Tag
Researchers find that 'typed decision models' (like Jev) follow option labels rather than option definitions, and trace the failure to how prompts are rendered rather than to constrained decision heads, introducing PolicyBench to demonstrate this 'option-label bias'.
This paper studies open-weight typed decision models (laya, von) built on Qwen2.5 and shows that their outputs are driven mostly by option labels rather than written definitions, an 'option-label bias' traced to prompt rendering rather than the constrained decision head. It introduces PolicyBench and a two-call test practitioners can use to detect the failure mode.
This paper introduces the Triple-Stream Stress (TSS) diagnostic framework to uncover lexical interference and label bias in mental health NLP, proposing the Degree of Divergence metric for label-source auditing.