Typed models can misinterpret option names despite type safety
Type-Safe Is Not Error-Free: A Constrained Decision Head Follows the Option Name, Not the Rubric Bound to It
Artificial Intelligence
Summary
Typed decision models promise to always produce outputs that follow a specific format, like choosing from set answers instead of random text. But the authors found that even when the model’s output fits the expected format perfectly, the model might misunderstand what those answers mean. By swapping the labels (like changing '0/1' to 'no/yes'), the model’s decisions reversed drastically, showing it followed the option names' wording more than their intended meanings. This means type safety alone doesn’t guarantee correct understanding or accurate decisions.
What this means in practice
- •For software developers: Check model output consistency beyond schema validation by testing how changes in option names affect decisions to ensure correct interpretation.
- •For machine learning engineers: Design and evaluate typed decision systems by considering semantic effects of option labels to avoid misleading accuracy despite type-safe outputs.
Authors
Yu Sun, Junhao Xu
Abstract
Typed decision models are built for settings where model outputs are consumed directly by software. Instead of generating free-form text, they return a decision over a predefined set of options. By construction, every output conforms to the required schema. Yet this guarantee does not tell us whether the model interprets the options as intended. We study Jev and two Jev-like models with open weights by changing how option names are assigned to rubrics. Each option consists of an option name and a textual rubric that defines what the option means. We change only which option name is assigned to each rubric; the question, state, rubric wording, and set of option names remain exactly the same. On 1200 workflow decisions with task-specific rubrics, renaming the two options from 0/1 to no/yes changes 70.4 more answers per hundred (95% CI: [67.6, 73.1]) and shifts AUC from .94 to .23, revealing a systematic reversal in the decision ranking rather than simple uncertainty. The same operation has little effect with neutral option names. This pattern holds across all 4 predicates, where the effect is at least 7.4x larger than under the neutral control, and becomes stronger as the number of options increases. The effect also depends on the read-out geometry: a second model family that mean-pools over the full option span flips 4.1x less often. The hosted model exhibits the same behavior: the swap changes AUC from .8146 to .5806 and produces 24x as many answer flips as its test-retest floor. In contrast, replacing the option names with random character strings returns all model families to the neutral-control regime without reducing accuracy. The failure therefore depends on the semantic polarity of the option names rather than on the renaming operation itself. Across all conditions, the type-error rate remains 0%, even when decision accuracy degrades substantially.