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

# Prompt-Space Meta-Learning Does Not Transfer Across Users: A Frozen-LLM Negative Result
Source: [https://arxiv.org/html/2609.01615](https://arxiv.org/html/2609.01615)
Liam Byrne, David Dylan, Orla Fitzgerald, Eoin Doyle, Ciara Nolan, Padraig Lynch, and Sinead GallagherD\. Dylan, L\. Byrne, C\. Nolan and S\. Gallagher are with Trinity College Dublin; O\. Fitzgerald and P\. Lynch are with University College Dublin; E\. Doyle is with Dublin City University, Dublin, Ireland\.

###### 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 withMuse\(Meta\-learned User\-adaptation via Shared Evolution\), 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 \(LaMP\-2categorization andLaMP\-3rating\) over 200 held\-out users each,Musedoes 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 \(Δ\\DeltaMAE\+0\.175\+0\.175,p<0\.001p<0\.001\)\. 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 \(p=0\.555p\{=\}0\.555onLaMP\-2,p=0\.622p\{=\}0\.622onLaMP\-3\), 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\.

###### Index Terms:

Few\-shot learning, frozen language models, large language models, meta\-learning, negative transfer, personalization, prompt optimization, transfer learning, user modeling\.

![Refer to caption](https://arxiv.org/html/2609.01615v1/x1.png)Figure 1:The headline negative result\. On both benchmarks, meta\-learning a shared adaptation prompt withMuse\(π⋆\\pi^\{\\star\}\) is statistically indistinguishable from applying its own un\-evolved seed prompt \(π0\\pi\_\{0\}\) and from a structure\-broken control whose persona policy was meta\-trained on deliberately mismatched user–support pairs \(*mismatch*\)\. All three also tie the strongest persona baseline \(Pag\)\. On theLaMP\-3rating task, plain few\-shot retrieval \(Rag\) wins decisively \(p<0\.001p<0\.001\); onLaMP\-2categorization all methods cluster near parity\. Bars are accuracy \(↑\\uparrow, left\) and mean absolute error \(↓\\downarrow, right\) over200200held\-out users; whiskers are95%95\\%bootstrap confidence intervals\. The phrase “meta\-learning does not beat its own seed or a structure\-broken control” is the thesis of this paper\.## IIntroduction

Personalization has become a central requirement for large language model \(LLM\) services: the same model is expected to serve millions of users, yet each user brings idiosyncratic preferences, labeling tendencies, and style\. Because retraining or even parameter\-efficient finetuning a hosted model*per user*is operationally infeasible at scale, a large and growing body of work personalizes a*frozen*backbone by manipulating its input—retrieving the user’s past interactions, distilling them into a compact persona, or prepending a learned soft or natural\-language prompt\[[49](https://arxiv.org/html/2609.01615#bib.bib7),[45](https://arxiv.org/html/2609.01615#bib.bib8),[84](https://arxiv.org/html/2609.01615#bib.bib10),[58](https://arxiv.org/html/2609.01615#bib.bib11),[85](https://arxiv.org/html/2609.01615#bib.bib13)\]\. A natural and attractive way to systematize this is to view it as a*meta\-learning*problem\. If we treat each user as a task, with the user’s labeled history as a support set and their future interactions as queries, then personalization is exactly the few\-shot, fast\-adaptation regime that meta\-learning was designed for\[[9](https://arxiv.org/html/2609.01615#bib.bib87),[15](https://arxiv.org/html/2609.01615#bib.bib99)\]\. One would like to meta\-learn, across a population of users, a single*adaptation policy*that maps any new user’s support set to a configuration of the frozen model that performs well on that user’s queries\. Because the backbone is frozen and only its textual input is mutable, the most direct instantiation of this idea lives in*prompt space*: meta\-learn a shared natural\-language “adaptation prompt” that, when combined with a user’s history, induces a per\-user persona\. This framing is appealing on several counts—it is backbone\-agnostic, it reuses the mature machinery of prompt optimization and evolution\[[88](https://arxiv.org/html/2609.01615#bib.bib76),[75](https://arxiv.org/html/2609.01615#bib.bib19),[1](https://arxiv.org/html/2609.01615#bib.bib17)\], and it promises amortized, transferable adaptation rather than a bespoke pipeline per user\.

This paper asks whether prompt\-space meta\-learning actually learns transferable cross\-user structure, or whether it merely appears to\. We address the question with a clean instantiation of the idea,Muse\(Meta\-learned User\-adaptation via Shared Evolution\), in which a single shared adaptation promptπ\\piis evolved over a meta\-train population of users using reflective prompt evolution in the style ofGepa\[[1](https://arxiv.org/html/2609.01615#bib.bib17)\], scored by each candidate’s accuracy \(or negative error\) on meta\-train users’ own held\-out queries, selected on a*disjoint*meta\-validation population, frozen asπ⋆\\pi^\{\\star\}, and applied zero\-shot to build a persona for each of200200held\-out test users\.Musesits in an apples\-to\-apples harness with a frozen Qwen3\-30B\-A3B backbone\[[74](https://arxiv.org/html/2609.01615#bib.bib32)\], deterministic benchmark scoring, and a battery of strong baselines \(no\-personalization, random\-user, profile\-augmented generation, distilled personas, and retrieval\-augmented few\-shot at several budgets\)\. To our knowledge, this is the first controlled head\-to\-head test ofGepa\-style prompt evolution*as cross\-user meta\-learning*for frozen\-LLM personalization\.

The findings are consistent and, at first sight, counterintuitive\.Musewith the evolvedπ⋆\\pi^\{\\star\}does*not*significantly outperform the same pipeline run with its hand\-written seed promptπ0\\pi\_\{0\}; the seed is in fact numerically better on both benchmarks\.Musealso does*not*beat a deliberately structure\-broken control in which the shared prompt is meta\-trained on*mismatched*user–support pairs, where each user’s persona is built from a different user’s history, and it is statistically tied with the strongest persona baselines\. On the rating benchmark, simple few\-shot retrieval dominates all persona methods by a wide and highly significant margin\. These are not isolated null results but symptoms of a single mechanism, which we name and quantify asmeta\-objective collapse\. The aggregate meta\-validation objective that the optimizer climbs is structure\-blind: it is statistically invariant to whether the user–support correspondence is genuine\. Deranging the user\-to\-support assignment during meta\-training, which destroys the adaptive signal by construction, changes the selection objective by an amount indistinguishable from zero\. An objective that cannot distinguish aligned from scrambled supports cannot, even in principle, be optimized into transferable adaptation; the optimizer can acquire only generic instruction polish, which the seed prompt already supplies, together with overfitting of a small \(3030\-user\) validation population\. This is a concrete instance of*proxy\-objective misalignment*: the meta\-validation proxy that is optimized is decoupled from the true goal of transferable per\-user adaptation\.

This is a negative result, and we report it as one: the headline numbers, the retrieval advantage on regression, and every confidence interval are stated directly\. The contribution lies in the rigor of the diagnosis and in the controls that produce it\. The same three controls, namely a seed\-prompt baseline, a wrong\-support derangement, and an invariance/oracle decomposition, constitute a reusable evaluation protocol that can be applied before attributing learned adaptation to a prompt\-space method\. Without them, the instruction polish and validation overfit thatMuseexhibits are easily mistaken for genuine cross\-user transfer\.[Figure˜1](https://arxiv.org/html/2609.01615#S0.F1)previews the finding\.

Contributions\.The main contributions of this paper are summarized as follows\.

- •Muse: a clean instantiation and test of prompt\-space meta\-learning for personalization\.We formalize users\-as\-tasks personalization of a frozen LLM and instantiate it withMuse, the first apples\-to\-apples test ofGepa\-style reflective prompt evolution as a cross\-user meta\-learner\.Museevolves one shared natural\-language adaptation prompt over a meta\-train user population, freezes it, and applies it zero\-shot to held\-out users \([Sections˜III](https://arxiv.org/html/2609.01615#S3)and[IV](https://arxiv.org/html/2609.01615#S4)\)\.
- •The meta\-objective collapse diagnosis via a derangement control\.Using a mismatch \(wrong\-support\) control, we show that the meta\-validation objective is*statistically invariant*to scrambling the user–support correspondence \(Δ\\Deltaaccuracy\+0\.033\+0\.033,95%95\\%CI\[−0\.050,\+0\.117\]\[\-0\.050,\+0\.117\],p=0\.555p\{=\}0\.555onLaMP\-2;Δ\\Deltareward\+0\.013\+0\.013,\[−0\.029,\+0\.054\]\[\-0\.029,\+0\.054\],p=0\.622p\{=\}0\.622onLaMP\-3\)\. An objective blind to this structure carries no exploitable cross\-user adaptive signal \([Section˜VII](https://arxiv.org/html/2609.01615#S7)\)\.
- •Quantified meta\-overfitting at small population scale\.We measure a uniform optimistic gap of≈9\\approx\\\!9points between meta\-validation and meta\-test onLaMP\-2, show that meta\-validation does not significantly rank held\-out test performance \(Spearmanρ=\+0\.47\\rho\{=\}\+0\.47,p=0\.20p\{=\}0\.20,ns\), and observe flat, non\-monotone support\-size and meta\-train\-size sweeps—the signature of selection\-set overfitting rather than learned adaptation \([Section˜VII](https://arxiv.org/html/2609.01615#S7)\)\.
- •An invariance/oracle decomposition and a reusable evaluation protocol\.We decompose test behavior into phrasing\-invariant and phrasing\-sensitive cases \(80%80\\%/72%72\\%invariant onLaMP\-2/LaMP\-3\), attribute the oracle headroom to phrasing variance rather than learned adaptation, and package the seed\-prompt, wrong\-support, and invariance/oracle controls into a protocol for distinguishing learned adaptation from instruction polish and validation overfit \([Sections˜VII](https://arxiv.org/html/2609.01615#S7)and[VIII](https://arxiv.org/html/2609.01615#S8)\)\.

Roadmap\.[Section˜II](https://arxiv.org/html/2609.01615#S2)surveys LLM personalization, meta\-learning and fast adaptation, in\-context learning as meta\-learning, prompt optimization, test\-time adaptation, and retrieval, and positions our study on the*learning*axis \(does the meta\-objective improve adaptation over a seed?\), which is orthogonal to the representation axis of distillation versus retrieval\.[Section˜III](https://arxiv.org/html/2609.01615#S3)formalizes users\-as\-tasks personalization\.[Section˜IV](https://arxiv.org/html/2609.01615#S4)detailsMuse\.[Section˜V](https://arxiv.org/html/2609.01615#S5)describes the benchmarks, baselines, backbone, metrics, and controls\.[Section˜VI](https://arxiv.org/html/2609.01615#S6)reports the main results\.[Section˜VII](https://arxiv.org/html/2609.01615#S7)establishes meta\-objective collapse through its three mechanistic legs plus the invariance/oracle decomposition\.[Section˜VIII](https://arxiv.org/html/2609.01615#S8)draws the protocol\-level lessons, states when prompt\-space meta\-learning can and cannot be expected to work, and lists limitations\.[Section˜IX](https://arxiv.org/html/2609.01615#S9)concludes\.

## IIRelated Work

We organize prior work along six threads that jointly define the setting and the methods we test: \(i\) LLM personalization and user modeling; \(ii\) meta\-learning and fast adaptation; \(iii\) in\-context learning understood as meta\-learning; \(iv\) prompt optimization and evolution; \(v\) test\-time adaptation; and \(vi\) retrieval and retrieval\-augmented generation\. Throughout, we distinguish two orthogonal axes\. The*representation*axis asks how a user’s history is encoded into the frozen model’s input \(e\.g\., a distilled persona versus retrieved exemplars\); the sibling study “Distill or Retrieve?” isolates that axis and we therefore reference it exactly once, below, to make clear that our contribution is on a different axis\. The*learning*axis—our focus—asks whether a meta\-objective that climbs over a user population improves adaptation beyond a fixed seed policy\. Our negative finding lives entirely on the learning axis\.

### II\-ALLM Personalization and User Modeling

The Language Model Personalization \(LaMP\) benchmark suite established a standard protocol in which each user supplies a profile of past, labeled interactions and the model must predict the user’s behavior on new items\[[49](https://arxiv.org/html/2609.01615#bib.bib7),[24](https://arxiv.org/html/2609.01615#bib.bib14)\]\. A dominant family of methods augments the prompt with the user’s profile, either directly or after summarization\. Profile\-Augmented Generation \(PAG\) distills a user’s history into a natural\-language profile with a fixed, hand\-written instruction\[[45](https://arxiv.org/html/2609.01615#bib.bib8)\]; retrieval\-of\-profile\-then\-generate pipelines select the most relevant past items before answering\[[48](https://arxiv.org/html/2609.01615#bib.bib9)\]\. PersonaAgent and related agentic personalizers iterate a persona through self\-feedback\[[84](https://arxiv.org/html/2609.01615#bib.bib10)\], while plug\-in user encoders such as P\-PLUG andUserLLMcompress a user into soft tokens or embeddings that condition the backbone\[[30](https://arxiv.org/html/2609.01615#bib.bib12),[38](https://arxiv.org/html/2609.01615#bib.bib59)\]\. Parameter\-side personalization, exemplified by OPPU and per\-user parameter\-efficient modules\[[58](https://arxiv.org/html/2609.01615#bib.bib11),[59](https://arxiv.org/html/2609.01615#bib.bib57),[14](https://arxiv.org/html/2609.01615#bib.bib58)\], finetunes a small adapter per user; these are powerful but assume write access to model weights, which the hosted\-frozen\-backbone setting precludes\. A complementary literature models users for recommendation and preference prediction\[[71](https://arxiv.org/html/2609.01615#bib.bib36),[81](https://arxiv.org/html/2609.01615#bib.bib61),[36](https://arxiv.org/html/2609.01615#bib.bib33)\], builds explicit long\-term memory of user state\[[40](https://arxiv.org/html/2609.01615#bib.bib24),[87](https://arxiv.org/html/2609.01615#bib.bib60)\], and studies persona consistency and role fidelity\[[83](https://arxiv.org/html/2609.01615#bib.bib54),[70](https://arxiv.org/html/2609.01615#bib.bib55),[60](https://arxiv.org/html/2609.01615#bib.bib37),[52](https://arxiv.org/html/2609.01615#bib.bib63),[68](https://arxiv.org/html/2609.01615#bib.bib64),[61](https://arxiv.org/html/2609.01615#bib.bib65),[50](https://arxiv.org/html/2609.01615#bib.bib66),[7](https://arxiv.org/html/2609.01615#bib.bib67)\]\. Surveys of LLM personalization and of personalized LLMs\[[85](https://arxiv.org/html/2609.01615#bib.bib13),[29](https://arxiv.org/html/2609.01615#bib.bib34),[65](https://arxiv.org/html/2609.01615#bib.bib62)\]catalog these approaches but, as we argue here, rarely separate genuine learned adaptation from prompt\-engineering artifacts\. OurMuseuses the same distilled\-persona representation as PAG and PersonaLink, but replaces their fixed adaptation instruction with a meta\-learned one, isolating exactly the “does meta\-learning help?” question\.

