Rationales Offer Minimal Accuracy Gains in Specialized QA Pipelines

A message-intervention study using DeepSeek across MuSiQue and HotpotQA shows that faithful rationales add almost no answer accuracy in reasoner-to-verifier pipelines. While corrupted rationales significantly shift support judgments by 10–55%, they only marginally affect final answer changes.