Prompt-Space Meta-Learning Does Not Transfer Across Users: A Frozen-LLM Negative Result
Not provided in the abstract
Abstract
The study investigates the effectiveness of personalizing a frozen large language model for individual users through meta-learning in prompt space, revealing that the approach does not yield transferable adaptation across users.
Reality Card
Muse, a method for evolving shared adaptation prompts, fails to improve user-specific performance compared to a non-evolved seed prompt and is outperformed by few-shot retrieval methods.
Muse does not significantly improve on its own un-evolved seed prompt or on a structure-broken control, with a Delta MAE of +0.175 on the rating task (p < 0.001).
The meta-validation objective is statistically invariant to genuine user-support correspondence, leading to non-transferable adaptation and overfitting.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.