An agent can decide whether to sample another answer, ask a verifier, or stop. If the budget is measured in tokens, those actions have one apparent exchange rate. If the budget is measured in energy, the hardware and serving regime enter the problem.
My energy-budget study asks a narrow deployment question: can a frozen controller regain a useful budget trade-off by recalibrating its action costs after a deployment change? The language model stays fixed. The intervention is in the controller’s cost model.
A price change the policy did not see
A model call’s cost depends on more than its text. The tested GPU and concurrency regimes changed measured joules per call substantially. A controller calibrated in a cheap regime can overspend in an expensive one, while an expensive-regime calibration can leave useful computation unused on a cheaper machine.
The study measured a frozen Qwen2.5-1.5B-Instruct workload across A10, H100, and L4 regimes, using long-window NVML energy measurements. Held-out replay then evaluated frozen protocols. The paper distinguishes attributed energy from net-above-idle energy and includes live checks; these are different accounting choices, not interchangeable physical quantities.
The result belongs to this workload and measurement protocol. It does not show that a small calibration universally predicts the energy of an arbitrary agent, model, or tool.
The first reserve guarded the wrong uncertainty
The original primary test was null. Within a regime, per-call energy was almost deterministic in output tokens. A reserve based on residual error in that relationship missed the dominant uncertainty: how long the next attempt would be.
That matters because a highly accurate price per token does not tell the controller how many tokens it is about to buy. A cost model can be well calibrated conditional on length and still produce poor budget decisions before length is known.
A token-quantile predictive reserve addressed this different uncertainty. In the confirmatory measured study, small target-regime calibration reduced violations for shifts to more expensive serving and sometimes improved accuracy when the target was cheaper. The paper reports the size of those effects and the regimes where a comparison missed its margin.
The weak self-verifier was another relevant component. Removing it improved energy use in several tested targets. Spending on verification only helps if that verifier supplies enough useful information to justify its own cost.
The deployment is part of the algorithm
This result changes how I would describe an “energy-aware” policy. Its behavior consists of the decision rule, the predictive reserve, the cost calibration, and the serving assumptions. Moving only the model checkpoint does not preserve that system.
An evaluation should therefore cross both workload and regime. Hold out questions, measure new containers, change concurrency, and report budget violations alongside task quality. A conservative policy can have an excellent violation rate simply by doing very little. An aggressive policy can have high accuracy by ignoring the budget.
It also matters whether replay matches live execution. A replayed sequence reuses measured calls under assumptions about their attribution and scheduling. A live run checks whether those assumptions remain adequate when the controller actually makes the decisions.
Where this meets physical AI
I am interested in the same pattern in a broader setting: a controller transfers, but its actions acquire different costs and consequences. A robot body changes the cost of a motion. A facility changes the value of a sensor reading. A serving stack changes the energy price of another attempt.
The analogy is useful only if it produces a test. For this study, the test was small calibration under a controlled regime change, with the policy frozen. For the broader systems, the corresponding question is which part can be recalibrated without relearning the whole controller, and how to detect that a small calibration is no longer enough.