Exploration Proof Of Concept

[rough note]

design setting where a subset of inputs/tasks artificially do not have successful strategies in router support. CVAE disentangles mixture of strategies and router intervention can force exploration of successful strategies on this input subset.

depending on how we wanna play things, and what to include in an initial paper submission, this kind of mini experiment can serve as a proof or concept in a closing discussion section (as opposed to one of the main experiments in the body of the paper; helps situate it wrt the paper's positioning for reviewing. don't want criticism of "this is only a proof of concept". the main body of the paper sets up the key problem (disentangling modes in language models; disentangled factorization) and proposes a methodological solution and evaluates it. intro hints at applications. discussion presents proof of concept for applications. paper is left self-contained with respect to the problem being formalized.