PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Liam Byrne, David Dylan, Orla Fitzgerald, Eoin Doyle, Ciara Nolan, Padraig Lynch, Sinead Gallagher

arXiv:2609.01615v1cs.LG

TL;DR

Prompt-space meta-learning promises transferable personalization for frozen LLMs, but evidence that its objective captures cross-user adaptation is limited. The paper tests this with MUSE, which evolves and freezes a shared adaptation prompt for unseen users, and finds no transferable advantage over controls, with retrieval winning on rating. The authors identify meta-objective collapse and propose controls for detecting it.

  • Problem

    The paper asks whether prompt-space meta-learning learns transferable cross-user structure or merely produces generic instruction quality.

  • Method

    MUSE evolves one shared natural-language adaptation prompt over meta-train users, freezes it, and applies it zero-shot to held-out users using a frozen LLM.

  • Results

    MUSE does not beat its seed prompt or mismatched-support control, while plain retrieval decisively dominates persona methods on the rating benchmark.

  • Takeaways & Limitations

    The results attribute the failure to meta-objective collapse and motivate seed-prompt, wrong-support, and invariance/oracle controls for evaluating learned adaptation.

  • Takeaways & Limitations

    The study covers one backbone family, two LaMP tasks, and natural-language GEPA-evolved prompts, so it does not claim universality across models, tasks, or prompt classes.

Abstract

from arXiv · show

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.

I. INTRODUCTION

The paper tests whether a shared prompt evolved across users learns transferable personalization for a frozen LLM. Across two benchmarks, MUSE shows no evidence of transferable adaptation, while controls diagnose a structure-blind meta-objective and provide a reusable evaluation protocol.

  • Method: MUSE evolves one shared natural-language adaptation prompt over meta-train users, freezes it, and applies it zero-shot to held-out users.The method is evaluated in a controlled frozen-backbone harness with matched persona pipelines.
  • Headline result: MUSE does not significantly outperform its un-evolved seed prompt, a mismatched-support control, or the strongest persona baselines on either benchmark.These comparisons isolate learned adaptation from instruction phrasing and selection effects.
  • Headline result: On LAMP-3 rating, plain few-shot retrieval decisively outperforms persona methods, while LAMP-2 categorization methods cluster near parity.Figure 1 reports results over 200 held-out users per benchmark.
  • Implications: The seed-prompt, wrong-support, and invariance/oracle controls form a reusable protocol for separating learned adaptation from instruction polish and validation overfit.The invariance/oracle decomposition reports 80%/72% phrasing-invariant cases on LAMP-2/LAMP-3.
  • Diagnosis: ∆accuracy +0.033, 95% CI [−0.050, +0.117], p=0.555 on LAMP-2; ∆reward +0.013, [−0.029, +0.054], p=0.622 on LAMP-3 when user-support correspondence is scrambled.The meta-validation objective is statistically invariant to genuine versus mismatched support correspondence.
  • Diagnosis: An optimistic gap of ≈9 points separates LAMP-2 meta-validation from meta-test, while meta-validation does not significantly rank held-out performance.The reported Spearman correlation is ρ=+0.47, p=0.20, NS.

II. RELATED WORK

Related work positions frozen-LLM personalization at the intersection of user modeling, meta-learning, in-context learning, prompt optimization, and retrieval. The paper distinguishes how histories are represented from whether a shared policy learns transferable adaptation across users.

  • Personalization: LLM personalization commonly augments a frozen model with users’ labeled histories through profiles, summaries, personas, or retrieval.LaMP formalizes prediction of each user’s behavior on new items from past labeled interactions.
  • Prompt optimization: MUSE replaces fixed persona instructions with a meta-learned instruction, isolating whether meta-learning improves adaptation rather than merely changing representation.The paper treats representation and learning as orthogonal axes.
  • Meta-learning: The users-as-tasks view treats each user as a task, the profile as a support set, and a shared policy as the mechanism for rapid adaptation.This directly transcribes the standard meta-learning setup to personalization.
  • Meta-learning: Meta-learning methods seek an inductive bias—such as an initialization, optimizer, or metric space—that enables adaptation from few examples.The paper notes that shortcut solutions or memorization can fail to produce genuine task-adaptive behavior.
  • In-context learning: In-context learning can be interpreted as implicit meta-learning, with models inferring latent tasks from examples and applying them to queries.Prior analyses show transformers can implement learning algorithms in their forward pass over context.

D. Prompt Optimization and Evolution

The paper situates MUSE as reflective prompt evolution applied to a shared cross-user adaptation policy, contrasting its aggregate objective with grounded adaptation and retrieval. Its controls expose when prompt optimization is structure-blind.

