Testing model reasoning via fact insertion in candidate profiles
September 25, 2026
A new evaluation method tests whether LLM rejections are based on actual missing facts by inserting those facts into rival profiles. In tests on 2WikiMultihopQA, supplying a named missing fact increased the model's choice by an odds ratio of 3.57.
HOW THIS AFFECTS YOU
●
researcherYou can use this insertion method to objectively verify if a model's reasoning is grounded in text or hallucinated.