### II\-BMeta\-Learning and Fast Adaptation

Meta\-learning seeks an inductive bias—an initialization, an optimizer, or a metric space—that enables rapid adaptation to new tasks from few examples\[[15](https://arxiv.org/html/2609.01615#bib.bib99)\]\. Gradient\-based methods learn an initialization from which a few SGD steps suffice: MAML and its relatives\[[9](https://arxiv.org/html/2609.01615#bib.bib87),[37](https://arxiv.org/html/2609.01615#bib.bib88),[27](https://arxiv.org/html/2609.01615#bib.bib90),[41](https://arxiv.org/html/2609.01615#bib.bib92),[4](https://arxiv.org/html/2609.01615#bib.bib91)\], and learned\-optimizer approaches\[[3](https://arxiv.org/html/2609.01615#bib.bib98),[43](https://arxiv.org/html/2609.01615#bib.bib89)\]\. Metric\-based methods learn an embedding in which simple nearest\-neighbor or prototype rules generalize\[[56](https://arxiv.org/html/2609.01615#bib.bib93),[62](https://arxiv.org/html/2609.01615#bib.bib94),[39](https://arxiv.org/html/2609.01615#bib.bib95)\], and memory\- or attention\-augmented learners amortize adaptation in a forward pass\[[51](https://arxiv.org/html/2609.01615#bib.bib96),[33](https://arxiv.org/html/2609.01615#bib.bib97)\]\. Continual and online variants study adaptation without catastrophic forgetting\[[19](https://arxiv.org/html/2609.01615#bib.bib100)\], and the paradigm has been applied to recommendation, machine translation, and control as a few\-shot user/task problem\[[25](https://arxiv.org/html/2609.01615#bib.bib103),[12](https://arxiv.org/html/2609.01615#bib.bib102),[79](https://arxiv.org/html/2609.01615#bib.bib101)\]\. The users\-as\-tasks view we adopt is the natural transcription of this literature to personalization: a user is a task, the profile is a support set, and we seek a shared policy that adapts fast\. A central and well\-known risk in meta\-learning is that the meta\-objective can be optimized in ways that do not produce genuine task\-adaptive behavior—“memorization” or shortcut solutions that exploit the meta\-training distribution rather than learning to adapt\[[41](https://arxiv.org/html/2609.01615#bib.bib92),[79](https://arxiv.org/html/2609.01615#bib.bib101)\]\. Our negative result is a sharp, prompt\-space manifestation of exactly this risk\.

### II\-CIn\-Context Learning as Meta\-Learning

A parallel line interprets in\-context learning \(ICL\) itself as implicit meta\-learning: a model trained on many sequences learns to infer a latent task from a few in\-context examples and apply it to a query\.MetaICLexplicitly meta\-trains a model to do ICL across tasks\[[32](https://arxiv.org/html/2609.01615#bib.bib104)\]; theoretical and mechanistic analyses show that transformers can implement learning algorithms \(e\.g\., gradient descent or ridge regression\) in their forward pass over the context\[[2](https://arxiv.org/html/2609.01615#bib.bib105),[63](https://arxiv.org/html/2609.01615#bib.bib106)\]\. This perspective is important for our study because it clarifies what a frozen backbone already does: given a user’s exemplars in context, the model performs a form of implicit adaptation*for free*\. The retrieval baselines we compare against exploit precisely this—they place raw user exemplars in the context and let the frozen model adapt implicitly—whereasMuseattempts to meta\-learn an*explicit*adaptation instruction on top\. One reading of our results is that the explicit prompt\-space meta\-objective adds little over the implicit adaptation the backbone already supplies\.

### II\-DPrompt Optimization and Evolution

Because the backbone is frozen, prompt\-space personalization inherits the toolkit of automatic prompt optimization\. Discrete search and reinforcement methods such as APE and OPRO search instruction text against a validation metric\[[88](https://arxiv.org/html/2609.01615#bib.bib76),[75](https://arxiv.org/html/2609.01615#bib.bib19)\];DSPycompiles and optimizes multi\-stage prompt programs\[[22](https://arxiv.org/html/2609.01615#bib.bib18)\]; and evolutionary or reflective methods—Promptbreeder, ExpeL, agentic memory induction, and notablyGepa’s genetic–Pareto reflective evolution—mutate prompts from natural\-language feedback on prior rollouts\[[8](https://arxiv.org/html/2609.01615#bib.bib77),[86](https://arxiv.org/html/2609.01615#bib.bib20),[69](https://arxiv.org/html/2609.01615#bib.bib21),[1](https://arxiv.org/html/2609.01615#bib.bib17),[82](https://arxiv.org/html/2609.01615#bib.bib22)\]\. Reflective self\-improvement of model outputs more broadly\[[31](https://arxiv.org/html/2609.01615#bib.bib15),[54](https://arxiv.org/html/2609.01615#bib.bib16),[11](https://arxiv.org/html/2609.01615#bib.bib74),[16](https://arxiv.org/html/2609.01615#bib.bib78),[80](https://arxiv.org/html/2609.01615#bib.bib81),[72](https://arxiv.org/html/2609.01615#bib.bib82)\]shares the same engine: an LLM critiques and rewrites text against a target\. OurMuseusesGepa’s loop verbatim, but elevates the optimized prompt from a per\-task strategy to a*shared, frozen, cross\-user adaptation policy*\. This is what makes it a meta\-learner rather than a per\-instance optimizer, and it is also what exposes the failure mode: when the optimization target is an aggregate over a user population, the target can be structure\-blind even when each per\-user evaluation looks reasonable\. We frame this failure as*proxy\-objective misalignment*, in the sense surveyed for LLM and agent objectives by\[[77](https://arxiv.org/html/2609.01615#bib.bib3)\]: the meta\-validation proxy thatGepamaximizes is decoupled from the true goal of transferable adaptation, and our derangement control*is*the operational demonstration of that proxy\-versus\-goal gap\.

### II\-ETest\-Time and Grounded Adaptation

Test\-time adaptation updates a model on the fly using the test input itself\. Test\-time training and entropy\-minimization methods such as TTT, Tent, and gradient\-edited variants adapt parameters per instance\[[57](https://arxiv.org/html/2609.01615#bib.bib107),[64](https://arxiv.org/html/2609.01615#bib.bib108),[34](https://arxiv.org/html/2609.01615#bib.bib109)\], while recent “thinking at test time” approaches scale inference\-time computation against a verifiable signal\[[55](https://arxiv.org/html/2609.01615#bib.bib85),[35](https://arxiv.org/html/2609.01615#bib.bib86),[78](https://arxiv.org/html/2609.01615#bib.bib80),[67](https://arxiv.org/html/2609.01615#bib.bib79)\]\. What unites the*successful*members of this family is a per\-instance, grounded feedback signal: a self\-consistency check, a verifier, a unit test, or a tool\-execution result that tells the adapter, for*this*input, whether it is on track\. The contrast with our setting is the crux of the paper\. Grounded, instance\-level test\-time adaptation—for example, agentic tool reasoning that is verified per call against execution feedback\[[76](https://arxiv.org/html/2609.01615#bib.bib2)\]—can genuinely improve because its objective is tied to a checkable outcome on each instance\.Muse’s meta\-objective is the opposite: it is an aggregate over a population, with no per\-instance grounding, and that is precisely why it collapses\. WhatMuselacks is exactly the instance\-level, verifiable feedback that makes grounded test\-time adaptation work\. We return to this contrast in[Section˜VIII](https://arxiv.org/html/2609.01615#S8)\.

### II\-FRetrieval and Retrieval\-Augmented Generation

The strongest baselines in our study are retrieval\-augmented few\-shot prompts, which fetch a user’s most relevant past items and place them, verbatim, in the context\. This builds on classical sparse and dense retrieval\[[46](https://arxiv.org/html/2609.01615#bib.bib29),[20](https://arxiv.org/html/2609.01615#bib.bib39),[17](https://arxiv.org/html/2609.01615#bib.bib41),[73](https://arxiv.org/html/2609.01615#bib.bib45),[10](https://arxiv.org/html/2609.01615#bib.bib46),[23](https://arxiv.org/html/2609.01615#bib.bib43)\], retrieval\-augmented language modeling and generation\[[26](https://arxiv.org/html/2609.01615#bib.bib38),[13](https://arxiv.org/html/2609.01615#bib.bib40),[6](https://arxiv.org/html/2609.01615#bib.bib44),[18](https://arxiv.org/html/2609.01615#bib.bib42),[42](https://arxiv.org/html/2609.01615#bib.bib50),[53](https://arxiv.org/html/2609.01615#bib.bib47),[21](https://arxiv.org/html/2609.01615#bib.bib48),[5](https://arxiv.org/html/2609.01615#bib.bib49)\], and demonstration selection for ICL\[[28](https://arxiv.org/html/2609.01615#bib.bib51),[47](https://arxiv.org/html/2609.01615#bib.bib52),[66](https://arxiv.org/html/2609.01615#bib.bib53)\]\. For personalization specifically, retrieving the user’s own history is a very strong, training\-free baseline because it preserves fine\-grained, instance\-level signal \(e\.g\., a user’s exact rating calibration\) that a distilled persona discards\. The orthogonal axis of*representation*—whether to distill a user into a compact persona or to retrieve raw exemplars—is studied directly by the sibling work “Distill or Retrieve?”; we cite it here once to delimit scope and to make explicit that the present paper does*not*re\-litigate that axis\. Our study holds the representation fixed \(a distilled persona, as in PAG/PersonaLink\) and varies only the*learning*of the adaptation policy, asking whether meta\-learning the policy beats a seed\. The retrieval baselines enter our tables as a reference point and, on the regression task, as the decisive winner—an honest fact we report rather than minimize\.

## IIIProblem Formulation

### III\-AUsers as Tasks

Letffdenote a*frozen*autoregressive LLM that maps a textual prompt to a textual output; its parameters are fixed and never updated\. Personalization must therefore act entirely throughff’s input\. We adopt an episodic, users\-as\-tasks formulation\. A useruuis a task with an associated distribution𝒟u\\mathcal\{D\}\_\{u\}over labeled items\(x,y\)\(x,y\), wherexxis an input \(e\.g\., a news article or a product review\) andyyis the label this user would assign \(a category or a rating\)\. Each user exposes a finite*support set*Su=\{\(xi,yi\)\}i=1muS\_\{u\}=\\\{\(x\_\{i\},y\_\{i\}\)\\\}\_\{i=1\}^\{m\_\{u\}\}\(their profile of past labeled interactions\) and is evaluated on held\-out*query*items drawn from the same𝒟u\\mathcal\{D\}\_\{u\}\. The population of users is partitioned into three mutually disjoint sets: a meta\-train set𝒰tr\\mathcal\{U\}\_\{\\text\{tr\}\}, a meta\-validation set𝒰val\\mathcal\{U\}\_\{\\text\{val\}\}, and a meta\-test set𝒰te\\mathcal\{U\}\_\{\\text\{te\}\}\. No user appears in more than one set, so generalization is measured across users, not merely across items within a user\.[Table˜I](https://arxiv.org/html/2609.01615#S3.T1)collects the notation\.

