GPEvac combines a graph neural network with a reinforcement-learning method called PPO. The preprint says it uses an edge-first message-passing scheme and a learnable virtual global node to handle both local and long-distance dependencies in a building. A permutation-invariant scoring mechanism is meant to let one learned policy work across different building layouts and sizes.

The authors report that in simulation, GPEvac outperformed intelligent baselines across distinct architectural layouts and reduced total threat exposure. They also state the system computes global evacuation routes in 14.73 milliseconds on local CPU hardware, which they say could support integration with live surveillance systems. The evidence is a preprint abstract; it does not report a real building trial, peer review, or independent replication.