Energy per Successful Goal: Goal-Level Energy Accounting for Agentic AI Systems
Independent Researcher(独立研究者);Texas A\&M University(德克萨斯A&M大学);Texas A\&M University Department of Computer Science(德克萨斯A&M大学计算机科学系)
Deepak Panigrahy, Aakash Tyagi, Stephen Abkin, David Kebo Houngninou, Rabi Mahapatra
备注:
v2: 37 pages. Adds ARM platform (NVIDIA Grace GB10) with GPU (DCGM) and network wait energy, nine task families, bootstrap confidence intervals, stochastic retry model, and new coauthors
当前的AI能量基准在单次模型调用或训练运行的粒度上测量能耗。对于经典的单轮工作负载,这种单位仍然一致。但对于智能体系统——其中单个用户目标可能触发多步编排、工具调用、重试和故障恢复循环——调用次数是实现产物而非任务属性,推理级归一化错误地表示了目标完成的能量成本。我们提出A-LEMS(智能体LLM能量测量系统),一个跨层测量框架,将AI能量核算单位从每次推理能量重新定义为每个成功目标能量(EpG)。EpG聚合所有执行尝试(包括失败和重试)的总工作流能量,归一化到成功完成的目标。A-LEMS通过时间边界模型、将RAPL信号映射到工作流级能量的五层观测管道,以及将每次测量绑定到硬件和运行时配置的可复现协议,形式化了能量归因。基于EpG,我们定义了编排开销指数(OOI),在相同任务标准下隔离编排相对于线性执行的能耗成本。在五个推理和三个工具增强任务族中,智能体工作流每个成功目标的平均能耗是线性基线的4.33倍(888.1 J vs 205.3 J)。这种开销由编排结构驱动,而非推理计算。对于工具增强任务,OOI反转至低于1.0倍:智能体执行比线性更便宜,确认该指标捕捉了编排结构而非固定的向上偏差。这些发现表明,每次推理能量对于智能体AI是不充分的。EpG和OOI为准确基准测试提供了测量基础,其中编排结构是能耗的主要决定因素。
英文摘要:
Current AI energy benchmarks measure consumption per model invocation. For agentic systems, where a single user goal may trigger multistep orchestration, tool calls, retries, and failure recovery, the invocation count is an implementation artifact rather than a task property, and per inference normalization misrepresents the energy cost of completing a goal. We present A-LEMS (Agentic LLM Energy Measurement System), a cross layer framework that redefines the unit of AI energy accounting as Energy per Successful Goal (EpG): total workflow energy across all attempts, including failures and retries, divided by successfully completed goals. A-LEMS combines a temporal boundary model, a five layer attribution pipeline mapping hardware signals (RAPL on x86; SPBM and DCGM on ARM) to workflow energy, and a reproducibility protocol binding every measurement to hardware, environment, and runtime configuration. Building on EpG, we define the Orchestration Overhead Index (OOI), the ratio of agentic to linear EpG on matched goals under identical success criteria. Across nine task families on Intel x86 and NVIDIA Grace GB10 ARM, CPU only OOI ranges 0.35x to 11.07x. Including GPU dynamic energy, which accounts for 83.8% of agentic EpG in local inference, widens this to 0.4x to 26.48x, so CPU only measurement understates overhead about twofold for reasoning tasks. For remote inference, network wait energy shifts OOI from 2.5x to 2.35x. The overhead arises from orchestration structure (planning, retries, recovery) that induces additional inference calls, not from costlier individual calls. The metric is not biased upward: single tool calculator dispatch inverts to OOI = 0.61x (95% CI [0.35, 0.97]) on x86. Goal level accounting via EpG and OOI provides a measurement foundation for benchmarking agentic AI.
