Learning to Adapt Cross-Domain Preferences via Meta-LoRA for LLM Personalization

arXiv cs.AI 论文

摘要

This paper introduces PAC-Bayes-regularized Meta-LoRA for cross-domain LLM personalization, enabling zero- and few-shot adaptation to user preferences while preventing overfitting under sparse evidence. Experiments on benchmarks like HiCUPID show significant improvements in cross-domain win rates and cold-start scenarios.

arXiv:2608.12389v1 Announce Type: new Abstract: Cross-domain zero- or few-shot personalization aims to generate user-preferred responses in unseen conversational domains from only a handful of target-domain interactions. Existing adaptation methods struggle to calibrate update magnitude under sparse evidence and thus overfit, whereas history-transfer methods often entangle user preferences with source-domain artifacts, yielding unreliable personalization priors and negative transfer. To calibrate adaptation to evidence quality, we propose PAC-Bayes-regularized Meta-LoRA, which uses a meta-learned LoRA initialization as both the adaptation start and prior center, while adjusting update strength according to support-set size and predictive uncertainty. This limits overfitting under sparse or ambiguous evidence while permitting stronger personalization as evidence grows. Controlled adaptation alone does not determine which preferences should transfer across domains or how they should be expressed. We therefore functionally decompose personalization priors into user and domain components, using a human-readable prompt for stable preferences and topology-preserving soft tokens for domain-specific hidden-space conditioning. Experiments across multiple benchmarks and personalization tasks show consistent gains over strong baselines. On HiCUPID, our method reduces cross-domain win-rate degradation by 47.9% relative to the best competing baseline and improves win rate by 110.2% under unseen-user cold start.
查看原文
查看缓存全文

缓存时间: 2026/08/14 09:25

# Learning to Adapt Cross-Domain Preferences via Meta-LoRA for LLM Personalization
Source: [https://arxiv.org/html/2608.12389](https://arxiv.org/html/2608.12389)
Xuefei Wang1,Jun Han1,Zixuan Wang1,Qingkai Zeng2, Xiao Wang1,Ruijie Wang1\\corresponding,Jianxin Li1

###### Abstract

Cross\-domain zero\- or few\-shot personalization aims to generate user\-preferred responses in unseen conversational domains from only a handful of target\-domain interactions\. Existing adaptation methods struggle to calibrate update magnitude under sparse evidence and thus overfit, whereas history\-transfer methods often entangle user preferences with source\-domain artifacts, yielding unreliable personalization priors and negative transfer\. To calibrate adaptation to evidence quality, we propose PAC\-Bayes\-regularized Meta\-LoRA, which uses a meta\-learned LoRA initialization as both the adaptation start and prior center, while adjusting update strength according to support\-set size and predictive uncertainty\. This limits overfitting under sparse or ambiguous evidence while permitting stronger personalization as evidence grows\. Controlled adaptation alone does not determine which preferences should transfer across domains or how they should be expressed\. We therefore functionally decompose personalization priors into user and domain components, using a human\-readable prompt for stable preferences and topology\-preserving soft tokens for domain\-specific hidden\-space conditioning\. Experiments across multiple benchmarks and personalization tasks show consistent gains over strong baselines\. On HiCUPID, our method reduces cross\-domain win\-rate degradation by 47\.9% relative to the best competing baseline and improves win rate by 110\.2% under unseen\-user cold start\.

## Introduction

Large Language Models \(LLMs\) have demonstrated strong general capabilities across a wide range of tasks\. However, a persistent gap remains between their generic competence and user satisfaction at the individual level\(Salemi and Zamani[2025](https://arxiv.org/html/2608.12389#bib.bib44); Moket al\.[2025b](https://arxiv.org/html/2608.12389#bib.bib43)\)\. This gap arises from the inherently personalized nature of human interaction: user preferences are high\-dimensional, latent, and context\-dependent\(Gaoet al\.[2026](https://arxiv.org/html/2608.12389#bib.bib116)\)\. They shape style, reasoning depth, evidence use, and constraint handling\. Personalized response generation therefore requires more than retrieving user facts or imitating past wording\(Liuet al\.[2025](https://arxiv.org/html/2608.12389#bib.bib111)\): it must identify stable user tendencies and realize them under the conventions of the current conversational domain\. Reliably transferring such preferences across domains remains a central challenge for user\-level personalization\(Linget al\.[2025](https://arxiv.org/html/2608.12389#bib.bib65)\)\.

