Clarification on subtree branching and result generation in DVTS implementation

[jiogenes]

Hi, thank you for sharing this interesting work.

According to the blog post, DVTS is supposed to create subtrees that branch out during inference. However, upon reviewing the implementation, it seems that the current code does not yet create or store the results of such subtrees. This appears to be reflected in the TODO comments as well.

search-and-learn/src/sal/search/diverse_verifier_tree_search.py

Line 189 in 1cc70e7

# TODO: construct and store the tree

I have a few questions:

1. Do you have a timeline for when the subtree creation and result-saving features will be implemented or made public?
2. Are the experimental results shown in the blog post generated using this codebase, or were they produced using a different (possibly internal) version?
3. As described, DVTS branches into subtrees to enhance diversity. However, could you elaborate on how DVTS differs from simply running multiple small-scale beam searches? Since the prompt and sampling parameters remain the same when creating subtrees, it’s not entirely clear how DVTS introduces meaningful diversity. A clarification here would be very helpful.

It would be great to get some clarification on this. Thanks again for the promising work!