Type-Safe Is Not Error-Free: A Constrained Decision... | AI Research

Type-Safe Is Not Error-Free: A Constrained Decision Head Follows the Option Name, Not the Rubric Bound to It
This paper investigates a critical vulnerability in "typed" decision models—AI systems designed to output structured, schema-compliant choices (like "yes" or "no") for software applications. While these models are guaranteed to return an output that fits a predefined format, this structural guarantee does not ensure the model is actually interpreting the task correctly. The authors demonstrate that these models often prioritize the "name" of an option (e.g., "yes") over the actual definition or "rubric" provided to explain what that option means. This leads to situations where a model remains perfectly "type-safe" while being fundamentally wrong in its decision-making. The ai search story also surfaces in Stanford AI discovery identifies natural weight..., adding another angle.

Testing the Model's Logic

To test how these models make decisions, the researchers conducted a "label-rubric swap." They kept the question, the state, and the definition (rubric) of each option identical, but swapped which name was attached to which definition. For example, if a model was asked to choose between "no" and "yes," the researchers would swap the definitions so that the name "no" was now tied to the definition of "yes," and vice versa. If a model were truly reading the definitions, its answers should remain consistent. Instead, the researchers found that the models consistently followed the option names, effectively ignoring the instructions provided in the rubrics.

The Impact of Semantic Polarity

The study reveals that the failure is driven by the "semantic polarity" of the names—the inherent meaning or bias associated with words like "yes," "no," "true," or "false." When researchers used neutral labels like "0/1" or "A/B," the models performed consistently regardless of the swap. However, when using polar labels, the models' performance plummeted. In some cases, the models did not just become uncertain; they exhibited "confident inversion," where they consistently chose the wrong answer because the name triggered a biased response that overrode the actual task requirements. The ai search story also surfaces in Google AI Releases TimesFM 3 for..., adding another angle.

Architectural Influence

The researchers tested three different model families and found that the "read-out geometry"—the specific way a model calculates its final score—influences how much it relies on names. Models that calculate scores by looking at a single "marker" token were highly susceptible to name-based bias. In contrast, models that "mean-pool" (averaging the score across the entire text of the name and the rubric) were more resistant to this effect, as the rubric text helped dilute the influence of the name. Regardless of the architecture, the "type-error rate" remained 0%, confirming that the models were always producing valid schema outputs, even when those outputs were logically incorrect.

Key Takeaways

The findings suggest that developers cannot rely on type-safety guarantees to ensure the accuracy of AI-driven decisions. The study shows that replacing meaningful names with random character strings eliminates the bias without hurting performance, proving that the models are exploiting the semantic weight of labels rather than the content of the instructions. Ultimately, the paper highlights that the way a model is built to "read" its options is a design choice that significantly impacts whether it follows the user's intent or its own internal biases. The same large language models question is explored in ActMap, which adds a research perspective. as detailed in the full paper on Arxiv