![Refer to caption](https://arxiv.org/html/2608.12389v1/x1.png)Figure 1:Illustration of the cross\-domain LLM personalization problem, which presents few\-shot overfitting and cross\-domain gap challenges\.This challenge is particularly acute when target\-domain interactions are scarce or absent\(Tanet al\.[2024a](https://arxiv.org/html/2608.12389#bib.bib109)\)\. Under zero\- or few\-shot conditions, limited observations provide incomplete and potentially ambiguous evidence about a user’s target\-domain preferences, making adaptation vulnerable to overfitting and preference distortion\(Linget al\.[2025](https://arxiv.org/html/2608.12389#bib.bib65); Zhanget al\.[2023](https://arxiv.org/html/2608.12389#bib.bib88)\)\. Existing methods lack principled control over update magnitude: aggressive updates may overwrite transferable personalization patterns, whereas conservative updates fail to capture user\-specific patterns\(Choiet al\.[2025a](https://arxiv.org/html/2608.12389#bib.bib112)\)\. Cross\-domain heterogeneity further complicates transfer\(Gaoet al\.[2026](https://arxiv.org/html/2608.12389#bib.bib116)\)\. The same preference may require different realizations across domains with distinct discourse conventions and pragmatic expectations; consequently, history\-based representations may entangle stable user tendencies with source\-domain artifacts, causing negative transfer or preference drift\. Together, evidence scarcity and domain heterogeneity remain key obstacles to robust cross\-domain personalization, as illustrated in Figure[1](https://arxiv.org/html/2608.12389#Sx1.F1)\.

Prior work personalizes LLMs through retrieval and profiling, parameter adaptation, personalized alignment, and inference\-time steering\. Retrieval\- and profile\-based methods may transfer source\-specific content rather than persistent preferences\(Salemiet al\.[2024b](https://arxiv.org/html/2608.12389#bib.bib49); Zhang[2024](https://arxiv.org/html/2608.12389#bib.bib48); Wanget al\.[2024](https://arxiv.org/html/2608.12389#bib.bib70); Namet al\.[2025](https://arxiv.org/html/2608.12389#bib.bib75)\)\. Parameter\-efficient fine\-tuning and meta\-learning can be dominated by atypical sparse examples and miscalibrate adaptation under domain shift\(Choiet al\.[2025b](https://arxiv.org/html/2608.12389#bib.bib45); Zhanget al\.[2025a](https://arxiv.org/html/2608.12389#bib.bib52); Singhet al\.[2025](https://arxiv.org/html/2608.12389#bib.bib94); Tanet al\.[2025](https://arxiv.org/html/2608.12389#bib.bib78)\)\. Alignment and steering control user\-level behavior but rarely model its realization under unseen\-domain conventions\(Caoet al\.[2024a](https://arxiv.org/html/2608.12389#bib.bib57); Chenet al\.[2024](https://arxiv.org/html/2608.12389#bib.bib76); Kimet al\.[2025](https://arxiv.org/html/2608.12389#bib.bib74); Zhuet al\.[2025](https://arxiv.org/html/2608.12389#bib.bib104)\)\. These paradigms therefore leave two related problems insufficiently addressed: adapting parameters according to scarce target evidence and constructing personalization priors that transfer reliably across domains\.

