Two problems sit on top of each other here, and the paper is unusually clear about both.
The first is that the people who need information out of a building model rarely know how to get it. An architect, a construction manager or a facilities manager wants to know how many fire doors are on level three. Getting that answer requires a query language or an API neither of them writes.
The second is worse. The same property is stored differently depending on which tool authored the model, which convention the team followed, and what language they worked in. Door width might be Width, or Rough Width, or NominalWidth, or Breite (B). There is no canonical place to look.
Existing systems translate a natural language question into one fixed query, which works until the model is organised differently than assumed. The same team's earlier agent had twenty-nine hand-coded tools, and 45% of its errors traced back to the limits of those tools.
The idea: stop assuming, start looking
Adaptive exploration means a CodeAct-style agent that writes and runs arbitrary Python against the live model, looks at what came back, and adjusts. It discovers how the model is structured rather than assuming it.
The loop is simple. The agent receives the question, the model path, its execution history and optionally some tool descriptions. It emits either a code action, meaning reasoning plus Python, or a final answer. The code runs in a sandbox with no network access, the output is appended to history, and an iteration cap forces the agent to abstain rather than loop forever.
The agent is not told where door width lives. It goes and finds out.
How it was tested
This is the most rigorous evaluation design in any of the five papers I have read this month. A three by four factorial ablation, run on the 514-question held-out half of a new benchmark the team released alongside the work.
That benchmark, ifc-bench v2, is 1,027 tasks across 37 IFC models from 21 projects. Crucially it spans Revit 2011 through 2025, ArchiCAD 11 through 25, Synchro and DDS-CAD, several IFC schema versions and multiple languages. It is built to be heterogeneous on purpose, which is the whole point.
Adaptive beat static by 36.8 to 38.5 percentage points across every augmentation configuration, at p below 0.001. The strong adaptive model reached 55 to 57% strict accuracy while abstaining on under 7% of questions. The static approach abstained on roughly half.
The result I would put in front of anyone choosing an AI vendor: the weaker adaptive model beat the stronger static one. Architecture beat capability.
The augmentation finding is the useful one
They tested two ways of helping the agent. Hybrid documentation retrieval, combining dense and BM25 search with reverse-question search, RRF fusion and a cross-encoder rerank. And domain tools, either hand-written or auto-generated by a seven-agent pipeline that identifies patterns, creates tools, debugs them and prunes the ones that earn their keep.
Read that last bar again. Giving the weaker model hand-written tools made it 10.2 points worse. The tools trapped it in unproductive loops, because a tool is an instruction about how to think, and a weak model follows it off a cliff.
For the strong model, augmentation moved the needle by under a point in either direction. All that tool engineering was noise.
Two things worth carrying into your own work
First, the authors expect the finding to generalise beyond Python. Whether the harness is an MCP tool server, SQL, Cypher or a CLI, the claim is that iterative execution with feedback is what matters, not the interface. If they are right, that is a design principle rather than a result.
Second, and this is the line I would underline: models that had been checked and corrected in Solibri scored noticeably higher. Model quality raised the ceiling on what any agent could extract.
That is a commercial observation dressed as a footnote. If cleaning a model measurably improves what AI can do with it, then getting models into a defined state is not preparatory work you do before the interesting project. It is the project.
The number that keeps this honest
55 to 57% strict accuracy. On a benchmark. Under lab conditions.
That is nowhere near enough for unsupervised use on anything safety-critical, and the authors say so. It is enough for a first pass that a person checks, which is the same conclusion every serious paper in this area reaches. It also covers IFC only, and tested one model family.