TABLE I:Notation used throughout the paper\.
### III\-BInner Adaptation and the Meta\-Objective

Adaptation to a user is mediated by a shared, learnable natural\-language*adaptation prompt*π\\pi, the meta\-parameter\. The inner adaptation operator𝒜\\mathcal\{A\}takesπ\\piand a user’s support set and produces a per\-user personaθu\\theta\_\{u\}by a single call to the frozen model,

θu=𝒜​\(π,Su\)=f​\(distill​\(π,Su\)\),\\theta\_\{u\}\\;=\\;\\mathcal\{A\}\(\\pi,S\_\{u\}\)\\;=\\;f\\big\(\\,\\textsc\{distill\}\(\\pi,S\_\{u\}\)\\,\\big\),\(1\)wheredistill​\(π,Su\)\\textsc\{distill\}\(\\pi,S\_\{u\}\)is a fixed template that injectsπ\\pias the controlling instruction over a digest of the user’s history and elicits a bounded persona \(a preference summary, a few exemplars, and a few decision rules; see[Section˜IV](https://arxiv.org/html/2609.01615#S4)\)\. The personaθu\\theta\_\{u\}is then prepended to a query and the frozen model answers,

y^=f​\(θu∥x\),\(x,y\)∈Qu,\\hat\{y\}\\;=\\;f\\big\(\\,\\theta\_\{u\}\\,\\\|\\,x\\,\\big\),\\qquad\(x,y\)\\in Q\_\{u\},\(2\)with “∥\\\|” denoting prompt concatenation\. The per\-user performance ofπ\\piis the expected item score on that user’s queries,

Ju​\(π\)=𝔼\(x,y\)∼Qu​\[ℓ​\(f​\(𝒜​\(π,Su\)∥x\),y\)\],J\_\{u\}\(\\pi\)\\;=\\;\\mathbb\{E\}\_\{\(x,y\)\\sim Q\_\{u\}\}\\Big\[\\,\\ell\\big\(f\(\\mathcal\{A\}\(\\pi,S\_\{u\}\)\\\|x\),\\,y\\big\)\\Big\],\(3\)whereℓ\\ellis accuracy for categorization and a bounded negative\-error reward for rating \(so that larger is better in both cases\)\. The*meta\-objective*is the population average over a user set𝒰\\mathcal\{U\},

ℳ𝒰​\(π\)=1\|𝒰\|​∑u∈𝒰Ju​\(π\)\.\\mathcal\{M\}\_\{\\mathcal\{U\}\}\(\\pi\)\\;=\\;\\frac\{1\}\{\|\\mathcal\{U\}\|\}\\sum\_\{u\\in\\mathcal\{U\}\}J\_\{u\}\(\\pi\)\.\(4\)Meta\-learning seeksπ⋆=arg⁡maxπ⁡ℳ𝒰tr​\(π\)\\pi^\{\\star\}=\\arg\\max\_\{\\pi\}\\,\\mathcal\{M\}\_\{\\mathcal\{U\}\_\{\\text\{tr\}\}\}\(\\pi\), with the finalπ⋆\\pi^\{\\star\}*selected*by its scoreℳ𝒰val\\mathcal\{M\}\_\{\\mathcal\{U\}\_\{\\text\{val\}\}\}on the disjoint meta\-validation users, and reported on the held\-out meta\-test users via the benchmark’s native metric\.

### III\-CWhat “Transfer” Requires, and the Structure\-Blindness Test

For prompt\-space meta\-learning to be meaningful,ℳ𝒰​\(π\)\\mathcal\{M\}\_\{\\mathcal\{U\}\}\(\\pi\)must rewardπ\\pifor producing personas that are genuinely*adapted*to the right user—that is, the objective must depend on the correspondence between a user’s support setSuS\_\{u\}and that user’s queriesQuQ\_\{u\}\. We make this requirement operational with a derangement\. Letσ\\sigmabe a permutation of𝒰tr\\mathcal\{U\}\_\{\\text\{tr\}\}with no fixed point, and define the*mismatched*meta\-objective

ℳ𝒰σ​\(π\)=1\|𝒰\|​∑u∈𝒰𝔼\(x,y\)∼Qu​\[ℓ​\(f​\(𝒜​\(π,Sσ​\(u\)\)∥x\),y\)\],\\mathcal\{M\}^\{\\sigma\}\_\{\\mathcal\{U\}\}\(\\pi\)\\;=\\;\\frac\{1\}\{\|\\mathcal\{U\}\|\}\\sum\_\{u\\in\\mathcal\{U\}\}\\mathbb\{E\}\_\{\(x,y\)\\sim Q\_\{u\}\}\\Big\[\\,\\ell\\big\(f\(\\mathcal\{A\}\(\\pi,S\_\{\\sigma\(u\)\}\)\\\|x\),\\,y\\big\)\\Big\],\(5\)in which each user’s persona is built from*another*user’s support setSσ​\(u\)S\_\{\\sigma\(u\)\}but still graded on the original user’s queriesQuQ\_\{u\}\. If the objective carries genuine cross\-user adaptive signal, then aligned supports must help relative to scrambled ones, i\.e\.,ℳ𝒰​\(π\)\>ℳ𝒰σ​\(π\)\\mathcal\{M\}\_\{\\mathcal\{U\}\}\(\\pi\)\>\\mathcal\{M\}^\{\\sigma\}\_\{\\mathcal\{U\}\}\(\\pi\)by a margin the optimizer can exploit\. We definemeta\-objective collapseas the empirical condition

ℳ𝒰val​\(π\)≈ℳ𝒰valσ​\(π\),\\mathcal\{M\}\_\{\\mathcal\{U\}\_\{\\text\{val\}\}\}\(\\pi\)\\;\\approx\\;\\mathcal\{M\}^\{\\sigma\}\_\{\\mathcal\{U\}\_\{\\text\{val\}\}\}\(\\pi\),\(6\)i\.e\., the meta\-objective is statistically invariant to whether the user–support correspondence is real\. Under[Equation˜6](https://arxiv.org/html/2609.01615#S3.E6), no optimizer can convertℳ\\mathcal\{M\}into transferable adaptation, because the quantity it climbs does not distinguish adaptation from its structure\-destroyed counterpart\.[Section˜VII](https://arxiv.org/html/2609.01615#S7)shows that[Equation˜6](https://arxiv.org/html/2609.01615#S3.E6)holds on both benchmarks\.

## IVMethod: Muse

Muse\(Meta\-learned User\-adaptation via Shared Evolution\) is the most direct instantiation of the formulation in[Section˜III](https://arxiv.org/html/2609.01615#S3)that is compatible with a hosted, frozen backbone\. It meta\-learns a*single*shared adaptation prompt over a user population, freezes it, and applies it zero\-shot to unseen users\. The only quantity that differs betweenMuseand a fixed\-instruction persona baseline such as PAG is*where the adaptation instruction comes from*: PAG uses a hand\-written instruction;Museuses a meta\-learned one\.[Figure˜2](https://arxiv.org/html/2609.01615#S4.F2)sketches the pipeline\.

meta\-train \(evolveπ\\pi\)meta\-test \(frozenπ⋆\\pi^\{\\star\}, zero\-shot\)Meta\-train users𝒰tr\\mathcal\{U\}\_\{\\text\{tr\}\}\(40\)supportSuS\_\{u\}/ queryQuQ\_\{u\}Gepaevolve\(reflective mutationof sharedπ\\pi\)scoreπ\\pion eachuser’s held\-outQuQ\_\{u\}[Equation˜3](https://arxiv.org/html/2609.01615#S3.E3)–\([4](https://arxiv.org/html/2609.01615#S3.E4)\)Meta\-val users𝒰val\\mathcal\{U\}\_\{\\text\{val\}\}\(30, disjoint\)Freezeπ⋆\\pi^\{\\star\}Held\-out useru∈𝒰teu\\in\\mathcal\{U\}\_\{\\text\{te\}\}\(200\)supportSuS\_\{u\}Distill personaθu=𝒜​\(π⋆,Su\)\\theta\_\{u\}=\\mathcal\{A\}\(\\pi^\{\\star\},S\_\{u\}\)1 frozen\-LLM callAnswer queryy^=f​\(θu∥x\)\\hat\{y\}=f\(\\theta\_\{u\}\\,\\\|\\,x\)1 frozen\-LLM callfeedbackselect bestπ\\piarg⁡max⁡ℳ𝒰val\\arg\\max\\mathcal\{M\}\_\{\\mathcal\{U\}\_\{\\text\{val\}\}\}zero\-shotFigure 2:TheMusepipeline\. A single shared adaptation promptπ\\piis evolved over a meta\-train user population withGepa\-style reflective mutation, where each candidateπ\\piis scored by the accuracy / negative\-error reward of the personas it induces on each meta\-train user’s own held\-out query slice \([Equations˜3](https://arxiv.org/html/2609.01615#S3.E3)and[4](https://arxiv.org/html/2609.01615#S3.E4)\)\. The bestπ\\pion the disjoint meta\-validation users is frozen asπ⋆\\pi^\{\\star\}and applied zero\-shot to each held\-out test user: one frozen\-LLM call distills the user’s support into a personaθu\\theta\_\{u\}, and one call answers the query under that persona\. The backbone is frozen throughout; only the textual prompt changes\.### IV\-AInner Loop: Persona Distillation underπ\\pi

Given a user’s support set, the inner loop issues a single call to the frozen model in whichπ\\piis injected as the controlling instruction over a digest of the user’s labeled history, and the model returns a bounded persona in a fixed JSON schema with three fields: a≤600\\leq\\\!600\-character preference summary, up to six induced exemplars \(each an input\-gist→\\tolabel pair\), and up to six short decision rules of the form “if*cue*then*label*”\. The persona is clipped to these bounds, rendered into a compact natural\-language preamble, and reused for every query of that user\. The schema, the parser, the clipping, and the answer call are shared verbatim with the PersonaLink/PAG harness so that the*only*delta across these methods is the adaptation instruction\. The task\-specific information \(the1515LaMP\-2categories or the11–55rating scale\) is supplied by the data digest and the fixed template, not byπ\\pi, so that a singleπ\\pican in principle serve both benchmarks\. Answering is a single greedy call with the persona prepended to the query; benchmark scoring is deterministic, with no LLM judge\.

### IV\-BOuter Loop: Shared Prompt Evolution

The meta\-parameterπ\\piis optimized withGepa’s genetic–Pareto reflective evolution\[[1](https://arxiv.org/html/2609.01615#bib.bib17)\], reused without modification to its core loop\. The optimizer maintains a growing pool of candidate prompts, seeded with the hand\-writtenπ0\\pi\_\{0\}\. Each iteration samples a parent from the pool by instance\-frontier \(Pareto\) coverage over the meta\-validation users, draws a meta\-train minibatch, and—if the parent is not already perfect on the minibatch—reflectively mutates it: the same frozen model is shown the currentπ\\pitogether with per\-user digests of where theπ\\pi\-built personas mispredicted, and is asked to rewriteπ\\piinto a better*general*recipe for persona construction \(how to infer stable preferences from limited history, surface label\-specific cues and user biases, turn patterns into crisp rules and well\-chosen exemplars, and break near\-ties\), under an explicit instruction not to specialize to any single user\. A child is accepted if it does not worsen the minibatch score; accepted children are re\-scored on the full meta\-validation set and added to the pool\. After the rollout budget is exhausted, the candidate with the best meta\-validation aggregate is frozen asπ⋆\\pi^\{\\star\}\.[Algorithm˜1](https://arxiv.org/html/2609.01615#alg1)states the procedure\. Two design points are essential to the honesty of the test\. First, the reflection model, the inner model, and the answer model are the*same*frozen backbone, so no extra capacity is smuggled in\. Second,π\\piis*shared*across all users and then frozen; it is never tuned to a test user, which is what makesMusea meta\-learner rather than a per\-user optimizer\.

Algorithm 1Muse: shared\-prompt meta\-learning for frozen\-LLM personalization\.0:frozen LLM

ff; meta\-train users

𝒰tr\\mathcal\{U\}\_\{\\text\{tr\}\}, meta\-val

𝒰val\\mathcal\{U\}\_\{\\text\{val\}\}; seed prompt

π0\\pi\_\{0\}; support cap

KK; query size; rollout budget

BB
1:pool

←\{π0\}\\leftarrow\\\{\\pi\_\{0\}\\\}; record

ℳ𝒰val​\(π0\)\\mathcal\{M\}\_\{\\mathcal\{U\}\_\{\\text\{val\}\}\}\(\\pi\_\{0\}\)via \([4](https://arxiv.org/html/2609.01615#S3.E4)\)

2:whilecalls

<B<Bdo

3:

π←\\pi\\leftarrowParetoSelect\(pool,

𝒰val\\mathcal\{U\}\_\{\\text\{val\}\}\)// instance\-frontier coverage

4:draw meta\-train minibatch

M⊆𝒰trM\\subseteq\\mathcal\{U\}\_\{\\text\{tr\}\}
5:for all

u∈Mu\\in Mdo

6:split profile into support

SuS\_\{u\}\(cap

KK\) and held\-out query

QuQ\_\{u\}
7:

θu←𝒜​\(π,Su\)\\theta\_\{u\}\\leftarrow\\mathcal\{A\}\(\\pi,S\_\{u\}\)// \([1](https://arxiv.org/html/2609.01615#S3.E1)\): 1 call

8:score

θu\\theta\_\{u\}on

QuQ\_\{u\}; collect mispredictions// \([3](https://arxiv.org/html/2609.01615#S3.E3)\)

9:endfor

10:ifminibatch score is not perfectthen

11:

π′←\\pi^\{\\prime\}\\leftarrowReflectMutate\(f,π,\{mispredictions\}\)\(f,\\pi,\\\{\\text\{mispredictions\}\\\}\)

12:accept

π′\\pi^\{\\prime\}if it does not worsen the minibatch score

13:if accepted: score

π′\\pi^\{\\prime\}on

𝒰val\\mathcal\{U\}\_\{\\text\{val\}\}; add to pool

14:endif

15:endwhile

16:

π⋆←arg⁡maxπ∈pool⁡ℳ𝒰val​\(π\)\\pi^\{\\star\}\\leftarrow\\arg\\max\_\{\\pi\\in\\text\{pool\}\}\\mathcal\{M\}\_\{\\mathcal\{U\}\_\{\\text\{val\}\}\}\(\\pi\)
17:freeze

π⋆\\pi^\{\\star\}
18:at test:for each

u∈𝒰teu\\in\\mathcal\{U\}\_\{\\text\{te\}\}, build

θu=𝒜​\(π⋆,Su\)\\theta\_\{u\}=\\mathcal\{A\}\(\\pi^\{\\star\},S\_\{u\}\)and answer

y^=f​\(θu∥x\)\\hat\{y\}=f\(\\theta\_\{u\}\\\|x\)
19:return

π⋆\\pi^\{\\star\}, meta\-test predictions

### IV\-CControls Built into the Method

Two controlled variants ofMuseare central to the diagnosis and are run through the identical pipeline\. Theseedcontrol \(Muse\-seed\) skips evolution entirely and applies the frozen, un\-meta\-learnedπ0\\pi\_\{0\}to the test users; it isolates how much of any apparent gain is generic instruction quality already present before learning\. Themismatchcontrol \(Muse\-mismatch\) meta\-trainsπ\\piunder the deranged objective of[Equation˜5](https://arxiv.org/html/2609.01615#S3.E5)—each meta\-train user’s persona is built from a deterministically shuffled other user’s support set, while still graded on the user’s own queries—then tests on the real users; it isolates whether the meta\-objective carries any cross\-user adaptive signal at all\. IfMuse\(π⋆\)\(\\pi^\{\\star\}\)cannot beatMuse\-seed, the evolution bought nothing the seed lacked; if it cannot beatMuse\-mismatch, the objective it climbed was structure\-blind\. Both turn out to hold\.