  • D. Prompt Optimization and Evolution: Prompt optimizers search instruction text against validation metrics using discrete, reinforcement, compilation, or evolutionary strategies.GEPA-style methods mutate prompts from natural-language feedback on prior performance.
  • D. Prompt Optimization and Evolution: MUSE uses GEPA’s reflective loop to evolve a shared, frozen prompt rather than a per-task strategy.Elevating the prompt to a cross-user policy makes it a meta-learner.
  • D. Prompt Optimization and Evolution: An aggregate population objective can be structure-blind even when per-user evaluation appears reasonable, creating proxy-objective misalignment.The paper operationalizes this gap with a user-support derangement control.
  • E. Test-Time and Grounded Adaptation: Grounded test-time adaptation succeeds by using per-instance, verifiable feedback such as self-consistency checks, verifiers, unit tests, or tool results.MUSE lacks this instance-level grounding because its objective aggregates over users.
  • F. Retrieval and Retrieval-Augmented Generation: Retrieval-augmented few-shot prompts place relevant past user items verbatim in context and serve as strong personalization baselines.Retrieval is a representation strategy distinct from the learning axis tested by MUSE.
  • F. Retrieval and Retrieval-Augmented Generation: On the regression task, retrieval is the decisive winner rather than an outcome the paper minimizes.The retrieval advantage provides the main comparison point for persona-based methods.
  • III. Users as Tasks: Users are modeled as tasks with support sets and held-out queries, partitioned into disjoint meta-train, meta-validation, and meta-test populations.No user appears in more than one split, so generalization is measured across users.

B. Inner Adaptation and the Meta-Objective

MUSE adapts each user through a shared natural-language prompt that distills the user’s support set into a persona, then uses that persona to answer queries. Its meta-objective averages per-user performance and selects the final prompt on disjoint validation users.

  • Inner adaptation: The shared adaptation prompt π and each user’s support set produce a per-user persona θu through a single frozen-model call.The distillation template controls how the user history is summarized into the persona.
  • Inner adaptation: The persona is prepended to each query, which the frozen model answers using prompt concatenation.
  • Meta-objective: Per-user performance is the expected item score on that user’s queries.
  • Meta-objective: Accuracy measures categorization, while a bounded negative-error reward measures rating, with larger values better in both tasks.
  • Meta-objective: The meta-objective averages user performance, and the final π⋆ is selected by its score on disjoint meta-validation users.

C. What “Transfer” Requires, and the Structure-Blindness Test

The transfer test asks whether the objective depends on matching each user’s support set to that user’s queries. A derangement breaks this correspondence, while meta-objective collapse means the aligned and mismatched objectives are statistically invariant.

  • Transfer requirement: Meaningful prompt-space meta-learning requires the objective to reward personas adapted to the correct user’s support and queries.
  • Structure-blindness test: The derangement builds each persona from another user’s support set while grading it on the original user’s queries.
  • Structure-blindness test: Genuine adaptive signal would make aligned supports outperform scrambled supports by a margin the optimizer can exploit.
  • Structure-blindness test: Meta-objective collapse is the condition that the objective is statistically invariant to whether user–support correspondence is real.
  • MUSE comparison: MUSE differs from a fixed-instruction persona baseline only in whether its adaptation instruction is meta-learned or hand-written.
  • MUSE comparison: The inner loop uses one frozen-model call to turn labeled history into a bounded persona containing preferences, exemplars, and decision rules.

B. Outer Loop: Shared Prompt Evolution

MUSE evolves one shared prompt with reflective optimization, selects it on validation users, and applies it to held-out users through persona construction and answering. Seed and mismatch controls test whether evolution adds adaptation beyond instruction quality or structure-blind optimization.

  • B. Outer Loop: Shared Prompt Evolution: GEPA-style evolution maintains candidate prompts, selects parents by validation-user Pareto coverage, and reflectively mutates them using meta-train errors.
  • C. Controls Built into the Method: MUSE-SEED skips evolution and applies the hand-written π0, isolating instruction quality already present before learning.
  • C. Controls Built into the Method: MUSE-MISMATCH evolves under a deterministic derangement and tests on real users, isolating whether the objective carries cross-user adaptive signal.
  • C. Controls Built into the Method: The MUSE pipeline scores candidate prompts on users’ held-out query slices, freezes the best validation prompt, and applies it zero-shot to test users.
  • C. Controls Built into the Method: Algorithm 1 initializes the pool with π0, iterates prompt selection and reflective mutation within a call budget, then freezes the validation-selected π⋆.
  • C. Controls Built into the Method: At test time, MUSE builds each user persona from π⋆ and support data before answering the query under that persona.
  • C. Controls Built into the Method: The evaluation uses LaMP-2 personalized news categorization and LaMP-3 personalized product-review rating, with distinct classification and regression metrics.