To address these problems, we propose a unified framework that combines PAC\-Bayes\-regularized Meta\-LoRA with dual\-channel prior conditioning\. Episodic meta\-learning over source domain tasks yields a LoRA initialization that captures transferable adaptation structure\. We use this initialization both as the starting point for adaptation and as the center of the PAC\-Bayes prior, thereby encouraging target\-domain updates to preserve transferable behavior and deviate only when supported by user evidence\. The anchoring strength is calibrated by adaptation evidence quality\. Throughout meta\-training and target adaptation, the model is conditioned on personalization priors, allowing limited target interactions to refine the personalization context rather than infer user preferences from scratch\.

Constructing these priors requires distinguishing what should transfer across domains from how it should be expressed in the target domain\. We therefore functionally decompose personalization priors into user and domain components and match each to an appropriate conditioning interface\. Stable user preferences are injected through a human\-readable prompt, preserving interpretability, while target\-domain conditions are injected as compact soft tokens that provide fine\-grained hidden\-space control\. This dual\-channel design avoids appending long histories and allows the same user preference to be realized differently across domains\.

Our main contributions are threefold:

- •We introduce PAC\-Bayes\-regularized Meta\-LoRA, which uses a meta\-learned LoRA initialization as both adaptation start and prior center, calibrating update strength by support size and predictive uncertainty\.
- •We propose functional user\-domain prior factorization with dual\-channel conditioning, injecting readable user preferences through prompts and compact domain conditions through soft tokens\.
- •Experiments show consistent improvements across cross\-domain, few\-shot, distant\-domain, and unseen\-user settings, reducing cross\-domain win\-rate degradation by 47\.9% and improving unseen\-user win rate by 110\.2%\.

![Refer to caption](https://arxiv.org/html/2608.12389v1/x2.png)Figure 2:Framework overview\. User history yields a textual promptPuP\_\{u\}, while graph\-guided domain composition produces soft tokensTtT\_\{t\}; both augment the input, while PAC\-Bayes\-regularized Meta\-LoRA controls target\-domain parameter adaptation\.
## Related Work

#### LLM Personalization\.

Existing methods condition LLMs on retrieved interactions, textual profiles, continuous prompts, user\-specific parameters, or inference\-time preference directions\(Salemiet al\.[2024a](https://arxiv.org/html/2608.12389#bib.bib108); Liuet al\.[2025](https://arxiv.org/html/2608.12389#bib.bib111); Lesteret al\.[2021](https://arxiv.org/html/2608.12389#bib.bib120); Li and Liang[2021](https://arxiv.org/html/2608.12389#bib.bib121); Tanet al\.[2024a](https://arxiv.org/html/2608.12389#bib.bib109); Caoet al\.[2024b](https://arxiv.org/html/2608.12389#bib.bib117)\)\. Retrieval and profiles are transparent but may carry source\-specific topics or wording into a new domain, while continuous prompts and adapters provide expressive conditioning but may entangle recurring user tendencies with their observed domain\-specific realizations\(Qiuet al\.[2025a](https://arxiv.org/html/2608.12389#bib.bib113); Choiet al\.[2025a](https://arxiv.org/html/2608.12389#bib.bib112)\)\. Alignment and decoding\-time steering control personalized behavior without parameter updates\(Balepuret al\.[2025](https://arxiv.org/html/2608.12389#bib.bib115); Buet al\.[2025](https://arxiv.org/html/2608.12389#bib.bib114)\), but primarily target observed preferences rather than their realization under unseen\-domain conventions\. Our work instead assigns user tendencies and domain realization conditions distinct functional roles and conditioning interfaces\.

#### Meta\-Learning and Controlled Adaptation\.

Meta\-learning learns transferable initializations or update rules for rapid adaptation from limited task data\(Finnet al\.[2017b](https://arxiv.org/html/2608.12389#bib.bib118)\)\. Personalized parameter\-efficient tuning similarly adapts compact user\-specific parameters while freezing the backbone\(Huet al\.[2022](https://arxiv.org/html/2608.12389#bib.bib39); Tanet al\.[2024a](https://arxiv.org/html/2608.12389#bib.bib109)\)\. These approaches emphasize adaptation speed and support\-set fit, but typically do not regulate parameter displacement according to support\-set size and model uncertainty\. PAC\-Bayes analysis relates empirical risk to prior–posterior complexity\(McAllester[1999](https://arxiv.org/html/2608.12389#bib.bib119)\), providing a principled basis for such control\. We instantiate this connection by centering the parameter prior at a source\-learned LoRA initialization and using the resulting complexity structure to guide evidence\-calibrated target adaptation\.

