[samuellll]

The work seems to generate per-instance weights that describe the features based on the effect that they have on the outcome. How would you propose to do that with two LLMs?