## VExperimental Setup

### V\-ABenchmarks and User Splits

We evaluate on two tasks from the LaMP suite\[[49](https://arxiv.org/html/2609.01615#bib.bib7)\]\.LaMP\-2is personalized news categorization: given an article, predict which of1515categories*this*user would assign; the metric is accuracy \(↑\\uparrow\) with macro\-F1 as a secondary measure\.LaMP\-3is personalized product\-review rating: predict the11–55score*this*user would give; the metrics are mean absolute error \(MAE,↓\\downarrow\) and root mean squared error \(RMSE,↓\\downarrow\)\. Each LaMP question identifier corresponds to one user with a profile of past labeled items\. From the development split we form a candidate user pool by keeping users with at least1616profile items and removing the200200designated test users; the meta\-train and meta\-validation users are drawn from this pool, yielding three mutually disjoint user sets:\|𝒰tr\|=40\|\\mathcal\{U\}\_\{\\text\{tr\}\}\|\{=\}40,\|𝒰val\|=30\|\\mathcal\{U\}\_\{\\text\{val\}\}\|\{=\}30, and\|𝒰te\|=200\|\\mathcal\{U\}\_\{\\text\{te\}\}\|\{=\}200per benchmark\. Disjointness is enforced at the user level, so all reported generalization is across users\. Within a user, the profile is split into a support pool and a held\-out query slice that never overlap; the test users’ single LaMP test item is never part of their profile\.[Table˜II](https://arxiv.org/html/2609.01615#S5.T2)summarizes the resulting per\-benchmark configuration; the two tasks share an identical user\-split protocol and differ only in label space \(a1515\-way category set versus a11–55ordinal scale\) and metric family\.

TABLE II:Dataset and user\-split statistics\. Both LaMP tasks share the same users\-as\-tasks protocol \(disjoint4040/3030/200200meta\-train/val/test users,≥16\\geq 16\-item profile filter\); they differ only in the label space and metric family\. “Support / query” are the per\-user inner\-loop slices; the single LaMP test item per test user is held out of the profile\.
### V\-BBackbone and Decoding

All methods share one frozen backbone, Qwen3\-30B\-A3B\[[74](https://arxiv.org/html/2609.01615#bib.bib32)\], served locally on four GPUs\. Decoding is greedy \(temperature0, single sample\) for every call—persona distillation, reflection, and answering—so that the only per\-method difference is the text prepended to an identical question\. We use no LLM judge; all scoring is the benchmark’s deterministic metric\. Because the serving stack is not bitwise\-deterministic across runs, we treat the*item*as the unit of analysis \(rather than averaging over decoding seeds\) and report paired, per\-item statistics, following our harness’s convention\.

### V\-CBaselines

We compareMuseagainst a ladder of training\-free and prompt\-optimization baselines, all in the same harness\.NoPersanswers with no personalization \(a zero\-knowledge lower bound\)\.RandomUserbuilds the persona from a*random other*user’s profile, controlling for the mere presence of persona text\.PAG\[[45](https://arxiv.org/html/2609.01615#bib.bib8)\]distills the user’s profile into a natural\-language persona with a fixed hand\-written instruction\.PersonaLinkis a self\-refining persona distiller; we report its first refinement round \(PersonaLinkr1\), with additional rounds in the appendix\.RAG\-kkretrieves the user’skkmost relevant past items and places them verbatim in the context as few\-shot exemplars, fork∈\{1,3,5,10,20\}k\\in\\\{1,3,5,10,20\\\}; this is the strongest training\-free baseline and the one that preserves fine\-grained instance\-level signal\. TheOraclerow in[Section˜VII](https://arxiv.org/html/2609.01615#S7)is a per\-item best\-of\-three over\{π⋆,seed,mismatch\}\\\{\\pi^\{\\star\},\\text\{seed\},\\text\{mismatch\}\\\}and is reported only as an upper bound on phrasing variance, not as a deployable method\.

### V\-DMetrics and Statistics

Point metrics are accuracy/macro\-F1 \(LaMP\-2\) and MAE/RMSE \(LaMP\-3\) over the200200test users\. We attach a95%95\\%bootstrap confidence interval to every metric\. Paired comparisons use the item as the unit: for classification we use McNemar’s test on per\-item correctness \(reporting the discordant countsAw/BwA\_\{\\text\{w\}\}/B\_\{\\text\{w\}\}\), and for regression a paired bootstrap \(10,00010\{,\}000resamples\) on per\-item absolute error, with the sign ofΔ\\DeltaMAE indicating which method is better\. We label any comparison withp≥0\.05p\\geq 0\.05asns\(not significant\)\. Rank correlations between meta\-validation and meta\-test use Spearmanρ\\rhoand Kendallτ\\tau\. For the collapse analysis we recompute the meta\-objective \([4](https://arxiv.org/html/2609.01615#S3.E4)\) and the deranged objective \([5](https://arxiv.org/html/2609.01615#S3.E5)\) on the3030meta\-validation users with a paired bootstrap on their difference\.

### V\-ESeed and Mismatch Controls

The two controls of[Section˜IV](https://arxiv.org/html/2609.01615#S4)are run end\-to\-end\. The seed control applies the frozenπ0\\pi\_\{0\}with no evolution\. The mismatch control evolvesπ\\piunder a fixed, deterministic derangement of the meta\-train user\-to\-support assignment \(no fixed points\), tests on the real users, and is otherwise identical toMuse\. We additionally sweep the support capK∈\{4,8,16,32\}K\\in\\\{4,8,16,32\\\}and the meta\-train size\|𝒰tr\|∈\{10,20,40\}\|\\mathcal\{U\}\_\{\\text\{tr\}\}\|\\in\\\{10,20,40\\\}, re\-evolving and re\-testing for each setting, to probe whether more support or more meta\-train users induces transfer\.

## VIResults

[Table˜IV](https://arxiv.org/html/2609.01615#S6.T4)reports the main comparison;[Figure˜1](https://arxiv.org/html/2609.01615#S0.F1)visualizes the headline rows\. We organize the findings as finding\-titled subsections\.[Table˜III](https://arxiv.org/html/2609.01615#S6.T3)attaches the95%95\\%bootstrap confidence intervals to the two headline metrics \(LaMP\-2accuracy andLaMP\-3MAE\); the intervals are wide relative to the inter\-method spread and overlap heavily across all persona methods andMusevariants, which is the interval\-level restatement of the null findings below—onlyRagonLaMP\-3sits in a visibly lower MAE band\.

TABLE III:Headline point estimates with95%95\\%bootstrap confidence intervals \(200200test users; intervals from the same item\-level resampling used for the paired tests\)\. ForLaMP\-2the metric is accuracy \(↑\\uparrow\); forLaMP\-3it is MAE \(↓\\downarrow\)\. The persona/Museintervals mutually overlap \(no learned\-adaptation separation\), whereasRag\-20onLaMP\-3clears the persona band\.TABLE IV:Main results over200200held\-out users per benchmark\.LaMP\-2: accuracy and macro\-F1 \(↑\\uparrow\)\.LaMP\-3: MAE and RMSE \(↓\\downarrow\)\.Musevariants are grouped at the bottom\.Boldmarks the best value in each column among deployable methods \(the per\-item Oracle is an upper bound on phrasing variance, not a method\)\. Key paired tests \([Sections˜VI](https://arxiv.org/html/2609.01615#S6)and[VII](https://arxiv.org/html/2609.01615#S7)\):Musevs\.Muse\-seedand vs\.Muse\-mismatchare*not significant*on both benchmarks;Musevs\.Rag\-20isnsonLaMP\-2butΔ​MAE=\+0\.175\\Delta\\text\{MAE\}\{=\}\{\+\}0\.175,95%95\\%CI\[\+0\.095,\+0\.260\]\[\{\+\}0\.095,\{\+\}0\.260\],p<0\.001p<0\.001onLaMP\-3\(retrieval wins regression\)\.Costis \#LLM calls per query\. Significance vs\.Muse\(π⋆\\pi^\{\\star\}\), per\-item paired tests \(McNemar for accuracy, paired bootstrap10410^\{4\}for MAE\):p∗⁣∗∗<0\.001\{\}^\{\*\*\*\}p<0\.001,p∗∗<0\.01\{\}^\{\*\*\}p<0\.01,p∗<0\.05\{\}^\{\*\}p<0\.05,ns\{\}^\{\\textsc\{ns\}\}not significant \(p≥0\.05p\\geq 0\.05\),=exact tie\.### VI\-AFinding 1: Meta\-Learning Does Not Beat Its Own Seed

The central comparison isMuse\(π⋆\)\(\\pi^\{\\star\}\)againstMuse\-seed, which differ*only*in whether the adaptation prompt was evolved\. Evolution does not help\. OnLaMP\-2,Musereaches accuracy0\.7650\.765versus the seed’s0\.7750\.775; the McNemar test on per\-item correctness gives discordant counts13/1513/15in the seed’s favor withp=0\.850p\{=\}0\.850\(ns\)\. OnLaMP\-3,Museattains MAE0\.4250\.425versus the seed’s0\.4000\.400, a paired\-bootstrapΔ​MAE=\+0\.025\\Delta\\text\{MAE\}\{=\}\{\+\}0\.025,95%95\\%CI\[−0\.045,\+0\.095\]\[\-0\.045,\+0\.095\],p=0\.523p\{=\}0\.523\(ns\)\. In both cases the un\-evolved seed is numerically*better*, and the difference is statistically indistinguishable from zero\. WhateverGepaextracted over the user population, it is not an improvement in per\-user adaptation that survives transfer to new users\.

### VI\-BFinding 2: Meta\-Learning Does Not Beat a Structure\-Broken Control

Muse\-mismatchmeta\-trains the shared prompt on*deliberately wrong*user–support pairs, so any genuine cross\-user signal in the objective has been destroyed by construction; ifMuselearned real adaptive structure it should beat this control\. It does not\. OnLaMP\-2,Muse0\.7650\.765versus mismatch0\.7400\.740isns\(McNemar16/1116/11,p=0\.441p\{=\}0\.441\); onLaMP\-3the two*tie*at MAE0\.4250\.425\(Δ​MAE=0\.000\\Delta\\text\{MAE\}\{=\}0\.000, CI\[−0\.075,\+0\.080\]\[\-0\.075,\+0\.080\],p=1\.000p\{=\}1\.000\)\. Meta\-training under a scrambled correspondence is as good as meta\-training under the real one—a direct symptom of the collapse we dissect in[Section˜VII](https://arxiv.org/html/2609.01615#S7)\.

### VI\-CFinding 3: Parity With Persona Baselines on Categorization

OnLaMP\-2, every distilled\-persona method—Muse, PAG, and PersonaLink—clusters in a narrow band around0\.740\.74–0\.780\.78accuracy and is mutually indistinguishable:Musevs\. PersonaLinkr1\(0\.7650\.765vs\.0\.7550\.755\) isns\(McNemar17/1517/15,p=0\.860p\{=\}0\.860\), andMusevs\. PAG \(0\.7650\.765vs\.0\.7450\.745\) isns\(18/1418/14,p=0\.596p\{=\}0\.596\)\. The meta\-learned prompt neither helps nor hurts relative to a hand\-written one\. All persona methods do clear the un\-personalized and random\-user floors \(NoPers0\.6450\.645, RandomUser0\.6650\.665\), so the personas*are*doing something; that “something” is simply not improved by meta\-learning the instruction\.

### VI\-DFinding 4: Retrieval Dominates on Regression

The one large, significant effect in the entire study runs*against*the persona paradigm\. OnLaMP\-3, retrieval\-augmented few\-shot prompts dominate: RAG\-20 attains MAE0\.2500\.250\(RMSE0\.5660\.566\) versusMuse’s0\.4250\.425, a pairedΔ​MAE=\+0\.175\\Delta\\text\{MAE\}\{=\}\{\+\}0\.175,95%95\\%CI\[\+0\.095,\+0\.260\]\[\{\+\}0\.095,\{\+\}0\.260\],p<0\.001p<0\.001\. Even modest retrieval budgets \(RAG\-3/5 at MAE≈0\.29\\approx\\\!0\.29\) beat all persona methods decisively\. The asymmetry withLaMP\-2, where retrieval is only marginally and non\-significantly ahead \(RAG\-200\.7900\.790vs\.Muse0\.7650\.765,ns\), is informative: distilling a user into a short natural\-language persona discards the fine\-grained rating calibration that raw few\-shot exemplars preserve\. We report this prominently rather than minimize it; on rating, the representation that keeps instance\-level signal wins, independently of whether the adaptation policy is learned\.[Figure˜3](https://arxiv.org/html/2609.01615#S6.F3)plots the retrieval ladder against the distilled\-persona band on both tasks and makes the asymmetry visual: onLaMP\-2the RAG curve threads straight through the persona band \(a tie\), whereas onLaMP\-3it pulls clear of the band for everyk≥3k\\geq 3and reaches its widest, significant gap atk=20k\{=\}20\.[Table˜V](https://arxiv.org/html/2609.01615#S6.T5)consolidates the head\-to\-head paired tests that back the four findings—eachΔ\\Delta, its95%95\\%confidence interval, and itspp\-value—in one place, so the reader can verify at a glance that every persona\-vs\-Musecomparison isnswhile onlyRag\-20onLaMP\-3clears significance\.

