credibility-augmented RAG

[ekmpa]

Objs:

• Figuring out how to best integrate credibility scores in RAG:
  • In choosing documents, or after retrieval?
  • Prompt format?
  • How much to weigh the scores?
• Assess credibility-augmented RAG:
  • Both the labels dataset, and our Credi-inferred scores
  • Improvements on performance? Assessing different set-ups on FACTS, maybe also look at:
    • DeepSearchQA
    • BrowseComp
    • LiveNewsBench
  • In situations of conflicting evidence in particular?
  • Open v. closed sources?

[ekmpa]

Also: FEVER

[ekmpa]

Look cross-domain.