Beyond Source Selection: Detecting and Resolving Conditional Knowledge Conflicts in Retrieval-Augmented Generation
Febin K Renu
Final-year research, Karunya Institute of Technology and Sciences
Abstract
Conflict-aware retrieval-augmented generation resolves disagreements between passages by selecting the most credible source. For a large and common class of disagreement this is the wrong operation: a general rule and its exception are both true under different scopes, and selection silently deletes the exception together with the condition that governs it. The work treats conditional conflicts as a first-class type, extracts the governing conditions, and composes a scoped answer that preserves every valid branch. It introduces the Suppression Rate metric suite, which makes the deletion measurable where answer-correctness cannot, and a benchmark of naturally occurring rule-and-exception pairs on which to measure it.
- retrieval-augmented generation
- knowledge conflicts
- conditional reasoning
- NLI
- evaluation
BibTeX
@misc{renu2026beyond,
title = {Beyond Source Selection: Detecting and Resolving Conditional Knowledge Conflicts in Retrieval-Augmented Generation},
author = {Febin K Renu},
booktitle = {Final-year research, Karunya Institute of Technology and Sciences},
year = {2026},
url = {https://github.com/febinrenu/ScopeCompose},
}