![Refer to caption](https://arxiv.org/html/2609.01615v1/x2.png)Figure 3:Retrieval budget versus the distilled\-persona band \(numbers from[Table˜IV](https://arxiv.org/html/2609.01615#S6.T4)\)\.*Left:*LaMP\-2accuracy \(↑\\uparrow\); the retrieval\-augmented few\-shot curve RAG\-kkrises withkkbut stays inside the shaded persona band \(Muse/seed/mismatch/PAG/PersonaLink span0\.7400\.740–0\.7750\.775\), so retrieval and personas are statistically tied \(RAG\-200\.7900\.790vs\.Muse0\.7650\.765,ns\)\.*Right:*LaMP\-3MAE \(↓\\downarrow, plotted so better is up\); even RAG\-3/5 \(MAE≈0\.29\\text\{MAE\}\\approx 0\.29\) leave the persona band, and RAG\-20 \(0\.2500\.250\) opens a significantΔ​MAE=\+0\.175\\Delta\\text\{MAE\}\{=\}\{\+\}0\.175gap overMuse\(0\.4250\.425,p<0\.001p<0\.001\)\. Distilling a user into a short persona discards the instance\-level rating calibration that raw exemplars preserve; the effect is large only on the ordinal task\.TABLE V:Consolidated paired significance tests behind[Section˜VI](https://arxiv.org/html/2609.01615#S6)\(Findings 1–4\)\. Each row is a comparison against the listed reference;Δ\\Deltais \(reference−\-Muse\) for accuracy and \(Muse−\-reference\) for MAE, signed so a positive value favorsMuse’s competitor\. Classification rows report McNemar discordant countsAw/BwA\_\{\\text\{w\}\}/B\_\{\\text\{w\}\}on per\-item correctness; regression rows report a paired bootstrap \(10410^\{4\}resamples\) on per\-item absolute error\. The only comparison that clearsp<0\.05p<0\.05isRag\-20onLaMP\-3; everyMuse\-vs\-persona andMuse\-vs\-control comparison isns\.

## VIIAnalysis: Meta\-Objective Collapse

The results of[Section˜VI](https://arxiv.org/html/2609.01615#S6)are not four separate nulls; they are consequences of one mechanism\. We establish*meta\-objective collapse*through three mechanistic legs—\(i\) no transfer over the seed, \(ii\) a structure\-blind objective, and \(iii\) meta\-overfitting at small population scale—and then decompose the residual test behavior with an invariance/oracle analysis and a semantic diff of the evolved prompts\.

### VII\-ALeg I: No Transfer —π⋆\\pi^\{\\star\}Behaves Like the Seed

If evolution had learned transferable structure,π⋆\\pi^\{\\star\}would behave differently—and better—thanπ0\\pi\_\{0\}on held\-out users\. Instead the two are behaviorally near\-identical\.[Figure˜6](https://arxiv.org/html/2609.01615#S7.F6)\(left\) decomposes the200200test users by how the prediction moves across\{π⋆,seed,mismatch\}\\\{\\pi^\{\\star\},\\text\{seed\},\\text\{mismatch\}\\\}\. OnLaMP\-2,80\.0%80\.0\\%of users are*phrasing\-invariant*\(132132invariant\-correct,2828invariant\-wrong\) and only4040\(20%20\\%\) ever “swing”; onLaMP\-3,72\.0%72\.0\\%are invariant \(9797/4747\) with5656\(28%28\\%\) swinging\. Among the swing users—the only ones the choice of prompt can affect—π⋆\\pi^\{\\star\}and the seed split essentially evenly: McNemar13/1513/15,p=0\.85p\{=\}0\.85onLaMP\-2and13/1913/19,p=0\.38p\{=\}0\.38onLaMP\-3, bothns, with the seed numerically ahead\. The evolved prompt is not a different adaptation policy; it is a paraphrase of the seed that the frozen backbone treats almost identically\.

![Refer to caption](https://arxiv.org/html/2609.01615v1/x3.png)Figure 4:The money figure: the meta\-objective is structure\-blind\.Recomputed meta\-validation objective on the3030meta\-val users for the evolved prompt with*aligned*\(real\) supports versus a*mismatched*\(deranged\) support assignment,[Equation˜4](https://arxiv.org/html/2609.01615#S3.E4)vs\. \([5](https://arxiv.org/html/2609.01615#S3.E5)\)\. Destroying the user–support correspondence does*not*lower the objective: the aligned score is \(numerically\) at least the mismatch score on*both*benchmarks and neither gap is significant—0\.8330\.833aligned vs\.0\.8000\.800mismatch,Δ=\+0\.033\\Delta\{=\}\{\+\}0\.033,95%95\\%CI\[−0\.050,\+0\.117\]\[\-0\.050,\{\+\}0\.117\],p=0\.55p\{=\}0\.55\(ns\) onLaMP\-2, and0\.8960\.896vs\.0\.8830\.883,Δ=\+0\.013\\Delta\{=\}\{\+\}0\.013,\[−0\.029,\+0\.054\]\[\-0\.029,\{\+\}0\.054\],p=0\.62p\{=\}0\.62\(ns\) onLaMP\-3\(the saved selection scores agree:0\.8460\.846/0\.8320\.832and0\.8910\.891/0\.8940\.894\)\. An objective statistically invariant to whether the correspondence is real cannot be optimized into transferable adaptation\.
### VII\-BLeg II: A Structure\-Blind Objective \(the Money Figure\)

The decisive evidence is in the objective itself\. We recompute, on the3030meta\-validation users, the aligned meta\-objective \([4](https://arxiv.org/html/2609.01615#S3.E4)\) and its deranged counterpart \([5](https://arxiv.org/html/2609.01615#S3.E5)\) for the evolved prompt, and compare them with a paired bootstrap on the per\-user difference \([Figure˜4](https://arxiv.org/html/2609.01615#S7.F4)\)\. The two are statistically indistinguishable, and on*both*benchmarks the aligned objective is \(numerically\) at least as high as the mismatched one\. OnLaMP\-2the aligned score is0\.8330\.833versus mismatch0\.8000\.800, a difference ofΔ=\+0\.033\\Delta\{=\}\{\+\}0\.033,95%95\\%CI\[−0\.050,\+0\.117\]\[\-0\.050,\+0\.117\],p=0\.555p\{=\}0\.555\(ns\); onLaMP\-3the aligned score is0\.8960\.896versus mismatch0\.8830\.883,Δ=\+0\.013\\Delta\{=\}\{\+\}0\.013,\[−0\.029,\+0\.054\]\[\-0\.029,\+0\.054\],p=0\.622p\{=\}0\.622\(ns\)\. TheGepaselection\-time scores tell the same story \(the saved selection scores agree:0\.8460\.846/0\.8320\.832onLaMP\-2and0\.8910\.891/0\.8940\.894onLaMP\-3\): aligned≈\\approxmismatch, and the objective is blind to the correspondence\. This is meta\-objective collapse in the precise sense of[Equation˜6](https://arxiv.org/html/2609.01615#S3.E6): the quantity the optimizer climbs does not distinguish a user’s persona built from their own history from one built from a stranger’s\. An optimizer maximizing such a quantity cannot, even in principle, acquire cross\-user adaptive structure—it can only move along directions to which the objective is sensitive, namely generic phrasing\. This is the operational meaning of*proxy\-objective misalignment*: the optimized proxy and the intended goal \(transferable adaptation\) have come apart, and the derangement control measures exactly the size of the gap \(here, indistinguishable from zero\)\.

### VII\-CLeg III: Meta\-Overfitting at Small Population Scale

The little that does separate candidate prompts is not predictive of held\-out performance—it is selection\-set overfitting\.[Figure˜5](https://arxiv.org/html/2609.01615#S7.F5)plots meta\-validation against meta\-test for nineLaMP\-2variants \(the seed, the mismatch control, and theKK\- and meta\-train\-size sweeps\)\. Three facts stand out\. First, there is a uniform*optimistic gap*: meta\-test trails meta\-validation by≈\+0\.094\\approx\\\!\+0\.094on average \(every variant lies below they=xy\{=\}xline\), exactly the over\-optimism expected when a small validation population is used to select among prompts\. Second, meta\-validation does*not*significantly rank meta\-test: Spearmanρ=\+0\.47\\rho\{=\}\+0\.47\(p=0\.20p\{=\}0\.20\) and Kendallτ=\+0\.38\\tau\{=\}\+0\.38\(p=0\.17p\{=\}0\.17\), bothnsand positive \(we explicitly do*not*claim a negative correlation\); strikingly, the best meta\-validation variant \(K=16K\{=\}16at0\.8720\.872\) yields a*below\-median*test accuracy of0\.7550\.755\. Third, the sweeps are flat and non\-monotone: accuracy wobbles inside0\.7450\.745/0\.7700\.770/0\.7550\.755/0\.7500\.750acrossK∈\{4,8,16,32\}K\\in\\\{4,8,16,32\\\}, and0\.7400\.740/0\.7750\.775/0\.7550\.755across\|𝒰tr\|∈\{10,20,40\}\|\\mathcal\{U\}\_\{\\text\{tr\}\}\|\\in\\\{10,20,40\\\}, with no trend that survives the±0\.06\\pm 0\.06McNemar noise band\. More support and more meta\-train users do not buy transfer; there is no scaling signal to ride\.

![Refer to caption](https://arxiv.org/html/2609.01615v1/x4.png)Figure 5:Meta\-overfitting\.*Left:*meta\-validation versus meta\-test accuracy for nineLaMP\-2variants; every point sits belowy=xy\{=\}x\(mean optimistic gap\+0\.094\+0\.094\), the rank correlation is positive but not significant \(ρ=\+0\.47\\rho\{=\}\+0\.47,p=0\.20p\{=\}0\.20;τ=\+0\.38\\tau\{=\}\+0\.38,p=0\.17p\{=\}0\.17\), and the highest meta\-val point \(K=16K\{=\}16\) has below\-median test accuracy\.*Right:*LaMP\-3meta\-val objective versus the held\-out1−MAE/41\{\-\}\\text\{MAE\}/4reward forπ⋆\\pi^\{\\star\}and the mismatch control are flat and near\-equal\. The selection objective does not predict held\-out generalization\.
### VII\-DInvariance and Oracle Decomposition

[Figure˜6](https://arxiv.org/html/2609.01615#S7.F6)\(right\) bounds what*any*choice among\{π⋆,seed,mismatch\}\\\{\\pi^\{\\star\},\\text\{seed\},\\text\{mismatch\}\\\}could buy\. A per\-item Oracle that picks the best of the three reaches accuracy0\.8600\.860onLaMP\-2and MAE0\.2450\.245onLaMP\-3, well above each individual method\. It is tempting to read this headroom as latent adaptation waiting to be unlocked\. It is not\. The*mismatch*prompt—whose supports were scrambled—is*in*the Oracle set, so the Oracle is rewarded whenever a scrambled\-support persona happens to phrase the answer in a way the backbone gets right\. The headroom is therefore*phrasing variance*across paraphrases of essentially the same policy, not learned cross\-user adaptation; no selector with access only to a user’s own support could realize it, because the three prompts do not encode different adaptive behavior, only different wordings\. The Oracle bound thus reinforces, rather than softens, the collapse diagnosis\.

![Refer to caption](https://arxiv.org/html/2609.01615v1/x5.png)Figure 6:Invariance and oracle\.*Left:*fraction of the200200test users that are invariant\-correct, invariant\-wrong, or swing across\{π⋆,seed,mismatch\}\\\{\\pi^\{\\star\},\\text\{seed\},\\text\{mismatch\}\\\};80%80\\%/72%72\\%are phrasing\-invariant onLaMP\-2/LaMP\-3, and among swing usersπ⋆\\pi^\{\\star\}does not beat the seed \(McNemar13/1513/15,p=0\.85p\{=\}0\.85;13/1913/19,p=0\.38p\{=\}0\.38; bothns\)\.*Right:*a per\-item Oracle over the three prompts reaches0\.8600\.860accuracy /0\.2450\.245MAE, but since the mismatch prompt is in the set, this headroom is phrasing variance, not learned adaptation\.
### VII\-EWhat the Evolved Prompt Actually Changed

A semantic analysis of the prompts corroborates the mechanism\. Embedding the prompts with a sentence encoder\[[44](https://arxiv.org/html/2609.01615#bib.bib31)\], the evolvedπ⋆\\pi^\{\\star\}is only moderately similar to the seed \(cosine0\.5130\.513onLaMP\-2,0\.5390\.539onLaMP\-3\), confirming thatGepadid rewrite the text substantially\. But the rewrites do*not*organize by adaptive content\. Across all evolved prompts \(the two benchmarks, the mismatch variants, and the sweeps\), the prompts cluster by*benchmark*, not by aligned\-versus\-mismatch: mean within\-benchmark cosine \(0\.7230\.723onLaMP\-2,0\.7290\.729onLaMP\-3\) exceeds the cross\-benchmark mean \(0\.6650\.665\)\. If the evolution had discovered genuine adaptive structure, the aligned and mismatched prompts would differ in kind; instead they are as similar as any two same\-benchmark prompts\. Reading the diff confirms the impression: the seed\-to\-π⋆\\pi^\{\\star\}edit reformats free prose into a structured JSON recipe and adds generic “polish” lines—“focus on stability over short\-term fluctuations”, “select examples that best represent each category”—i\.e\., better instruction formatting, not a new adaptation mechanism\. The evolution found the directions the objective rewards \(phrasing, formatting\) and could not find the direction it does not reward \(cross\-user adaptation\), exactly as collapse predicts\.[Table˜VI](https://arxiv.org/html/2609.01615#S7.T6)lays out the cosine geometry that grounds this reading: the seed\-to\-π⋆\\pi^\{\\star\}similarity is only moderate \(the rewrite is real\), yet the aligned and mismatch prompts sit at within\-benchmark similarity, and the within\-benchmark similarity exceeds the cross\-benchmark similarity on both tasks\. The prompts organize by*benchmark*, not by whether their training supports were aligned—the signature of a benchmark\-shaped phrasing edit rather than a discovered adaptive mechanism\.