B. Backbone and Decoding

The experiments use a shared frozen backbone and deterministic benchmark scoring across a disjoint users-as-tasks split. Results compare persona methods with retrieval and report item-level uncertainty and paired tests.

  • B. Backbone and Decoding: All methods use the frozen Qwen3-30B-A3B backbone with greedy decoding, so methods differ only in prepended text.
  • B. Backbone and Decoding: Both LaMP tasks use disjoint 40/30/200 meta-train, validation, and test users, with profiles containing at least 16 items.
  • B. Backbone and Decoding: The benchmark holds out one test item per test user from the profile while separating each profile into support and query slices.
  • C. Baselines: Baselines include no personalization, random-user personas, fixed-instruction personas, self-refining personas, and retrieved few-shot exemplars.
  • B. Backbone and Decoding: Metrics are accuracy and macro-F1 for LaMP-2 and MAE and RMSE for LaMP-3, evaluated over 200 test users with 95% bootstrap intervals.
  • B. Backbone and Decoding: Paired analyses use item-level McNemar tests for classification and 10,000-resample paired bootstraps for regression.
  • E. Seed and Mismatch Controls: The mismatch control uses a fixed no-fixed-point derangement, while support caps and meta-train sizes are swept to probe transfer.
  • E. Seed and Mismatch Controls: Persona and MUSE intervals overlap, whereas RAG-20 on LaMP-3 clears the persona interval.

VI. RESULTS

Across both benchmarks, evolving MUSE’s shared prompt does not improve over the seed or mismatched-support control. Retrieval is statistically tied on LaMP-2 but decisively better than persona methods on LaMP-3.

  • Finding 1: Meta-Learning Does Not Beat Its Own Seed: MUSE reaches LaMP-2 accuracy 0.765 versus the seed’s 0.775, with the difference statistically indistinguishable from zero.The McNemar test gives p=0.850.
  • Finding 2: Meta-Learning Does Not Beat a Structure-Broken Control: MUSE and the mismatched-support control are statistically tied on both benchmarks.LaMP-2 reports 0.765 versus 0.740 (p=0.441), while LaMP-3 ties at MAE 0.425 (p=1.000).
  • Finding 3: Parity With Persona Baselines on Categorization: LaMP-2 persona methods cluster around 0.74–0.78 accuracy and are mutually indistinguishable, despite exceeding un-personalized and random-user floors.MUSE versus PersonaLinkr1 is 0.765 versus 0.755, and MUSE versus PAG is 0.765 versus 0.745; both comparisons are non-significant.
  • Finding 4: Retrieval Dominates on Regression: LaMP-3 retrieval dominates personas: RAG-20 attains MAE 0.250 versus MUSE’s 0.425, with ∆MAE=+0.175 and p < 0.001.Even RAG-3/5 at MAE ≈0.29 beat all persona methods decisively.

VII. ANALYSIS: META-OBJECTIVE COLLAPSE

The analysis identifies meta-objective collapse: aligned and mismatched supports receive statistically indistinguishable validation scores, while prompt choices largely affect phrasing-invariant behavior rather than adaptation.

  • Analysis: Meta-Objective Collapse: Meta-objective collapse is diagnosed through no transfer over the seed, a structure-blind objective, meta-overfitting, and invariance/oracle analysis.These mechanisms explain the related null findings rather than treating them as independent outcomes.
  • Leg I: No Transfer — π⋆ Behaves Like the Seed: 80.0% of LaMP-2 users and 72.0% of LaMP-3 users are phrasing-invariant across evolved, seed, and mismatch prompts.Among users whose predictions swing, the evolved prompt does not beat the seed on either benchmark.
  • Leg II: A Structure-Blind Objective: The aligned and mismatched meta-validation objectives are statistically indistinguishable: LaMP-2 ∆=+0.033, p=0.555; LaMP-3 ∆=+0.013, p=0.622.The aligned score is numerically at least as high as the mismatched score on both benchmarks.
  • Leg II: A Structure-Blind Objective: Because the objective cannot distinguish genuine from scrambled supports, optimizing it cannot acquire cross-user adaptive structure and instead remains sensitive to generic phrasing.The paper characterizes this as proxy-objective misalignment between validation optimization and transferable adaptation.

C. Leg III: Meta-Overfitting at Small Population Scale

Meta-validation overfits a small selection population and fails to predict held-out performance; oracle headroom is attributed to phrasing variance rather than learned adaptation.

