Feedback for “Indirect Reasoning with LLMs”

It does not work in my task.

My task is to tell whether a 'hypothesis' is entailed in a 'premise' or not. My hypothesis is a triple [subject, predicate, obeject], while my premise is a long text. My prompt firstly asks the LLM to find the sentence relevant to the hypothesis, then shorten the sentence, finally use NOT hypothesis and NOT premise for reasoning.

But the issue is with the sentence shortening: LLM cannot shorten the sentence correctly: it may combine the hypothesis into it !!

Any ideas to deal with lengthy premises with a lot of irrelevant information?

The issue names the “Indirect Reasoning with LLMs” material but no file, test, or entry point. Start by reviewing that material and reproducing the reported prompt workflow; done would require an agreed documentation change or guidance for handling lengthy premises.

Written by the indexing model from the issue text.