What Do Rationales Communicate? A Message-Intervention Study in Role-Specialized QA
Read the original on arXiv AI →The study investigates the role of rationales in role‑specialized question‑answering pipelines by fixing evidence and candidate answers while varying only the rationale sent from a reasoner to a verifier. Experiments on 400 examples from MuSiQue, HotpotQA, and 2WikiMultiHopQA show that faithful rationales add almost no answer accuracy, whereas corrupted rationales significantly alter support judgments, especially under explicit rationale‑checking prompts. Human audits reveal that many corrupted rationales lead to mistrust or rejection, highlighting the importance of evaluating rationale sharing as a verification mechanism rather than merely a means to improve accuracy.
Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.