Appier finds 28 LLMs guess when no answer exists

Appier tested 28 leading LLMs and found accuracy fell 30% to 50% when “none of the above” was correct. Its research release says Direct Preference Optimization improved recognition of those no-answer cases by nearly 30 percentage points, while a companion paper found some models’ reasoning language differed from their response language in more than 90% of cases. Builders should add abstention and language-routing tests to agent evaluations before trusting answer accuracy; AISI’s evaluation work shows why a passing sandbox is not enough.