Question on multiple treatments for causal forest

Incredible package folks!

Currently I'm trying to understand how econml.grf.CausalForest handles multiple treatments, but am struggling to find it in the code. Most of the cited papers by Athey seem to focus on binary treatments only.

Say I have a one hot encoded matrix of 5 different treatments: T, where T is of shape (n_samples x n_treatments). I would have assumed that I would find the same treatment effect for the first treatment regardless of whether I used T[0] as my treatment vector, or whether I used the full matrix T. Using T[0] would encode whether a sample was treated with the first treatment, but if I used the full one hot encoded matrix (n_samples x 5 treatments) the treatment effects for the first treatment is different!

Why is this the case? For multiple treatments, what does the algorithm assign as the control? I cannot find anything in the code or documentation

Thanks in advance

Start with the econml.grf.CausalForest code and documentation, then compare its treatment handling with the cited papers. Clarify how one-hot encoded multiple treatments are interpreted, how treatment effects differ from a single treatment vector, and which treatment serves as control; document the behavior and supporting examples.