义j-Do演算

Sridhar Mahadevan

机构 * Adobe Research and University of Massachusetts(Adobe研究院和马萨诸塞大学)

AI总结 该论文将Pearl的do-演算推广到拓扑斯因果模型的直觉主义设置,引入j-do-演算,定义相关稳定性,给出推理规则并证明合理性,后续配套论文将阐述从数据估计实体及其实例化方法与实验结果。

Comments Substantially revised. The \(j\)-do-calculus results are now derived from chartwise Pearl do-calculus and sheaf descent. We remove the unsupported general Kan-extension duality, make all hypotheses explicit, correct categorical and probabilistic formulations, and add a detailed revision log

详情

展开后加载摘要…

URL PDF HTML 收藏