T2SPO: Trajectory-to-Step Policy Optimization for Agentic Reinforcement Learning
State Key Laboratory of Novel Software Technology, Nanjing University(南京大学计算机软件新技术国家重点实验室);School of Intelligence Science and Technology, Nanjing University(南京大学智能科学与技术学院);ByteDance(字节跳动)
Agent-UCT: Upper Confidence Bounds Applied to Trees for Cost-Aware Agentic Workflow Optimization
The University of Hong Kong(香港大学);School of Artificial Intelligence, Jiangxi Science and Technology Normal University(江西科技师范大学人工智能学院);The Hong Kong University of Science and Technology(香港科技大学);University of Science and Technology of China(中国科学技术大学);New York University(纽约大学)
Establishing Best Practices for Building Rigorous Agentic Benchmarks
UIUC(伊利诺伊大学);Stanford University(斯坦福大学);University of California, Berkeley(加州大学伯克利分校);Yale University(耶鲁大学);Princeton University(普林斯顿大学);MIT(麻省理工学院);Transluce;ML Commons;Amazon(亚马逊);UK AI Safety Institute(英国人工智能安全研究所);University of Oxford(牛津大学)
A Dual-Helix Governance Approach Towards Reliable Agentic Artificial Intelligence for WebGIS Development
Geographic Information Systems Center, Florida International University(佛罗里达国际大学地理信息系统中心);Geospatial Analytics, Technology and Open Research Lab, University of Florida(佛罗里达大学地理空间分析、技术与开放研究实验室)
MATRA: Modeling the Attack Surface of Agentic AI Systems -- OpenClaw Case Study
DistriNet, KU Leuven(DistriNet,根特大学);IBM Research(IBM研究院);CISPA Helmholtz Center for Information Security(CISPA海德堡信息安全中心);Imperial College London(伦敦帝国理工学院);A*STAR Institute for Infocomm Research(A*STAR信息与通信研究机构)
Agentic AI Software Engineers: Programming with Trust
National University of Singapore(新加坡国立大学);Carnegie Mellon University, KBR Inc., NASA Ames(卡内基梅隆大学、KBR公司、NASA阿姆斯特朗实验室);CISPA Helmholtz Center for Information Security, University of Stuttgart(CISPA海姆霍兹信息安全中心、斯图加特大学);Columbia University(哥伦比亚大学)
EDGE: Experience-Distillation for Guided Exploration in Agentic Reinforcement Learning
School of Artificial Intelligence, University of Chinese Academy of Sciences(中国科学院大学人工智能学院);Institute of Automation, Chinese Academy of Sciences(中国科学院自动化研究所);School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences(中国科学院大学先进交叉科学学院);Wuhan AI Research(武汉人工智能研究院)
System of Agentic AI for the Discovery of Metal-Organic Frameworks
University of California, Berkeley(加州大学伯克利分校);Lawrence Berkeley National Laboratory(伯克利国家实验室);Google DeepMind(谷歌DeepMind);Kavli Energy NanoScience Institute(Kavli能源纳米科学研究所)
Beyond Shortest Path: Agentic Vehicular Routing with Semantic Context
Institute of Computing (IC)(计算学院);Hub de Inteligência Artificial e Arquiteturas Cognitivas (H.IAAC)(人工智能与认知架构中心);Universidade Estadual de Campinas (UNICAMP)(坎皮纳斯州立大学)
Agentic Learner with Grow-and-Refine Multimodal Semantic Memory
Nanjing University of Science and Technology(南京理工大学);Baidu Inc(百度公司);AIML, Adelaide University(AIML,阿德莱德大学);Singapore University of Technology and Design(新加坡科技设计大学)
VASP Agent: An Agentic Framework for Autonomous First-principles Calculations
Shanghai Artificial Intelligence Laboratory(上海人工智能实验室);Department of Computer Science and Technology(计算机科学与技术系);School of Physical Science and Technology(物理科学与技术学院);Department of Automation(自动化系);Beijing National Research Center for Information Science and Technology (BNRist)(北京信息科学与技术国家研究中心);School of Chemistry and Chemical Engineering(化学与化工学院);Key Laboratory of Computational Physical Sciences (Ministry of Education)(计算物理科学重点实验室);The Chinese University of Hong Kong(香港中文大学);Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences(中国科学院深圳先进技术研究所)
IneqAgent: Agentic Inequality Proving in Lean with Symbolic Computation
School of Software Engineering, East China Normal University(华东师范大学软件工程学院);School of Computer and Information Science, Southwest University(西南大学计算机与信息科学学院)
TheoremForge: Scaling up Formal Data Synthesis with Low-Budget Agentic Workflow
Gaoling School of Artificial Intelligence, Renmin University of China(中国人民大学全球化人工智能学院);Beijing Key Laboratory of Research on Large Models(北京大模型研究关键实验室);Engineering Research Center of Next-Generation Intelligent Search(下一代智能搜索与推荐工程研究中心)