Decision models—non-generative scorers that return a probability over caller-supplied options and emit no text—are increasingly used as routers, classifiers, and guardrails. Because they generate nothing, their behavior is shaped entirely by the inputs they score, including the language of the test prompt.

A new paper on arXiv investigates this sensitivity by focusing on Bengali idioms. The authors study how changing the test language influences whether these models lean toward literal or idiomatic readings of the phrases. The title's phrase "literal lock-in" suggests that certain language conditions make the models more prone to selecting literal meanings.

The work highlights a practical concern for deploying decision models in multilingual settings: even when the model itself is language-agnostic in output, the test language can bias its scoring. The abstract does not provide full results, but the framing points to a need for careful prompt design when using such models for idiom-heavy or culturally specific content.