TABLE VI:Semantic geometry of the evolved prompts \(cosine similarity of sentence\-encoder embeddings\[[44](https://arxiv.org/html/2609.01615#bib.bib31)\]\)\. The seed\-to\-π⋆\\pi^\{\\star\}similarity is only moderate \(substantial rewrite\), but the aligned\-vs\-mismatch similarity is at the within\-benchmark level, and within\-benchmark exceeds cross\-benchmark\. Prompts cluster by benchmark, not by whether their meta\-training supports were aligned—consistent with a phrasing/formatting edit, not a learned adaptation mechanism\.

## VIIIDiscussion

### VIII\-AA Reusable Evaluation Protocol

The most transferable product of this study is methodological\. The same three controls that diagnose collapse here can pre\-empt false positives anywhere prompt\-space methods claim to “learn to adapt” to users or tasks\.\(1\) Seed control:run the identical pipeline with the un\-optimized seed prompt; if the optimized prompt does not beat it, the optimization bought instruction polish, not learning\.\(2\) Wrong\-support \(derangement\) control:recompute the selection objective with the support\-to\-user correspondence scrambled; if the objective does not drop, it is structure\-blind and cannot encode adaptation\.\(3\) Invariance/oracle decomposition:measure how many test instances are invariant to the choice of prompt, and check whether any oracle headroom comes from genuinely different policies or merely from paraphrase variance \(which is exposed by including a structure\-broken variant in the oracle set\)\. Each control is cheap—it reuses the existing pipeline—and each isolates a distinct confound \(prior instruction quality, objective structure, phrasing variance\)\. We recommend reporting all three before attributing gains to learned adaptation\. Their absence is, we suspect, why prompt\-space personalization results can look stronger than they are: instruction polish and small\-population validation overfit are easy to mistake for transfer\.

### VIII\-BWhen Prompt\-Space Meta\-Learning Can and Cannot Work

Our negative result is specific, and its boundary is informative\. The failure is not that frozen\-LLM personalization is hopeless—the persona methods clearly beat the un\-personalized floor, and retrieval wins outright on rating—but that*meta\-learning a shared adaptation instruction against an aggregate population objective*does not add transferable structure on top\. The diagnosis points directly at the missing ingredient:*grounded, per\-instance feedback*\. Test\-time adaptation methods that succeed do so because their objective is tied to a checkable signal on each instance—a verifier, a self\-consistency vote, a tool execution\. Grounded agentic adaptation, where each step is verified against per\-call execution feedback \([Section˜II](https://arxiv.org/html/2609.01615#S2)\), improves precisely because its reward is instance\-level and structure\-bearing\.Muse’s meta\-objective is the opposite: a population average with no per\-instance grounding, which is why scrambling the correspondence leaves it unchanged\. We therefore expect prompt\-space meta\-learning to transfer when, and roughly only when, the meta\-objective can be made sensitive to the user–instance correspondence—for example by per\-user verifiable rewards, contrastive objectives that explicitly penalize wrong\-user personas, or richer support sets that make the right user distinguishable\. Absent such grounding, the objective collapses to phrasing and a hand\-written seed is as good as evolution\. This also reconciles our finding with the implicit\-meta\-learning view of ICL: the frozen backbone already adapts to a user’s in\-context exemplars, so an explicit, ungrounded prompt\-space objective has little additional structure to grab\.

### VIII\-CThreats to Validity and Limitations

We delimit the claims carefully\.*\(i\) Backbone\.*We use a single backbone family \(Qwen3\-30B\-A3B\); a larger or differently\-tuned model might extract signal a smaller one cannot, though the structure\-blindness of the objective \([Equation˜6](https://arxiv.org/html/2609.01615#S3.E6)\) is a property of the*objective*, not the model, and would have to be broken by changing the objective, not merely the backbone\.*\(ii\) Benchmarks\.*We study two LaMP tasks \(categorization and rating\); other personalization tasks—generation, dialogue style, long\-form preference—may behave differently, and we do not claim universality\.*\(iii\) Prompt class\.*Musemeta\-learns a*natural\-language*adaptation prompt evolved byGepa; soft\-prompt or gradient\-based meta\-learners, or contrastive objectives, are outside our scope and are the natural next test of whether grounding the objective restores transfer\.*\(iv\) Determinism\.*The serving stack is not bitwise\-deterministic, so we use the item as the unit of analysis and rely on paired tests; we reportnswhere warranted and never convert a non\-significant difference into a claim\.*\(v\) Scope of the positive claim\.*The defensible positive contribution is the diagnostic protocol and the collapse mechanism, not a new state\-of\-the\-art personalizer\. We present it as such\.

## IXConclusion

We asked whether prompt\-space meta\-learning genuinely learns to adapt a frozen LLM across users, and answered, with controls, that on two standard personalization benchmarks it does not\.Muse—a clean instantiation ofGepa\-style shared\-prompt evolution as a cross\-user meta\-learner—fails to beat its own un\-evolved seed prompt, fails to beat a structure\-broken control that meta\-trains on scrambled user–support pairs, ties the strongest persona baselines, and is decisively dominated by plain retrieval on rating\. The unifying cause is*meta\-objective collapse*: the aggregate meta\-validation objective is statistically invariant to whether the user–support correspondence is real, so it carries no exploitable cross\-user adaptive signal, and the optimizer can only buy generic instruction polish \(already present in a hand\-written seed\) plus overfitting of a small validation population\. We quantified each leg—no transfer over the seed, a structure\-blind objective, an≈9\\approx\\\!9\-point optimistic gap with non\-predictive selection, and80%80\\%/72%72\\%phrasing invariance with oracle headroom attributable to paraphrase variance—and distilled a reusable evaluation protocol \(seed, wrong\-support, and invariance/oracle controls\) so that the community can separate learned adaptation from instruction polish and validation overfit\. The constructive reading is a hypothesis with teeth: what prompt\-space meta\-learning lacks is grounded, per\-instance feedback, and supplying it is the most promising route to making the meta\-objective informative again\.

## Appendix AFull Sweep Numbers

[Table˜VII](https://arxiv.org/html/2609.01615#A1.T7)reports the support\-size \(KK\) and meta\-train\-size sweeps onLaMP\-2that underlie[Figure˜5](https://arxiv.org/html/2609.01615#S7.F5), including each variant’sGepa\-selection meta\-validation score, held\-out test accuracy and macro\-F1, and the optimistic gap\. Higher meta\-validation does not imply higher test accuracy; the highest meta\-val variant \(K=16K\{=\}16,0\.8720\.872\) has below\-median test accuracy \(0\.7550\.755\)\.

TABLE VII:LaMP\-2support\-size and meta\-train\-size sweeps \(200200test users\)\. “meta\-val” is theGepa\-selection aggregate on the3030meta\-val users \(the optimized quantity\), not the test metric\. “gap”==meta\-val−\-test accuracy\.VariantKK\|𝒰tr\|\|\\mathcal\{U\}\_\{\\text\{tr\}\}\|meta\-valAcc↑\\uparrowmacro\-F1gapMuse\(π⋆\\pi^\{\\star\}\)24400\.8460\.7650\.5010\.081Muse\-mismatch24400\.8320\.7400\.4330\.092Muse\-KK44400\.8300\.7450\.4440\.085Muse\-KK88400\.8500\.7700\.5130\.080Muse\-KK1616400\.8720\.7550\.4870\.117Muse\-KK3232400\.8590\.7500\.4950\.109Muse\-mt1024100\.8470\.7400\.4480\.107Muse\-mt2024200\.8590\.7750\.5080\.084Muse\-mt4024400\.8430\.7550\.5060\.088
## Appendix BHyperparameters and Implementation

[Table˜VIII](https://arxiv.org/html/2609.01615#A2.T8)lists the configuration\. The reflection, inner\-distillation, and answer calls all use the same frozen backbone at temperature0\. The mismatch control uses a fixed, seeded derangement \(no fixed points\) of the meta\-train user\-to\-support assignment\. The benchmark scoring is deterministic; no LLM judge is used\. Ablation prompts are saved to label\-specific sidecar files so that the canonicalπ⋆\\pi^\{\\star\}and its meta\-test cells are never overwritten\.

TABLE VIII:Key hyperparameters and protocol settings\.
## Appendix CEvolved Prompt Texts

For reproducibility we summarize the seed and evolved prompts\. The seedπ0\\pi\_\{0\}is a single free\-prose instruction directing the model to read a user’s labeled history and build a persona that captures stable preferences, label\-driving cues, systematic tendencies, a few general decision rules, and representative exemplars, emphasizing the user’s*decision policy*over memorization\. The evolvedLaMP\-2promptπ⋆\\pi^\{\\star\}reformats this into a structured JSON recipe with fields for preferences \(“extract core topics and trends…focusing on stability over short\-term fluctuations”\), categorization cues \(“identify key labels and surface specific cues…be strict but fair”\), exemplar selection, decision rules, and tie disambiguation\. The evolvedLaMP\-3prompt similarly enumerates fields for stable preferences, noise filtering, label\-specific cues, systematic strictness/leniency, decision rules, exemplars, and near\-tie disambiguation\. The mismatch\-trained prompts are of the same character \(generic, benchmark\-shaped recipes\)\. Consistent with[Section˜VII](https://arxiv.org/html/2609.01615#S7), the edits are formatting and generic guidance, not benchmark\-transcending adaptive structure: the evolved prompts cluster by benchmark \(within\-benchmark cosine0\.7230\.723/0\.7290\.729\>\>cross\-benchmark0\.6650\.665\) rather than by whether their training supports were aligned, and cosine\(π0,π⋆\)\(\\pi\_\{0\},\\pi^\{\\star\}\)is only0\.5130\.513/0\.5390\.539\.