  • C. Leg III: Meta-Overfitting at Small Population Scale: Meta-test accuracy trails meta-validation by approximately +0.094 on average across nine LaMP-2 variants.Every variant lies below the y=x line.
  • C. Leg III: Meta-Overfitting at Small Population Scale: Meta-validation does not significantly rank meta-test accuracy: Spearman ρ=+0.47, p=0.20, and Kendall τ=+0.38, p=0.17.The highest meta-validation variant, K=16 at 0.872, has below-median test accuracy of 0.755.
  • Invariance and Oracle: The per-item oracle reaches accuracy 0.860 on LaMP-2 and MAE 0.245 on LaMP-3, but its inclusion of the mismatch prompt makes the headroom phrasing variance.The oracle therefore reinforces rather than softens the collapse diagnosis.
  • Semantic Analysis: Evolved prompts cluster by benchmark rather than by aligned-versus-mismatch status, despite substantial textual rewriting.This semantic pattern corroborates the structure-blind objective mechanism.

A. A Reusable Evaluation Protocol

The paper proposes three controls for distinguishing learned cross-user adaptation from instruction polish, structure-blind objectives, and phrasing variance. It argues that grounded, per-instance feedback is the promising route to restoring informative prompt-space meta-learning objectives.

  • A. A Reusable Evaluation Protocol: Seed control tests whether an optimized prompt beats the un-optimized seed, separating adaptation learning from prior instruction quality.Failure to improve over the seed indicates that optimization bought instruction polish rather than learned adaptation.
  • A. A Reusable Evaluation Protocol: Wrong-support control scrambles user–support assignments to test whether the selection objective is sensitive to correspondence structure.If the objective does not drop under scrambling, it cannot encode transferable adaptation.
  • A. A Reusable Evaluation Protocol: Invariance/oracle decomposition measures prompt-invariant instances and tests whether oracle headroom reflects distinct policies or paraphrase variance.Including a structure-broken prompt in the oracle set exposes gains attributable merely to phrasing differences.
  • A. A Reusable Evaluation Protocol: The protocol is reusable because all three controls reuse the existing pipeline and isolate distinct confounds.The controls target prior instruction quality, objective structure, and phrasing variance, respectively.
  • A. A Reusable Evaluation Protocol: Prompt-space meta-learning is expected to transfer when its meta-objective is sensitive to user–instance correspondence through grounded feedback.Suggested routes include per-user verifiable rewards, contrastive penalties for wrong-user personas, and richer support sets.
  • A. A Reusable Evaluation Protocol: The study’s claims are bounded to one backbone family, two LaMP tasks, and natural-language GEPA-evolved prompts.Soft prompts, gradient-based learners, contrastive objectives, other tasks, and other backbones remain outside the demonstrated scope.

APPENDIX A FULL SWEEP NUMBERS

The appendix reports sweep metrics, protocol settings, and prompt examples. The sweeps show that optimizing meta-validation does not ensure higher held-out accuracy, while the evolved prompts mainly add structured, benchmark-shaped guidance.

  • APPENDIX A FULL SWEEP NUMBERS: 0.872 meta-validation for K=16 has below-median test accuracy of 0.755 on LaMP-2.Meta-validation is the GEPA-selection aggregate over 30 users, whereas test accuracy is measured on 200 test users.
  • APPENDIX A FULL SWEEP NUMBERS: Higher meta-validation does not imply higher LaMP-2 test accuracy across the support-size and meta-train-size sweeps.The reported gap is meta-val − test accuracy, emphasizing that the optimized quantity is not the test metric.
  • APPENDIX A FULL SWEEP NUMBERS: The mismatch control uses a fixed seeded derangement with no fixed points, and all reflection, distillation, and answer calls use the same frozen backbone at temperature 0.Benchmark scoring is deterministic and does not use an LLM judge.
  • APPENDIX A FULL SWEEP NUMBERS: The seed prompt instructs the model to construct a persona from labeled history using stable preferences, label cues, decision rules, and exemplars.It emphasizes modeling the user’s decision policy rather than memorizing examples.
  • APPENDIX A FULL SWEEP NUMBERS: The evolved prompts reformat persona construction into structured JSON recipes with preferences, categorization cues, rules, exemplars, and tie disambiguation.The LaMP-3 recipe similarly covers stable preferences, noise filtering, label-specific cues, strictness or leniency, and near-tie decisions.
  • APPENDIX A FULL SWEEP NUMBERS: Aligned and mismatch-trained prompts cluster by benchmark rather than by support alignment, consistent with phrasing and formatting edits.Their aligned-versus-mismatch similarity is at the within-benchmark level, while within-benchmark similarity exceeds cross-benchmark similarity.
Loading 2609.01615v1…