How is skill learning in SISL connected to and make use of the various modules such as InverseMode, ContrastiveModel, TL_model, RL_model, AL_model trained in SISA?

[cjmdd]

Hello, I have reviewed the source code and understand the architecture of the SISA module, including the encoder-decoder structure, complete graph construction, kNN graph filtering, coding tree initialization, hierarchical abstraction via structural entropy minimization, directed transition graph GdirG_{dir}Gdir​ construction, structural entropy computation on directed graphs, transition probability calculation, clustering loss for the abstract MDP, and other loss computations.

However, I have the following questions:

These modules only appear during the training phase — how are they reflected or utilized in the decision-making (inference) phase?
How is skill learning in SISL connected to and make use of the various modules such as InverseMode, ContrastiveModel, TL_model, RL_model, AL_model trained in SISA?