## References

- \[1\]L\. A\. Agrawal, S\. Tan, D\. Soylu, N\. Ziems, R\. Khare, K\. Opsahl\-Ong, A\. Singhvi, H\. Shandilya, M\. J\. Ryan, M\. Jiang,et al\.\(2025\)GEPA: reflective prompt evolution can outperform reinforcement learning\.arXiv preprint arXiv:2507\.19457\.Cited by:[§I](https://arxiv.org/html/2609.01615#S1.p1.1),[§I](https://arxiv.org/html/2609.01615#S1.p2.3),[§II\-D](https://arxiv.org/html/2609.01615#S2.SS4.p1.1),[§IV\-B](https://arxiv.org/html/2609.01615#S4.SS2.p1.7)\.
- \[2\]\(2023\)What learning algorithm is in\-context learning? investigations with linear models\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2211\.15661Cited by:[§II\-C](https://arxiv.org/html/2609.01615#S2.SS3.p1.1)\.
- \[3\]M\. Andrychowicz, M\. Denil, S\. Gómez, M\. W\. Hoffman, D\. Pfau, T\. Schaul, B\. Shillingford, and N\. de Freitas\(2016\)Learning to learn by gradient descent by gradient descent\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:1606\.04474Cited by:[§II\-B](https://arxiv.org/html/2609.01615#S2.SS2.p1.1)\.
- \[4\]A\. Antoniou, H\. Edwards, and A\. Storkey\(2019\)How to train your MAML\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:1810\.09502Cited by:[§II\-B](https://arxiv.org/html/2609.01615#S2.SS2.p1.1)\.
- \[5\]A\. Asai, Z\. Wu, Y\. Wang, A\. Sil, and H\. Hajishirzi\(2024\)Self\-RAG: learning to retrieve, generate, and critique through self\-reflection\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2310\.11511Cited by:[§II\-F](https://arxiv.org/html/2609.01615#S2.SS6.p1.1)\.
- \[6\]S\. Borgeaud, A\. Mensch, J\. Hoffmann, T\. Cai, E\. Rutherford, K\. Millican, G\. van den Driessche, J\. Lespiau, B\. Damoc, A\. Clark,et al\.\(2022\)Improving language models by retrieving from trillions of tokens\.InInternational Conference on Machine Learning \(ICML\),Note:arXiv:2112\.04426Cited by:[§II\-F](https://arxiv.org/html/2609.01615#S2.SS6.p1.1)\.
- \[7\]J\. Chen, X\. Wang, R\. Xu, S\. Yuan, Y\. Zhang, W\. Shi, J\. Xie, S\. Li, R\. Yang, T\. Zhu, A\. Chen, N\. Li, L\. Chen, C\. Hu, S\. Wu, S\. Ren, Z\. Fu, and Y\. Xiao\(2024\)From persona to personalization: a survey on role\-playing language agents\.Transactions on Machine Learning Research \(TMLR\)\.Note:arXiv:2404\.18231Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[8\]C\. Fernando, D\. Banarse, H\. Michalewski, S\. Osindero, and T\. Rocktäschel\(2023\)Promptbreeder: self\-referential self\-improvement via prompt evolution\.arXiv preprint\.Note:arXiv:2309\.16797Cited by:[§II\-D](https://arxiv.org/html/2609.01615#S2.SS4.p1.1)\.
- \[9\]C\. Finn, P\. Abbeel, and S\. Levine\(2017\)Model\-agnostic meta\-learning for fast adaptation of deep networks\.InInternational Conference on Machine Learning \(ICML\),Note:arXiv:1703\.03400Cited by:[§I](https://arxiv.org/html/2609.01615#S1.p1.1),[§II\-B](https://arxiv.org/html/2609.01615#S2.SS2.p1.1)\.
- \[10\]T\. Formal, B\. Piwowarski, and S\. Clinchant\(2021\)SPLADE: sparse lexical and expansion model for first stage ranking\.InProceedings of the 44th International ACM SIGIR Conference \(SIGIR\),pp\. 2288–2292\.Note:arXiv:2107\.05720Cited by:[§II\-F](https://arxiv.org/html/2609.01615#S2.SS6.p1.1)\.
- \[11\]Z\. Gou, Z\. Shao, Y\. Gong, Y\. Shen, Y\. Yang, N\. Duan, and W\. Chen\(2023\)CRITIC: large language models can self\-correct with tool\-interactive critiquing\.arXiv preprint\.Note:arXiv:2305\.11738Cited by:[§II\-D](https://arxiv.org/html/2609.01615#S2.SS4.p1.1)\.
- \[12\]J\. Gu, Y\. Wang, Y\. Chen, V\. O\. K\. Li, and K\. Cho\(2018\)Meta\-learning for low\-resource neural machine translation\.InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),Note:arXiv:1808\.08437Cited by:[§II\-B](https://arxiv.org/html/2609.01615#S2.SS2.p1.1)\.
- \[13\]K\. Guu, K\. Lee, Z\. Tung, P\. Pasupat, and M\. Chang\(2020\)REALM: retrieval\-augmented language model pre\-training\.InInternational Conference on Machine Learning \(ICML\),Note:arXiv:2002\.08909Cited by:[§II\-F](https://arxiv.org/html/2609.01615#S2.SS6.p1.1)\.
- \[14\]X\. Han, B\. Guo, Y\. Jung, B\. Yao, Y\. Zhang, X\. Liu, and C\. Guo\(2023\)PersonaPKT: building personalized dialogue agents via parameter\-efficient knowledge transfer\.InProceedings of the Fourth Workshop on Simple and Efficient Natural Language Processing \(SustaiNLP\),Note:arXiv:2306\.08126Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[15\]T\. M\. Hospedales, A\. Antoniou, P\. Micaelli, and A\. J\. Storkey\(2022\)Meta\-learning in neural networks: a survey\.IEEE Transactions on Pattern Analysis and Machine Intelligence \(TPAMI\)44\(9\),pp\. 5149–5169\.Note:arXiv:2004\.05439Cited by:[§I](https://arxiv.org/html/2609.01615#S1.p1.1),[§II\-B](https://arxiv.org/html/2609.01615#S2.SS2.p1.1)\.
- \[16\]J\. Huang, S\. S\. Gu, L\. Hou, Y\. Wu, X\. Wang, H\. Yu, and J\. Han\(2022\)Large language models can self\-improve\.arXiv preprint\.Note:arXiv:2210\.11610Cited by:[§II\-D](https://arxiv.org/html/2609.01615#S2.SS4.p1.1)\.
- \[17\]G\. Izacard, M\. Caron, L\. Hosseini, S\. Riedel, P\. Bojanowski, A\. Joulin, and E\. Grave\(2022\)Unsupervised dense information retrieval with contrastive learning\.Transactions on Machine Learning Research \(TMLR\)\.Note:arXiv:2112\.09118Cited by:[§II\-F](https://arxiv.org/html/2609.01615#S2.SS6.p1.1)\.
- \[18\]G\. Izacard and E\. Grave\(2021\)Leveraging passage retrieval with generative models for open domain question answering\.InProceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics \(EACL\),pp\. 874–880\.Note:arXiv:2007\.01282Cited by:[§II\-F](https://arxiv.org/html/2609.01615#S2.SS6.p1.1)\.
- \[19\]K\. Javed and M\. White\(2019\)Meta\-learning representations for continual learning\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:1905\.12588Cited by:[§II\-B](https://arxiv.org/html/2609.01615#S2.SS2.p1.1)\.
- \[20\]V\. Karpukhin, B\. Oğuz, S\. Min, P\. Lewis, L\. Wu, S\. Edunov, D\. Chen, and W\. Yih\(2020\)Dense passage retrieval for open\-domain question answering\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),pp\. 6769–6781\.Note:arXiv:2004\.04906Cited by:[§II\-F](https://arxiv.org/html/2609.01615#S2.SS6.p1.1)\.
- \[21\]O\. Khattab, K\. Santhanam, X\. L\. Li, D\. Hall, P\. Liang, C\. Potts, and M\. Zaharia\(2022\)Demonstrate\-search\-predict: composing retrieval and language models for knowledge\-intensive NLP\.arXiv preprint\.Note:arXiv:2212\.14024Cited by:[§II\-F](https://arxiv.org/html/2609.01615#S2.SS6.p1.1)\.
- \[22\]O\. Khattab, A\. Singhvi, P\. Maheshwari, Z\. Zhang,et al\.\(2024\)DSPy: compiling declarative language model calls into self\-improving pipelines\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§II\-D](https://arxiv.org/html/2609.01615#S2.SS4.p1.1)\.
- \[23\]O\. Khattab and M\. Zaharia\(2020\)ColBERT: efficient and effective passage search via contextualized late interaction over BERT\.InProceedings of the 43rd International ACM SIGIR Conference \(SIGIR\),pp\. 39–48\.Note:arXiv:2004\.12832Cited by:[§II\-F](https://arxiv.org/html/2609.01615#S2.SS6.p1.1)\.
- \[24\]I\. Kumar, S\. Viswanathan, S\. Yerra, A\. Salemi, R\. A\. Rossi, F\. Dernoncourt, H\. Deilamsalehy, X\. Zhang, S\. Agarwal, N\. L\. Burhanuddin,et al\.\(2024\)LongLaMP: a benchmark for personalized long\-form text generation\.arXiv preprint arXiv:2407\.11016\.Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[25\]H\. Lee, J\. Im, S\. Jang, H\. Cho, and S\. Chung\(2019\)MeLU: meta\-learned user preference estimator for cold\-start recommendation\.InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining \(KDD\),Note:arXiv:1908\.00413Cited by:[§II\-B](https://arxiv.org/html/2609.01615#S2.SS2.p1.1)\.
- \[26\]P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni, V\. Karpukhin, N\. Goyal, H\. Küttler, M\. Lewis, W\. Yih, T\. Rocktäschel, S\. Riedel, and D\. Kiela\(2020\)Retrieval\-augmented generation for knowledge\-intensive NLP tasks\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:2005\.11401Cited by:[§II\-F](https://arxiv.org/html/2609.01615#S2.SS6.p1.1)\.
- \[27\]Z\. Li, F\. Zhou, F\. Chen, and H\. Li\(2017\)Meta\-SGD: learning to learn quickly for few\-shot learning\.arXiv preprint\.Note:arXiv:1707\.09835Cited by:[§II\-B](https://arxiv.org/html/2609.01615#S2.SS2.p1.1)\.
- \[28\]J\. Liu, D\. Shen, Y\. Zhang, B\. Dolan, L\. Carin, and W\. Chen\(2022\)What makes good in\-context examples for GPT\-3?\.InProceedings of Deep Learning Inside Out \(DeeLIO\),pp\. 100–114\.Note:arXiv:2101\.06804Cited by:[§II\-F](https://arxiv.org/html/2609.01615#S2.SS6.p1.1)\.
- \[29\]J\. Liu, Z\. Qiu, Z\. Li, Q\. Dai, J\. Zhu, M\. Hu, M\. Yang, T\. Chua, and I\. King\(2025\)A survey of personalized large language models: progress and future directions\.arXiv preprint\.Note:arXiv:2502\.11528Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[30\]J\. Liu, Y\. Zhu, S\. Wang, X\. Wei, E\. Min, Y\. Lu, S\. Wang, D\. Yin, and Z\. Dou\(2025\)LLMs\+\+persona\-plug==personalized LLMs\.arXiv preprint arXiv:2409\.11901\.Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[31\]A\. Madaan, N\. Tandon, P\. Gupta, S\. Hallinan, L\. Gao,et al\.\(2023\)Self\-refine: iterative refinement with self\-feedback\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Cited by:[§II\-D](https://arxiv.org/html/2609.01615#S2.SS4.p1.1)\.
- \[32\]S\. Min, M\. Lewis, L\. Zettlemoyer, and H\. Hajishirzi\(2022\)MetaICL: learning to learn in context\.InProceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics \(NAACL\),Note:arXiv:2110\.15943Cited by:[§II\-C](https://arxiv.org/html/2609.01615#S2.SS3.p1.1)\.
- \[33\]N\. Mishra, M\. Rohaninejad, X\. Chen, and P\. Abbeel\(2018\)A simple neural attentive meta\-learner\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:1707\.03141Cited by:[§II\-B](https://arxiv.org/html/2609.01615#S2.SS2.p1.1)\.
- \[34\]E\. Mitchell, C\. Lin, A\. Bosselut, C\. Finn, and C\. D\. Manning\(2022\)Fast model editing at scale\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2110\.11309Cited by:[§II\-E](https://arxiv.org/html/2609.01615#S2.SS5.p1.1)\.
- \[35\]N\. Muennighoff, Z\. Yang, W\. Shi, X\. L\. Li, L\. Fei\-Fei, H\. Hajishirzi, L\. Zettlemoyer, P\. Liang, E\. Candès, and T\. Hashimoto\(2025\)S1: simple test\-time scaling\.arXiv preprint\.Note:arXiv:2501\.19393Cited by:[§II\-E](https://arxiv.org/html/2609.01615#S2.SS5.p1.1)\.
- \[36\]S\. Mysore, Z\. Lu, M\. Wan, L\. Yang, S\. Menezes, T\. Baghaee, E\. B\. Gonzalez, J\. Neville, and T\. Safavi\(2024\)PEARL: personalizing large language model writing assistants with generation\-calibrated retrievers\.InProceedings of the 1st Workshop on Customizable NLP \(CustomNLP4U\),Note:arXiv:2311\.09180Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[37\]A\. Nichol, J\. Achiam, and J\. Schulman\(2018\)On first\-order meta\-learning algorithms\.arXiv preprint\.Note:arXiv:1803\.02999Cited by:[§II\-B](https://arxiv.org/html/2609.01615#S2.SS2.p1.1)\.
- \[38\]L\. Ning, L\. Liu, J\. Wu, N\. Wu, D\. Berlowitz, S\. Prakash, B\. Green, S\. O’Banion, and J\. Xie\(2024\)User\-LLM: efficient LLM contextualization with user embeddings\.arXiv preprint\.Note:arXiv:2402\.13598Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[39\]B\. N\. Oreshkin, P\. Rodríguez López, and A\. Lacoste\(2018\)TADAM: task dependent adaptive metric for improved few\-shot learning\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:1805\.10123Cited by:[§II\-B](https://arxiv.org/html/2609.01615#S2.SS2.p1.1)\.
- \[40\]C\. Packer, S\. Wooders, K\. Lin, V\. Fang, S\. G\. Patil, I\. Stoica, and J\. E\. Gonzalez\(2023\)MemGPT: towards LLMs as operating systems\.arXiv preprint arXiv:2310\.08560\.Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[41\]A\. Raghu, M\. Raghu, S\. Bengio, and O\. Vinyals\(2020\)Rapid learning or feature reuse? towards understanding the effectiveness of MAML\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:1909\.09157Cited by:[§II\-B](https://arxiv.org/html/2609.01615#S2.SS2.p1.1)\.
- \[42\]O\. Ram, Y\. Levine, I\. Dalmedigos, D\. Muhlgay, A\. Shashua, K\. Leyton\-Brown, and Y\. Shoham\(2023\)In\-context retrieval\-augmented language models\.Transactions of the Association for Computational Linguistics \(TACL\)11,pp\. 1316–1331\.Note:arXiv:2302\.00083Cited by:[§II\-F](https://arxiv.org/html/2609.01615#S2.SS6.p1.1)\.
- \[43\]S\. Ravi and H\. Larochelle\(2017\)Optimization as a model for few\-shot learning\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§II\-B](https://arxiv.org/html/2609.01615#S2.SS2.p1.1)\.
- \[44\]N\. Reimers and I\. Gurevych\(2019\)Sentence\-BERT: sentence embeddings using siamese BERT\-networks\.Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing \(EMNLP\)\.Cited by:[§VII\-E](https://arxiv.org/html/2609.01615#S7.SS5.p1.8),[TABLE VI](https://arxiv.org/html/2609.01615#S7.T6)\.
