Prompt-Space Meta-Learning Does Not Transfer Across Users: A Frozen-LLM Negative Result

A new study from arXiv (2609.01615v1) finds that prompt-space meta-learning does not transfer across users for personalizing frozen large language models, with the proposed Muse method failing to significantly improve over baselines on LaMP-2 and LaMP-3 benchmarks across 200 held-out users each. The authors attribute the failure to 'meta-objective collapse,' where the meta-validation objective is statistically invariant to genuine user-support correspondence (p=0.555 on LaMP-2, p=0.622 on LaMP-3), and Muse is dominated by plain few-shot retrieval on the rating task (Delta MAE +0.175, p < 0.001).

arXiv:2609.01615v1 Announce Type: new Abstract: Personalizing a frozen large language model LLM to individual users is often framed as a meta-learning problem in prompt space: each user is a task, and one seeks a shared natural-language adaptation policy that, given a handful of the user's labeled interactions, configures the frozen model for that user. The framing is attractive because it is backbone-agnostic and reuses the machinery of prompt optimization, yet the field rarely tests whether the optimized meta-objective encodes transferable cross-user adaptation rather than generic instruction quality. We study this question with Muse Meta-learned User-adaptation via Shared Evolution , which evolves a single shared adaptation prompt over a meta-train user population by reflective prompt evolution, freezes it, and applies it zero-shot to held-out users; matched controls isolate learning from confounds of phrasing and selection. On two standard personalization benchmarks LaMP-2 categorization and LaMP-3 rating over 200 held-out users each, Muse does not significantly improve on its own un-evolved seed prompt or on a structure-broken control that meta-trains on mismatched user-support pairs, and is dominated by plain few-shot retrieval on the rating task Delta MAE +0.175, p < 0.001 . We attribute these outcomes to a single mechanism, meta-objective collapse: the meta-validation objective is statistically invariant to whether the user-support correspondence is genuine p=0.555 on LaMP-2, p=0.622 on LaMP-3 , so it cannot be optimized into transferable adaptation and instead rewards instruction polish and validation overfitting. The seed-prompt, wrong-support, and invariance-oracle controls form a reusable protocol that separates learned adaptation from these confounds.