Studies how uncertain or ambiguous rationales affect reasoning performance, and proposes a simple way for models to choose among reasoning paths when rationale quality is inconsistent.
Overview
Natural language rationales are often treated as clean supervision for reasoning. In practice, those explanations can be incomplete, uncertain, or ambiguous—and models may still be asked to trust them.
This paper studies how ambiguous rationales affect reasoning performance and proposes a simple way for models to choose among reasoning paths when rationale quality is inconsistent.
Why it matters
If explanation quality varies, training and evaluation that ignore that noise can overstate reasoning ability. Understanding and handling rationale ambiguity is a step toward more reliable language-based reasoning systems.
Paper
The full paper is embedded below. You can also download the PDF or view it on arXiv.