Evaluating AI reasoning often depends on human-written reasoning chains as reference answers. Yet, as a new arXiv paper points out, the logical support for each step in those chains is rarely verified. If the references themselves contain gaps, any evaluation built on them may be unreliable.
The paper, 'Verifying Human-Written Reasoning References,' situates this problem relative to prior neuro-symbolic approaches to reasoning formalization. The abstract notes that such approaches have formalized some aspects of this area, but the text is cut off before the contribution is fully described.
The key takeaway is that verification of reference reasoning is an overlooked but critical step in AI evaluation. The paper's focus on this gap suggests a direction for making benchmarks more trustworthy.