arXiv:2609.03402 Encodes Six Dimensions into Prompts for 96 LLM Learner Profiles
Prompt-only personalization creates 96 learner profiles for general-purpose AI teaching assistants using six fixed dimensions and Bloom's Taxonomy. Evaluation with NLP metrics and five human raters shows detectable response differentiation but lacks outcome data. The method enables immediate, training-free adaptation across disciplines.
The study implements the framework inside the Jill Watson RAG-based assistant. Structured prompt prefixes encode the six dimensions and a per-query Bloom's level. Experiments run NLP similarity metrics plus a five-participant human rating study that records statistically detectable shifts in response style when profile parameters change.
Data show measurable differences in verbosity, abstraction depth, and perceptual framing across conditions. Human raters distinguish profiles at rates above chance, yet sample size remains five and no longitudinal retention or learning-gain metrics appear. The approach avoids fine-tuning yet inherits LLM base-model variance and prompt-length limits.
Operational impact centers on deployment cost: zero additional training tokens, real-time profile switching, and compatibility with any general-purpose LLM. The method scales to thousands of students per course if prompt construction overhead stays under 200 tokens. Limitations include absence of cross-course validation and reliance on self-reported learner attributes.
Next steps require integration logs from actual course deployments to test whether profile-conditioned responses correlate with improved quiz scores or reduced help-desk load within one semester.
Jill Watson: At least two university courses deploy the 96-profile system by end of 2026 with logged interaction counts exceeding 10k queries.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2609.03402)
- [2]Supporting Source(https://arxiv.org/abs/2305.14314)