## Methodology

Given a useruuwith source\-domain historyℋu\\mathcal\{H\}\_\{u\}observed during training, we consider a target domaindtd\_\{t\}that is unseen during training\. Our goal is to generate a user\-aligned responseyty\_\{t\}to a queryxtx\_\{t\}in the sparsely observed target domaindtd\_\{t\}\. Let𝒮=\{\(xi,yi\)\}i=1K\\mathcal\{S\}=\\\{\(x\_\{i\},y\_\{i\}\)\\\}\_\{i=1\}^\{K\}denote the target\-domain support set, whereK=0K=0andK\>0K\>0correspond to the zero\-shot and few\-shot settings, respectively\. Complete task and information\-access definitions are in Appendix B\.

We propose a unified framework consisting of PAC\-Bayes\-Regularized Meta\-LoRA and dual\-channel personalization priors, as illustrated in Figure[2](https://arxiv.org/html/2608.12389#Sx1.F2)\. Episodic meta\-learning first learns a transferable LoRA initializationθ0\\theta\_\{0\}, used directly with the injected priors in the zero\-shot setting and as the starting point for a few inner\-loop updates on𝒮\\mathcal\{S\}in the few\-shot setting\. We further useθ0\\theta\_\{0\}as the center of a PAC\-Bayes parameter prior and calibrate the adaptation strength to the available support evidence, limiting excessive parameter drift under sparse or uncertain observations\.

The personalization condition is jointly constructed through two functionally complementary channels: a user\-side prior and a domain\-side prior\. The former is derived fromℋu\\mathcal\{H\}\_\{u\}and injected as a textual prompt, whereas the latter is constructed from source\-domain characteristics and target\-context evidence and injected as soft tokens\. Together, they enable user preferences to be expressed appropriately under the conventions of the target domain\.

### PAC\-Bayes\-Regularized Meta\-LoRA

When target\-domain support examples are scarce or ambiguous, direct fine\-tuning may overfit incidental patterns and impair its generalization to future interactions\. Although standard Meta\-LoRA provides an initialization for rapid adaptation, its inner loop typically optimizes only the support loss and does not regulate how far parameters should move under limited evidence\. We therefore anchor target domain LoRA updates to the meta\-initialization and adapt the strength and geometry of this anchor to the available evidence\.

#### Meta\-Learned Prior and PAC\-Bayes\-Guided Constraint\.

Letθ=\{θl\}l=1L\\theta=\\\{\\theta\_\{l\}\\\}\_\{l=1\}^\{L\}denote the LoRA parameters inserted intoLLTransformer layers, letθ0=\{θ0,l\}l=1L\\theta\_\{0\}=\\\{\\theta\_\{0,l\}\\\}\_\{l=1\}^\{L\}be the initialization learned through source\-domain episodic meta\-training, and letg​\(l\)g\(l\)index the depth group containing layerll\. Our goal is to retain the rapid adaptation enabled byθ0\\theta\_\{0\}while preventing a small target support set from driving the model toward incidental patterns\. PAC\-Bayes is well suited to this objective because it jointly relates support\-set fit, deviation from a target\-support\-independent prior, and the number of available examples\. We therefore useθ0\\theta\_\{0\}as both the starting point for task\-specific adaptation and the center of a Gaussian parameter prior:

