公司1 次关联
NVIDIA
Implied by 'Jetson Nano', a product developed and marketed by NVIDIA.
正在加载行业信息...
正在加载行业信息...
公司、人物、产品、技术和开源项目的持续记录。
Implied by 'Jetson Nano', a product developed and marketed by NVIDIA.
Manufacturer of the 'Orange Pi 5 Pro' single-board computer used in evaluation.
Implied by 'Pixel 8' and 'Pixel 8 Pro' devices, which are manufactured and branded by Google.
A model family evaluated in the study for cross-device latency profiling.
A runtime-aware latency prediction framework for LLM deployment on heterogeneous edge devices, using gated fusion of static descriptors and dynamic hardware telemetry, and separating inference into prefill/decode phases.
A deployed physical kiosk system in scenic areas used for public-space gesture interaction, serving as the empirical basis for the failure analysis.
A proposed software abstraction layer that bridges the hand-landmark model and interaction tasks, enabling stable event confirmation by addressing temporal, geometric, and lifecycle failures.
A computer vision model that detects and tracks hand landmarks in video frames, serving as the foundational perception component in the gesture interaction pipeline.
A gesture interaction system designed for public spaces such as scenic kiosks, exhibition halls, and service terminals, focusing on event-level confirmation rather than frame-level recognition.
The alignment of predictive behavior between unimodal (text-only, vision-only) and fused multimodal reasoning; the core detection signal leveraged by FlowGuard.
Internal representations derived from Partial Information Decomposition within FlowGuard, used to quantify cross-modal redundancy, synergy, and dominance.
An open-access repository where the FlowGuard paper (arXiv:2607.21600v1) is published.
AI systems that process and reason across multiple modalities (e.g., text and vision); the target domain secured by FlowGuard.
A theoretical framework used to quantify cross-modal redundancy, synergy, and modality-specific dominance; serves as the foundation for FlowGuard's FlowVectors.
A lightweight inference-time framework for detecting harmful inputs in multimodal large language models by monitoring internal cross-modal consistency using FlowVectors derived from Partial Information Decomposition.
A hybrid multi-agent system combining predictive AI with LLM-based reasoning, evaluated on wildfire risk; inherits and reveals structural limitations of upstream signals rather than correcting them.
A GRU-based predictive model evaluated in the study; achieves strong local monotonicity but fails to produce well-distributed risk levels.
A novel evaluation framework proposed in the paper to assess whether increases in predicted risk scores consistently correspond to increases in observed operational load (e.g., number of fires, intervention time, deployed resources).
An expert-based wildfire risk index used in the French Alpes-Maritimes department, exhibiting the most balanced monotonic behavior across the full risk scale despite poor classification metrics.
The organization hosting the FlowEvo GitHub repository.
A grade-school math word problem benchmark for evaluating quantitative reasoning in language models.
A code generation benchmark consisting of programming problems with function signatures and test cases.
An interactive text-based environment benchmark for evaluating agent task-solving capabilities.
A training-free framework for self-evolving LLM agents that co-evolves workflows and executable skills via workflow-to-skill compilation, skill-to-workflow feedback, and skill curation.
一种无需训练的智能体框架，通过工作流到技能编译、技能到工作流反馈和技能筛选三个机制，使大语言模型智能体在不更新模型参数的情况下持续积累和复用任务求解能力。
本文提出的单调评估框架的开源实现，托管于GitHub。
基于门控循环单元（GRU）的野火风险预测模型，在局部具有强单调性，但风险等级分布不均。
一种用于评估野火风险信号的新型单调性评估框架，衡量预测风险分数的增加是否与观测到的运营负荷（如火灾数量、响应时间、部署资源）的增加一致。
FlowGuard中提出的向量表示，受部分信息分解启发，用于刻画多模态融合预测是否与单模态语义证据保持对齐。
该论文在arXiv平台发布的预印本，编号2607.21601v1，主题为公共空间手势交互的故障分析与运行时抽象
以用户意图是否稳定转化为可执行交互事件为判定标准的新型评估维度
将手势交互建模为单帧手部关键点识别任务的传统评估范式
部署于景区自助终端、展览馆和服务终端等开放环境中的基于手势识别的交互系统
论文提出的介于手势关键点模型与交互任务之间的运行时抽象机制，用于弥合帧级识别与事件级交互之间的鸿沟
一种免训练的探针引导 KV 残差修正方法，针对智能体记忆系统设计，通过估算每个记忆单元共享的偏移量来校正所有复用 token，在仅刷新 10-30% 缓存的情况下实现接近全量重计算的性能，预填充速度提升 2-3.5 倍，且与 KV 缓存量化正交兼容。
用于评估文本到图像生成模型组合能力的基准测试集。
一种通过测试时奖励对齐实现组合文生图的免训练框架，使用模型内在奖励处理联合分布与单概念分布之间的重叠模式，结合KL约束目标函数和两种引导策略的混合方法，提升组合对齐并保持图像质量。
一类基于图结构进行消息传递的神经网络模型，依赖局部邻域聚合实现信息传播，其长距离传播性能受图拓扑结构显著制约。
一种从图的归一化拉普拉斯矩阵计算单个标量的方法，无需模型训练或标注数据，可在数秒内预测图神经网络在给定消息传递深度下的长距离准确率，用于评估图拓扑对GNN长程传播能力的限制。
HierFlow所基于的核心方法论，即在智能体工作流合成中耦合拓扑结构搜索与执行过程搜索的分层范式。
一种无需训练、测试时分层搜索架构，通过拓扑与执行交织的搜索范式自动化智能体工作流设计，结合反馈引导的拓扑调整、MCTS启发的子工作流优化及智能门控模块以提升效率。
SCOPE与SCION相关代码、证据链接输出、日志、运行清单及基准包的开源仓库。
SCION的紧凑开源模型变体，减少对专有大语言模型的依赖。
SCION的轻量变体，在SCOPE核心套件上取得最高F1分数，优于Text2Onto-style、LLM-only及extract-then-aggregate等基线方法。
Schema Construction and Induction with Ontology Normalization，一种可审计的参考管道（非新抽取架构），通过训练文本构建候选空间，并在严格JSON契约下约束命名、合并、过滤、验证和保守融合操作。
Schema Construction and Ontology-induction Pipeline Evaluation，一个仅基于训练文本的基准，用于从原始文本进行模式归纳与可选的模式融合评估，构建自24个公开信息抽取源，核心目标为事件类型及事件内论元角色的抽取。
AgentKVShift 技术的首次公开论文，发布于 arXiv 预印本平台，标识符为 2607.21604v1。
题为《Coupled Hierarchical Search over Topology and Execution for Agentic Workflow Synthesis》的预印本论文，首次公开提出HierFlow技术。