‘Vibe Building’: when designing a structure that can withstand earthquakes and wind, it is a physics engine that makes the call, not the LLM

In PE-Loop, the language model suggests revisions and features a deterministic calculation engine. Relying on an LLM for this assessment results in 58 per cent of non-compliant buildings being overlooked.

Agents who automatically design buildings often produce models that are plausible to the eye, but rarely verified against seismic and wind-resistant standards. A team of researchers formalizes this problem under the name "Vibe Building" and proposes PE-Loop (Physics-Engine-in-the-Loop), an agent in which a deterministic physics engine is the only source of evaluation.

A strict division of labour

The language model merely proposes discrete changes: choice of cross-sections in a catalogue, topology, bracing system. Regulatory constraints are translated into a reward by the engine, and each design is then validated by seismic and wind time calculations kept away from the loop, as well as by a constructability check. On VB-Bench, a test bench of 3,577 buildings falling under six families of standards, PE-Loop obtains the best verified success rate with three of the four LLMs tested and the best seismic passage rate with all four.

The most telling experiment consists of replacing the verdict of the engine with an LLM judge, everything else being unchanged: 58.43% of the accepted designs are then found to be non-compliant. For the authors, reliability in structural engineering depends less on the power of the model that proposes than on an arbitrator that it cannot influence.