- \[45\]C\. Richardson, Y\. Zhang, K\. Gillespie, S\. Kar, A\. Singh, Z\. Raeesy, O\. Z\. Khan, and A\. Sethy\(2023\)Integrating summarization and retrieval for enhanced personalization via large language models\.InProceedings of the 2023 EMNLP Workshop on Customizable NLP,Cited by:[§I](https://arxiv.org/html/2609.01615#S1.p1.1),[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1),[§V\-C](https://arxiv.org/html/2609.01615#S5.SS3.p1.5)\.
- \[46\]S\. Robertson and H\. Zaragoza\(2009\)The probabilistic relevance framework: BM25 and beyond\.Foundations and Trends in Information Retrieval3\(4\),pp\. 333–389\.Cited by:[§II\-F](https://arxiv.org/html/2609.01615#S2.SS6.p1.1)\.
- \[47\]O\. Rubin, J\. Herzig, and J\. Berant\(2022\)Learning to retrieve prompts for in\-context learning\.InProceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics \(NAACL\-HLT\),pp\. 2655–2671\.Note:arXiv:2112\.08633Cited by:[§II\-F](https://arxiv.org/html/2609.01615#S2.SS6.p1.1)\.
- \[48\]A\. Salemi, S\. Kallumadi, and H\. Zamani\(2024\)Optimization methods for personalizing large language models through retrieval augmentation\.Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval \(SIGIR\)\.Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[49\]A\. Salemi, S\. Mysore, M\. Bendersky, and H\. Zamani\(2024\)LaMP: when large language models meet personalization\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(ACL\),pp\. 7370–7392\.Cited by:[§I](https://arxiv.org/html/2609.01615#S1.p1.1),[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1),[§V\-A](https://arxiv.org/html/2609.01615#S5.SS1.p1.14)\.
- \[50\]V\. Samuel, H\. P\. Zou, Y\. Zhou, S\. Chaudhari, A\. Kalyan, T\. Rajpurohit, A\. Deshpande, K\. Narasimhan, and V\. Murahari\(2024\)PersonaGym: evaluating persona agents and LLMs\.arXiv preprint\.Note:arXiv:2407\.18416Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[51\]A\. Santoro, S\. Bartunov, M\. Botvinick, D\. Wierstra, and T\. Lillicrap\(2016\)Meta\-learning with memory\-augmented neural networks\.InInternational Conference on Machine Learning \(ICML\),Cited by:[§II\-B](https://arxiv.org/html/2609.01615#S2.SS2.p1.1)\.
- \[52\]Y\. Shao, L\. Li, J\. Dai, and X\. Qiu\(2023\)Character\-LLM: a trainable agent for role\-playing\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),Note:arXiv:2310\.10158Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[53\]W\. Shi, S\. Min, M\. Yasunaga, M\. Seo, R\. James, M\. Lewis, L\. Zettlemoyer, and W\. Yih\(2024\)REPLUG: retrieval\-augmented black\-box language models\.InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics \(NAACL\),Note:arXiv:2301\.12652Cited by:[§II\-F](https://arxiv.org/html/2609.01615#S2.SS6.p1.1)\.
- \[54\]N\. Shinn, F\. Cassano, E\. Berman, A\. Gopinath, K\. Narasimhan, and S\. Yao\(2023\)Reflexion: language agents with verbal reinforcement learning\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Cited by:[§II\-D](https://arxiv.org/html/2609.01615#S2.SS4.p1.1)\.
- \[55\]C\. Snell, J\. Lee, K\. Xu, and A\. Kumar\(2024\)Scaling LLM test\-time compute optimally can be more effective than scaling model parameters\.arXiv preprint\.Note:arXiv:2408\.03314Cited by:[§II\-E](https://arxiv.org/html/2609.01615#S2.SS5.p1.1)\.
- \[56\]J\. Snell, K\. Swersky, and R\. S\. Zemel\(2017\)Prototypical networks for few\-shot learning\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:1703\.05175Cited by:[§II\-B](https://arxiv.org/html/2609.01615#S2.SS2.p1.1)\.
- \[57\]Y\. Sun, X\. Wang, Z\. Liu, J\. Miller, A\. A\. Efros, and M\. Hardt\(2020\)Test\-time training with self\-supervision for generalization under distribution shifts\.InInternational Conference on Machine Learning \(ICML\),Note:arXiv:1909\.13231Cited by:[§II\-E](https://arxiv.org/html/2609.01615#S2.SS5.p1.1)\.
- \[58\]Z\. Tan, Q\. Liu, Y\. Huang, C\. Yang, and M\. Jiang\(2024\)Democratizing large language models via personalized parameter\-efficient fine\-tuning\.Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing \(EMNLP\)\.Cited by:[§I](https://arxiv.org/html/2609.01615#S1.p1.1),[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[59\]Z\. Tan, Z\. Liu, and M\. Jiang\(2024\)Personalized pieces: efficient personalized large language models through collaborative efforts\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),Note:arXiv:2406\.10471Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[60\]Y\. Tseng, Y\. Huang, T\. Hsiao, W\. Chen, C\. Huang, Y\. Meng, and Y\. Chen\(2024\)Two tales of persona in LLMs: a survey of role\-playing and personalization\.InFindings of the Association for Computational Linguistics: EMNLP 2024,Note:arXiv:2406\.01171Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[61\]Q\. Tu, S\. Fan, Z\. Tian, and R\. Yan\(2024\)CharacterEval: a chinese benchmark for role\-playing conversational agent evaluation\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(ACL\),Note:arXiv:2401\.01275Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[62\]O\. Vinyals, C\. Blundell, T\. Lillicrap, K\. Kavukcuoglu, and D\. Wierstra\(2016\)Matching networks for one shot learning\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:1606\.04080Cited by:[§II\-B](https://arxiv.org/html/2609.01615#S2.SS2.p1.1)\.
- \[63\]J\. von Oswald, E\. Niklasson, E\. Randazzo, J\. Sacramento, A\. Mordvintsev, A\. Zhmoginov, and M\. Vladymyrov\(2023\)Transformers learn in\-context by gradient descent\.InInternational Conference on Machine Learning \(ICML\),Note:arXiv:2212\.07677Cited by:[§II\-C](https://arxiv.org/html/2609.01615#S2.SS3.p1.1)\.
- \[64\]D\. Wang, E\. Shelhamer, S\. Liu, B\. Olshausen, and T\. Darrell\(2021\)Tent: fully test\-time adaptation by entropy minimization\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2006\.10726Cited by:[§II\-E](https://arxiv.org/html/2609.01615#S2.SS5.p1.1)\.
- \[65\]H\. Wang, R\. Wang, F\. Mi, Y\. Deng, Z\. Wang, B\. Liang, R\. Xu, and K\. Wong\(2023\)Cue\-CoT: chain\-of\-thought prompting for responding to in\-depth dialogue questions with LLMs\.InFindings of the Association for Computational Linguistics: EMNLP 2023,Note:arXiv:2305\.11792Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[66\]L\. Wang, N\. Yang, and F\. Wei\(2024\)Learning to retrieve in\-context examples for large language models\.InProceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics \(EACL\),Note:arXiv:2307\.07164Cited by:[§II\-F](https://arxiv.org/html/2609.01615#S2.SS6.p1.1)\.
- \[67\]X\. Wang, J\. Wei, D\. Schuurmans, Q\. Le, E\. Chi, S\. Narang, A\. Chowdhery, and D\. Zhou\(2023\)Self\-consistency improves chain of thought reasoning in language models\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2203\.11171Cited by:[§II\-E](https://arxiv.org/html/2609.01615#S2.SS5.p1.1)\.
- \[68\]Z\. M\. Wang, Z\. Peng, H\. Que, J\. Liu, W\. Zhou, Y\. Wu, H\. Guo, R\. Gan, Z\. Ni, J\. Yang, M\. Zhang, Z\. Zhang, W\. Ouyang, K\. Xu, S\. W\. Huang, J\. Fu, and J\. Peng\(2024\)RoleLLM: benchmarking, eliciting, and enhancing role\-playing abilities of large language models\.InFindings of the Association for Computational Linguistics: ACL 2024,Note:arXiv:2310\.00746Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[69\]Z\. Z\. Wang, J\. Mao, D\. Fried, and G\. Neubig\(2024\)Agent workflow memory\.arXiv preprint arXiv:2409\.07429\.Cited by:[§II\-D](https://arxiv.org/html/2609.01615#S2.SS4.p1.1)\.
- \[70\]T\. Wolf, V\. Sanh, J\. Chaumond, and C\. Delangue\(2019\)TransferTransfo: a transfer learning approach for neural network based conversational agents\.arXiv preprint\.Note:arXiv:1901\.08149Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[71\]L\. Wu, Z\. Zheng, Z\. Qiu, H\. Wang, H\. Gu, T\. Shen, C\. Qin, C\. Zhu, H\. Zhu, Q\. Liu, H\. Xiong, and E\. Chen\(2023\)A survey on large language models for recommendation\.arXiv preprint\.Note:arXiv:2305\.19860Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[72\]T\. Wu, W\. Yuan, O\. Golovneva, J\. Xu, Y\. Tian, J\. Jiao, J\. Weston, and S\. Sukhbaatar\(2024\)Meta\-rewarding language models: self\-improving alignment with LLM\-as\-a\-meta\-judge\.arXiv preprint\.Note:arXiv:2407\.19594Cited by:[§II\-D](https://arxiv.org/html/2609.01615#S2.SS4.p1.1)\.
- \[73\]L\. Xiong, C\. Xiong, Y\. Li, K\. Tang, J\. Liu, P\. N\. Bennett, J\. Ahmed, and A\. Overwijk\(2021\)Approximate nearest neighbor negative contrastive learning for dense text retrieval\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2007\.00808Cited by:[§II\-F](https://arxiv.org/html/2609.01615#S2.SS6.p1.1)\.
- \[74\]A\. Yang, A\. Li, B\. Yang, B\. Zhang,et al\.\(2025\)Qwen3 technical report\.arXiv preprint arXiv:2505\.09388\.Cited by:[§I](https://arxiv.org/html/2609.01615#S1.p2.3),[§V\-B](https://arxiv.org/html/2609.01615#S5.SS2.p1.1)\.
- \[75\]C\. Yang, X\. Wang, Y\. Lu, H\. Liu, Q\. V\. Le, D\. Zhou, and X\. Chen\(2024\)Large language models as optimizers\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§I](https://arxiv.org/html/2609.01615#S1.p1.1),[§II\-D](https://arxiv.org/html/2609.01615#S2.SS4.p1.1)\.
- \[76\]S\. Yang, S\. C\. Han, Y\. Ding, S\. Wang, and E\. Hovy\(2026\)ToolTree: efficient LLM agent tool planning via dual\-feedback monte carlo tree search and bidirectional pruning\.Note:ICLR 2026External Links:2603\.12740Cited by:[§II\-E](https://arxiv.org/html/2609.01615#S2.SS5.p1.1)\.
- \[77\]S\. Yang, S\. C\. Han, S\. Wang, Y\. Li, Y\. Ding, and E\. Hovy\(2026\)Toward understanding misalignment in LLM agents: a survey of taxonomy, causes, mitigation, and evaluation\.Note:ACL ARR 2026External Links:[Link](https://openreview.net/forum?id=zzTEGP2BYa)Cited by:[§II\-D](https://arxiv.org/html/2609.01615#S2.SS4.p1.1)\.
- \[78\]S\. Yao, D\. Yu, J\. Zhao, I\. Shafran, T\. L\. Griffiths, Y\. Cao, and K\. Narasimhan\(2023\)Tree of thoughts: deliberate problem solving with large language models\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:2305\.10601Cited by:[§II\-E](https://arxiv.org/html/2609.01615#S2.SS5.p1.1)\.
- \[79\]T\. Yu, D\. Quillen, Z\. He, R\. Julian, K\. Hausman, C\. Finn, and S\. Levine\(2019\)Meta\-World: a benchmark and evaluation for multi\-task and meta reinforcement learning\.InConference on Robot Learning \(CoRL\),Note:arXiv:1910\.10897Cited by:[§II\-B](https://arxiv.org/html/2609.01615#S2.SS2.p1.1)\.
- \[80\]W\. Yuan, R\. Y\. Pang, K\. Cho, S\. Sukhbaatar, J\. Xu, and J\. Weston\(2024\)Self\-rewarding language models\.arXiv preprint\.Note:arXiv:2401\.10020Cited by:[§II\-D](https://arxiv.org/html/2609.01615#S2.SS4.p1.1)\.
- \[81\]K\. Zhang, Y\. Kang, F\. Zhao, and X\. Liu\(2023\)LLM\-based medical assistant personalization with short\- and long\-term memory coordination\.arXiv preprint\.Note:arXiv:2309\.11696Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[82\]Q\. Zhang, C\. Hu,et al\.\(2025\)Agentic context engineering: evolving contexts for self\-improving language models\.arXiv preprint arXiv:2510\.04618\.Cited by:[§II\-D](https://arxiv.org/html/2609.01615#S2.SS4.p1.1)\.
- \[83\]S\. Zhang, E\. Dinan, J\. Urbanek, A\. Szlam, D\. Kiela, and J\. Weston\(2018\)Personalizing dialogue agents: i have a dog, do you have pets too?\.InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics \(ACL\),Note:arXiv:1801\.07243Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[84\]W\. Zhang, X\. Tang, Y\. Wang, Z\. Liu, X\. Wang, and P\. S\. Yu\(2025\)PersonaAgent: when large language model agents meet personalization at test time\.arXiv preprint arXiv:2506\.06254\.Cited by:[§I](https://arxiv.org/html/2609.01615#S1.p1.1),[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[85\]Z\. Zhang, R\. A\. Rossi, B\. Kveton, Y\. Shao, D\. Yang, H\. Zamani,et al\.\(2024\)Personalization of large language models: a survey\.arXiv preprint arXiv:2411\.00027\.Cited by:[§I](https://arxiv.org/html/2609.01615#S1.p1.1),[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[86\]A\. Zhao, D\. Huang, Q\. Xu, M\. Lin, Y\. Liu, and G\. Huang\(2024\)ExpeL: LLM agents are experiential learners\.InAAAI Conference on Artificial Intelligence,Cited by:[§II\-D](https://arxiv.org/html/2609.01615#S2.SS4.p1.1)\.
- \[87\]W\. Zhong, L\. Guo, Q\. Gao, H\. Ye, and Y\. Wang\(2024\)MemoryBank: enhancing large language models with long\-term memory\.InProceedings of the AAAI Conference on Artificial Intelligence \(AAAI\),Note:arXiv:2305\.10250Cited by:[§II\-A](https://arxiv.org/html/2609.01615#S2.SS1.p1.1)\.
- \[88\]Y\. Zhou, A\. I\. Muresanu, Z\. Han, K\. Paster, S\. Pitis, H\. Chan, and J\. Ba\(2023\)Large language models are human\-level prompt engineers\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2211\.01910Cited by:[§I](https://arxiv.org/html/2609.01615#S1.p1.1),[§II\-D](https://arxiv.org/html/2609.01615#S2.SS4.p1.1)\.