P0=∏l=1L𝒩​\(θ0,l,σg​\(l\)2​I\)​Qθ=∏l=1L𝒩​\(θl,σg​\(l\)2​I\),KL​\(Qθ∥P0\)=∑l=1L∥θl−θ0,l∥222​σg​\(l\)2\.\\begin\{array\}\[\]\{c\}\\displaystyle P\_\{0\}=\\prod\_\{l=1\}^\{L\}\\mathcal\{N\}\\\!\\bigl\(\\theta\_\{0,l\},\\,\\sigma\_\{g\(l\)\}^\{2\}I\\bigr\)\\;\\;Q\_\{\\theta\}=\\prod\_\{l=1\}^\{L\}\\mathcal\{N\}\\\!\\bigl\(\\theta\_\{l\},\\,\\sigma\_\{g\(l\)\}^\{2\}I\\bigr\),\\\\\[4\.0pt\] \\displaystyle\\mathrm\{KL\}\\\!\\bigl\(Q\_\{\\theta\}\\\|P\_\{0\}\\bigr\)=\\sum\_\{l=1\}^\{L\}\\frac\{\\lVert\\theta\_\{l\}\-\\theta\_\{0,l\}\\rVert\_\{2\}^\{2\}\}\{2\\,\\sigma\_\{g\(l\)\}^\{2\}\}\.\\end\{array\}\(1\)The depth grouping,θ0\\theta\_\{0\}, and the group\-wise variances\{σg2\}\\\{\\sigma\_\{g\}^\{2\}\\\}are learned exclusively from source\-domain tasks and fixed before𝒮\\mathcal\{S\}is observed\. Conditioned on the source training data,P0P\_\{0\}is therefore independent of the target support set, and its KL complexity measures a depth\-group\-weighted displacement from the transferable meta\-initialization\.

For the PAC\-Bayes analysis, we map the token\-averaged negative log\-likelihood to\[0,1\]\[0,1\]using a monotone mapping fixed from source\-domain data; actual adaptation continues to optimize the original negative log\-likelihood\. Consider a support set𝒮=\{\(xi,yi\)\}i=1K\\mathcal\{S\}=\\\{\(x\_\{i\},y\_\{i\}\)\\\}\_\{i=1\}^\{K\}ofK≥1K\\geq 1interactions drawn independently from the target user–domain distribution\. LetR^𝒮t​\(Qθ\)\\widehat\{R\}\_\{\\mathcal\{S\}\_\{t\}\}\(Q\_\{\\theta\}\)denote the empirical risk of the randomized predictor obtained by sampling parameters fromQθQ\_\{\\theta\}, and letR​\(Qθ\)R\(Q\_\{\\theta\}\)denote its expected risk on a future interaction\. For anyδ∈\(0,1\)\\delta\\in\(0,1\), with probability at least1−δ1\-\\deltaover𝒮\\mathcal\{S\}, the following standard PAC\-Bayes bound holds simultaneously for allQθQ\_\{\\theta\}in the above family\(Maurer[2004](https://arxiv.org/html/2608.12389#bib.bib47)\):

R​\(Qθ\)≤R^𝒮​\(Qθ\)\+KL​\(Qθ∥P0\)\+ln⁡\(\(K\+1\)/δ\)2​K\.R\(Q\_\{\\theta\}\)\\leq\\hat\{R\}\_\{\\mathcal\{S\}\}\(Q\_\{\\theta\}\)\+\\sqrt\{\\frac\{\\mathrm\{KL\}\(Q\_\{\\theta\}\\\|P\_\{0\}\)\+\\ln\\\!\\bigl\(\(K\{\+\}1\)/\\delta\\bigr\)\}\{2K\}\}\.\(2\)Thus, the same displacement fromθ0\\theta\_\{0\}contributes more strongly to the bound when fewer support examples are available\. LetAθ=KL⁡\(Qθ∥P0\)\+ln⁡\(\(K\+1\)/δ\)A\_\{\\theta\}=\\operatorname\{KL\}\(Q\_\{\\theta\}\\\|P\_\{0\}\)\+\\ln\(\(K\+1\)/\\delta\)\. For anyβ\>0\\beta\>0,Aθ/\(2​K\)≤Aθ/\(4​β​K\)\+β/2\\sqrt\{A\_\{\\theta\}/\(2K\)\}\\leq A\_\{\\theta\}/\(4\\beta K\)\+\\beta/2\. For fixedKK,δ\\delta, andβ\\beta, removing terms independent ofθ\\thetaleavesKL⁡\(Qθ∥P0\)/\(4​β​K\)\\operatorname\{KL\}\(Q\_\{\\theta\}\\\|P\_\{0\}\)/\(4\\beta K\), motivating theK−1K^\{\-1\}\-scaled anchor in the adaptation objective below\. Since response generation uses the posterior meanθ\\thetarather than samples fromQθQ\_\{\\theta\}, this derivation determines the prior\-centered geometry and support\-size dependence of our deterministic adaptation surrogate\.

