What Do Rationales Communicate? A Message-Intervention Study in Role-Specialized QA Pipelines

Forum topic · 小凯 · 2026-10-05

Summary

This post summarizes an NLP paper (arXiv:2610.00018) by Jiameng Zhang and Hongqiu Wu that examines what rationale messages actually contribute in role-specialized question-answering pipelines, where a reasoner passes rationales to a verifier. The authors introduce a message-intervention diagnostic: evidence and candidate answers are fixed while only the rationale crossing the reasoner-to-verifier boundary is varied. Testing 400 samples from MuSiQue, HotpotQA, and 2WikiMultiHopQA with DeepSeek as both generator and verifier, they find faithful rationales add almost no answer accuracy over no rationale, yet corrupted rationales strongly alter support judgments. Under a blind verifier prompt, harmless paraphrases shift support by only 0–2.5% while corrupted rationales shift it by 10–22%; an explicit rationale-checking prompt amplifies this to 34–55%. Final answer changes are smaller (2–30%), and only 2.9–35.3% of corrupted support flips coincide with answer changes. Human auditing shows 16 of 42 effective corruptions were over-trust of contaminated rationales, and blind human reviewers rejected 9/10 corrupted rationales that the model accepted. The paper argues rationale sharing should be evaluated as a verification-message mechanism, not merely a path to higher answer accuracy.

论文概要

研究领域: NLP 作者: Jiameng Zhang, Hongqiu Wu 发布时间: 2026-10-05 arXiv: 2610.00018

Overview

Key Findings

Setup: 400 examples from MuSiQue, HotpotQA, and 2WikiMultiHopQA, with DeepSeek as both generator and verifier.
Faithful rationales add almost no answer accuracy over no rationale at all, while corrupted rationales strongly alter support judgments.
Under a blind verifier prompt, harmless paraphrases shift support by only 0–2.5%, whereas corrupted rationales shift support by 10–22%.
An explicit rationale-checking prompt amplifies the same pattern to 34–55%.
Final answer movement is smaller (2–30%), and only 2.9–35.3% of corrupted support flips coincide with answer changes.

Human Audit

Of 42 effective corruptions, 16 were cases of over-trust in contaminated rationales.
Blind human reviewers rejected 9 out of 10 corrupted rationales that the model accepted, or flagged them as unclear.

Boundary Analysis

Cross-model and cross-task boundary checks reveal when the rationale channel is active, amplified, inert, or folded into task labels.

Conclusion

Rationale sharing should be evaluated as a verification-message mechanism, not merely as a path toward higher answer accuracy.

Original Abstract (excerpt)

Role-specialized QA pipelines increasingly pass rationales from a reasoner to a verifier, but it is unclear what this message actually buys: better answers, stronger support assessment, or a new failure surface. We introduce a message-intervention diagnostic that fixes the evidence and candidate answer while varying only the rationale passed across the reasoner-to-verifier boundary. On 400 MuSiQue, HotpotQA, and 2WikiMultiHopQA examples with DeepSeek as generator and verifier, faithful rationales add almost no answer accuracy over no rationale, while corrupted rationales strongly alter support judgments...

Paper link: arXiv:2610.00018

--- *Auto-collected on 2026-10-05*