Role prompts do not limit ai reasoning to assigned skill levels
When a Kindergartener Solves Calculus: Measuring Capability Leakage in Role-Prompted Reasoning Models
Computation and LanguageArtificial IntelligenceHuman-Computer Interaction
Summary
When AI models are told to pretend to be learners at different school levels, they still often show knowledge better than that level. The authors created a test to measure this mismatch between the role the AI is given and the skills it displays. They found that models write in the right style but still solve problems at a higher level than expected. They propose a new method called Injection that helps the AI better match its ability to the role it’s asked to play, without losing accuracy.
What this means in practice
- •For conversational ai developers: Control chatbot responses to maintain consistency between style and skill level of the assigned persona, preventing unintended expertise reveal.
- •For educational software teams: Ensure AI tutors simulate appropriate student knowledge when prompted as learners at different grades to create believable interactive learning experiences.
Authors
Pakhapoom Sarapat, Saksorn Ruangtanusak, Kunat Pipatanakul, Pittawat Taveekitworachai
Abstract
We investigate the problem of role-capability leakage (RCL), in which a role-prompted reasoning model generates convincing in-role text while continuing to exhibit capabilities on benchmarks that exceed those implied by the assigned role. For example, when a model is prompted to assume the role of a kindergarten student, one might expect its performance on a mathematics benchmark to reflect kindergarten-level ability rather than expert-level proficiency in solving calculus problems. We introduce RoleCapBench, a curriculum-grounded benchmark for evaluating RCL across six educational roles and four assessment levels spanning elementary school through A-level, and use it to evaluate three open-weight reasoning models. We find that although the models can generate stylistically convincing in-role responses, they consistently fail to align their underlying capabilities with their assigned roles. Naive role prompting yields strong role-voice scores of 1.218--1.389 while retaining above-role accuracy of 0.811--0.898. RCL persists across a range of prompting conditions, including prompts that explicitly instruct the model to match the role's capability level. To mitigate this problem, we propose Injection, an inference-time intervention that combines explicit, role-specific capability guidelines with a guiding prefilled response prefix. Injection improves role-capability alignment across models, reducing above-role accuracy by up to 0.562 while preserving in-role accuracy with a marginal drop of less than 0.058 across most models. All artifacts, including scripts and evaluation data, will be released upon acceptance.