#### Evidence\- and Structure\-Aware Adaptation Objective\.

The preceding PAC\-Bayes analysis relates adaptation complexity to the amount of target evidence, but support\-set size alone does not indicate how informative that evidence is for the current model\. We therefore characterize effective adaptation evidence through two complementary signals:KKprovides a coarse measure of preference\-evidence coverage, while predictive entropy captures model\-relative uncertainty about the observed responses\. Consistent with uncertainty\-aware calibration, we treat entropy as a property of the model’s predictive distribution\(Krishnan and Tickoo[2020](https://arxiv.org/html/2608.12389#bib.bib83)\)\. ForK\>0K\>0, letH~​\(𝒮;θ0\)∈\[0,1\]\\widetilde\{H\}\(\\mathcal\{S\};\\theta\_\{0\}\)\\in\[0,1\]denote the average token\-level predictive entropy underθ0\\theta\_\{0\}, normalized by the maximum vocabulary entropy\. We define

γ​\(𝒮\)=1K​\[1\+α​H~​\(𝒮;θ0\)\],α≥0,\\gamma\(\\mathcal\{S\}\)=\\frac\{1\}\{K\}\\left\[1\+\\alpha\\,\\widetilde\{H\}\(\\mathcal\{S\};\\theta\_\{0\}\)\\right\],\\qquad\\alpha\\geq 0,\(3\)whereα\\alphacontrols the entropy\-dependent increase in anchoring strength and is selected using source\-domain validation tasks before target adaptation\. The unit offset preserves the PAC\-Bayes\-motivatedK−1K^\{\-1\}scaling whenα=0\\alpha=0\. A high entropy value means thatθ0\\theta\_\{0\}distributes probability mass across many plausible continuations; with only a few observations, the support loss then provides less decisive evidence for how far the model should move from the transferable initialization\. Accordingly, higher uncertainty strengthens the anchor without preventing adaptation altogether\.

The group\-wise variancesσg​\(l\)2\\sigma\_\{g\(l\)\}^\{2\}allocate source\-learned adaptation freedom across network depths\. Within each layer, different parameter directions can have substantially different effects on the support loss\. Letℓi​\(θ0\)\\ell\_\{i\}\(\\theta\_\{0\}\)be the token\-averaged generation loss on theii\-th support interaction\. We estimate a diagonal empirical Fisher proxy atθ0\\theta\_\{0\}as

gi,l=∇θlℓi​\(θ0\),F^l=Diag⁡\(1K​∑i=1Kgi,l⊙2\)\.g\_\{i,l\}=\\nabla\_\{\\theta\_\{l\}\}\\ell\_\{i\}\(\\theta\_\{0\}\),\\qquad\\widehat\{F\}\_\{l\}=\\operatorname\{Diag\}\\left\(\\frac\{1\}\{K\}\\sum\_\{i=1\}^\{K\}g\_\{i,l\}^\{\\odot 2\}\\right\)\.\(4\)This squared\-gradient estimate provides a tractable, support\-conditioned approximation to local anisotropic sensitivity\. Larger entries identify directions in which the support loss is more sensitive to parameter changes, allowing the regularizer to constrain these directions more strongly while retaining greater flexibility elsewhere\. We computeF^l\\widehat\{F\}\_\{l\}once atθ0\\theta\_\{0\}and keep it fixed throughout the inner\-loop updates\. DefiningΔ​θl=θl−θ0,l\\Delta\\theta\_\{l\}=\\theta\_\{l\}\-\\theta\_\{0,l\}, the resulting adaptation objective is

