A-LEMS: Redefining AI Energy Metrics for Agentic Systems
A new paper on arXiv (2605.22883) introduces A-LEMS (Agentic LLM Energy Measurement System), a framework that shifts AI energy accounting from per-inference to Energy per Successful Goal (EpG). Current benchmarks measure consumption per model invocation, which fails for agentic systems where a single goal triggers multi-step orchestration, tool calls, retries, and failure-recovery cycles. A-LEMS aggregates total workflow energy across all attempts, normalized by successfully completed goals, using a temporal boundary model and a five-layer observation pipeline mapping RAPL signals. The system aims to provide a task-property-based metric rather than an implementation artifact.
Key facts
- arXiv paper 2605.22883 introduces A-LEMS
- A-LEMS stands for Agentic LLM Energy Measurement System
- New metric: Energy per Successful Goal (EpG)
- Current benchmarks measure per invocation, not per goal
- Agentic systems involve multi-step orchestration, tool calls, retries, failure-recovery
- EpG aggregates energy across all execution attempts
- Normalized by successfully completed goals
- Uses temporal boundary model and five-layer observation pipeline
Entities
Institutions
- arXiv