What Do Rationales Communicate? A Message-Intervention Study in Role-Specialized QA
Not provided in the abstract
Abstract
This study investigates the impact of rationales in role-specialized QA systems, revealing that faithful rationales add little to answer accuracy while corrupted rationales significantly alter support judgments.
Reality Card
Rationale sharing should be evaluated as a verification-message mechanism rather than solely for improving answer accuracy.
Corrupted rationales shift support judgments by 10-22%, and an explicit rationale-checking prompt amplifies this effect to 34-55%.
The study's findings may be limited by the specific datasets used (MuSiQue, HotpotQA, 2WikiMultiHopQA) and the reliance on a single model (DeepSeek).
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.