ℒadapt​\(θ;𝒮\)=ℒ^𝒮​\(θ\)\+γ​\(𝒮\)​∑l=1LΔ​θl⊤​F^l​Δ​θl2​σg​\(l\)2,\\mathcal\{L\}\_\{\\mathrm\{adapt\}\}\(\\theta;\\mathcal\{S\}\)=\\widehat\{\\mathcal\{L\}\}\_\{\\mathcal\{S\}\}\(\\theta\)\+\\gamma\(\\mathcal\{S\}\)\\sum\_\{l=1\}^\{L\}\\frac\{\\Delta\\theta\_\{l\}^\{\\top\}\\widehat\{F\}\_\{l\}\\Delta\\theta\_\{l\}\}\{2\\sigma\_\{g\(l\)\}^\{2\}\},\(5\)whereℒ^𝒮​\(θ\)\\widehat\{\\mathcal\{L\}\}\_\{\\mathcal\{S\}\}\(\\theta\)is the average token\-level generation loss on the support set\.γ​\(𝒮\)\\gamma\(\\mathcal\{S\}\)controls the overall anchoring strength,σg​\(l\)2\\sigma\_\{g\(l\)\}^\{2\}allocates adaptation freedom across depth groups, andF^l\\widehat\{F\}\_\{l\}shapes the constraint across local parameter directions\. The PAC\-Bayes derivation supplies the prior\-centered geometry andK−1K^\{\-1\}dependence, while predictive entropy and empirical Fisher calibrate the deterministic surrogate to the current support evidence\. BecauseF^l\\widehat\{F\}\_\{l\}is support\-dependent, it is used only to reweight the deterministic adaptation surrogate and is not part of the target\-support\-independent priorP0P\_\{0\}\. Full derivations are provided in Appendix C\.

#### Episodic Meta\-Optimization and Target\-Domain Adaptation\.

During source\-domain meta\-training, each user–domain pair defines a taskτ\\tauwith disjoint support and query sets,𝒮τ\\mathcal\{S\}\_\{\\tau\}and𝒬τ\\mathcal\{Q\}\_\{\\tau\}\. Starting fromθ0\\theta\_\{0\}, the inner loop performsMMgradient steps onℒadapt​\(θ;𝒮τ\)\\mathcal\{L\}\_\{\\mathrm\{adapt\}\}\(\\theta;\\mathcal\{S\}\_\{\\tau\}\), after which the outer loop minimizes the generation loss of the adapted parameters on𝒬τ\\mathcal\{Q\}\_\{\\tau\}and jointly updatesθ0\\theta\_\{0\}and the group\-wise variances\{σg2\}\\\{\\sigma\_\{g\}^\{2\}\\\}\. Because query examples are excluded from within\-task adaptation, the outer objective learns a prior center and depth\-wise adaptation freedoms that improve generation on new interactions after constrained adaptation, rather than merely fitting the support set\. For each episode,H~​\(𝒮τ;θ0\)\\widetilde\{H\}\(\\mathcal\{S\}\_\{\\tau\};\\theta\_\{0\}\)and\{F^l\}\\\{\\widehat\{F\}\_\{l\}\\\}are computed once atθ0\\theta\_\{0\}and held fixed across all inner\-loop steps\. Gradients are stopped only through the construction of these entropy and Fisher statistics; this does not freezeθ0\\theta\_\{0\}, which remains an outer\-loop parameter and serves as both the initialization and anchor center\. The same within\-task objective and update rule are used during source\-domain meta\-training and few\-shot target\-domain adaptation\. Training and inference algorithms are given in Appendix F\.

