Expanding Merlin-Arthur Classifiers

Can we stabilise the Merlin-Arthur setup for complex dataset/text data?

The idea is to expand Merlin-Arthur Classification setup in the following ways:

  1. Larger Datasets
    • Streetview Housenumbers, where we can compare with bounding boxes
    • Medicine Data, where Explainability Methods had been challenged in the past *
  2. Text-based Explanations
    • Have agents explain their rationale via text
    • Train Arthur as LLM to play a simple game (e.g. Connect Four) in some text based fashion
    • Merlin and Morgana are implemented as LLMs reasoners explaining their reasoning for the next move and Arhtur has to decide with which move he goes.