Thirty training prompts. That is all it took to push an open-weight reasoning model within 0.4 kg-CO2 of the theoretical optimum for building energy storage control, according to a new arXiv preprint [S1]. The fine-tuned model cut simulated emissions from 70.5 to 61.2 kg-CO2, landing just shy of the 60.8 kg-CO2 ceiling that exact dynamic programming proves is the best possible answer [S1]. How a handful of prompts did what years of bespoke control engineering has struggled to scale is the question every building operator should be asking.

Why building energy control is stuck

Model predictive control and reinforcement learning are the two techniques engineers reach for when they want a building to store thermal energy cheaply: charge a heat pump when electricity is clean or cheap, discharge when it is not [S1]. Both work in the lab. Both are hard to scale across buildings, the authors note, because each building needs its own model, its own tuning, its own months of commissioning [S1].

The preprint, posted on arXiv on 15 July and classified under artificial intelligence and machine learning, has not been peer-reviewed [S1]. But the approach it describes is simple enough to grasp: take a reasoning model that can already plan in plain text, and teach it to plan better using rewards computed from a known optimal solution.

The 30-prompt method

To adapt an open-weight reasoning model, the team employed reinforcement learning with verifiable rewards (RLVR) [S1]. The key lies in the reward signal. The team executed exact offline dynamic programming, a top-tier optimisation technique, on an intentionally straightforward office-building thermal energy storage test case with a known, computable optimum [S1]. These DP action values were then translated into dense rewards for each potential action the model could select [S1].

In plain terms: for every decision the model considered, the reward function could tell it exactly how good or bad that decision was relative to the optimum. No guessing, no human preference labels, no vague feedback. Just a number.

Reinforcement fine-tuning used only 30 training prompts [S1]. The model learned to act as an upper-level scheduler that reads text-based descriptions of the building state and weather forecasts, then outputs hourly heat-pump setpoints [S1].

What it means

The result is a model that emits 61.2 kg-CO2 on the benchmark, down from 70.5 kg-CO2 before fine-tuning and within touching distance of the 60.8 kg-CO2 dynamic programming optimum [S1]. That is a 13% emissions cut from a model that was already producing setpoints, pushed closer to perfect by 30 examples.

But the more interesting finding is what the fine-tuning actually changed. An analysis of the model's traces reveals that RFT failed to impart a novel strategy [S1]. Instead, it reinforced existing planning behaviors, such as evaluating alternative actions, anticipating future states, and verifying physical feasibility [S1]. The model previously exhibited these behaviors, albeit irregularly; the verifiable rewards served to solidify these beneficial routines.

Reasoning itself matters. GPT-5, a frontier reasoning model, nearly matched the DP and model predictive control baselines without any task-specific training [S1]. GPT-4o, a non-reasoning model, produced higher emissions than a building with no thermal storage at all, meaning its setpoint choices were worse than not trying [S1]. The gap between those two models points to inference-time reasoning, the step-by-step deliberation a model does before answering, as the ingredient that makes language models viable for this kind of scheduling [S1].

The reinforced planning patterns held up under forecast errors and an unseen thermal energy storage condition [S1]. They also carried over to a battery scheduling task, though the battery's different structure limited the gains [S1].

What it means for business

For a facilities manager or a two-person energy consultancy, the appeal is obvious if the method scales. Today, getting MPC working in a new building means building a custom model of that building's thermal dynamics, then solving an optimisation problem every hour. It works, but the engineering cost per building is high, which is why MPC lives in large commercial towers rather than suburban offices.

The preprint suggests a different path: a reasoning model that reads a text description of the building state and a weather forecast, then writes a setpoint schedule. If the DP-based reward approach generalises, a building operator could fine-tune a model for each site with a small number of prompts rather than a full custom control system. The 30-prompt figure is the number to watch because it implies the training cost per building could be low.

For vendors of building management software, this is a signal that open-weight reasoning models are approaching the point where they can replace hand-coded optimisation for routine scheduling. The open-weight framing matters: the model runs on infrastructure the operator controls, with no per-query API fee and no dependency on a single provider's pricing.

That said, the benchmark is deliberately simple [S1]. A real building has occupancy changes, equipment failures, mixed-use zones, and weather that does not match forecasts. The emissions figures are simulated, not measured from a physical deployment [S1].

What we don't know yet

The preprint does not name the specific open-weight model used [S1], making it impossible to assess cost, hardware requirements, or whether the result depends on a particular model family. The GPT-5 and GPT-4o comparisons rely on the authors' reporting, with no disclosed details about prompt formatting, temperature settings, or how many calls each model made [S1].

The benchmark is a single, deliberately simple office-building scenario [S1]. Whether the 30-prompt approach works on a complex multi-zone building, or across buildings with different heating and cooling systems, is untested. The battery task transfer showed limited gains, suggesting the method does not automatically generalise to every storage problem [S1].

The authors themselves frame this as a starting point. They say the results motivate higher-fidelity tests of whole-building control and the development of scalable verifiers for city-scale energy management [S1]. That is a research agenda, not a product.

The next concrete signal to watch for is whether the authors or others release code and a named model, and whether a follow-up applies the method to a real building with measured energy data rather than simulation. Until then, the 30-prompt result is promising but provisional.

If this kind of reporting helps you make better decisions, subscribe to keep reading.

Sources

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.