### Graph\-Guided Domain Prior Injection

Cross\-domain personalization requires two functionally distinct forms of conditioning: relatively stable user preference tendencies and target\-domain conditions governing how those tendencies should be realized\. We therefore adopt a functional, rather than semantic, factorization of personalization priors\. A textual summary of the user’s cross\-domain historyℋu\\mathcal\{H\}\_\{u\}specifies stable preference tendencies and is injected as a prompt to preserve semantic transparency\. The domain prior instead specifies continuous realization conditions, such as information organization, discourse conventions, and response structure, and is injected as soft tokens that directly condition hidden\-state computation\. Their distinction concerns their roles and conditioning interfaces during generation, rather than mutually exclusive semantic content\. We treat the textual user prompt as standard conditioning and focus below on constructing the domain prior from source\-domain relations and injecting it through topology\-preserving soft tokens\.

#### Reliable Multi\-View Domain Graph and Target\-Conditioned Composition\.

For each source domaind∈𝒟sd\\in\\mathcal\{D\}\_\{s\}, we construct three complementary observable views that describe what the domain discusses, how its interactions are conventionally organized, and what response forms are typically expected\. The semantic\-content view captures topics, entities, and domain\-specific concepts; the interaction\-pragmatics view characterizes discourse organization and expression conventions; and the response\-structure view summarizes statistical output patterns such as response length and paragraph or list organization\.

Because these views contain heterogeneous feature types, each view is first processed by a geometry\-compatible transformation to obtain a comparable view\-specific embedding\. For every domain–view pair, we estimate a reliability score that increases with effective sample size and decreases with estimation uncertainty\. Each domain forms a node in a view\-specific graph, and each edge combines the similarity between the corresponding embeddings with the reliability of the participating estimates\. Edges whose reliability\-adjusted similarity falls below a fixed threshold are removed, and normalized weighted neighborhood aggregation is applied over the remaining edges to reduce noise without propagating unreliable relations\. After mapping the resulting view\-specific node embeddings to probability measures on a shared feature support, we compute an entropy\-regularized Wasserstein barycenter using normalized reliability weights and map the resulting measure back to a domain representationzd∈ℝhzz\_\{d\}\\in\\mathbb\{R\}^\{h\_\{z\}\}, wherehzh\_\{z\}is the representation dimensionality\. This fusion preserves the relational geometry expressed by the heterogeneous views while reducing the influence of sparse or high\-variance estimates\. The resulting source\-domain representations form the prior bankℬ=\{zd:d∈𝒟s\}\\mathcal\{B\}=\\\{z\_\{d\}:d\\in\\mathcal\{D\}\_\{s\}\\\}\.

To learn how source\-domain priors should be composed for an unseen domain, we adopt leave\-one\-domain\-out training\. Each source domainddis treated in turn as a pseudo\-target, while the remaining source domains are available as references\. During each leave\-one\-domain\-out episode, the pseudo\-target node and its incident edges are masked from graph propagation and excluded from the candidate bank; its full representation is used only as the reconstruction target\.

A target\-evidence representationz^d\\widehat\{z\}\_\{d\}is constructed from the pseudo\-target domain description and simulated sparse interactions\. Retrieval ranks the available source nodes using their graph relations, estimation reliability, and the uncertainty of the target evidence, producing a candidate set𝒩k​\(d\)⊆𝒟s∖\{d\}\\mathcal\{N\}\_\{k\}\(d\)\\subseteq\\mathcal\{D\}\_\{s\}\\setminus\\\{d\\\}containing at mostkkdomains and retrieval confidence scoresωd​d′\\omega\_\{dd^\{\\prime\}\}for candidatesd′∈𝒩k​\(d\)d^\{\\prime\}\\in\\mathcal\{N\}\_\{k\}\(d\)\. Target\-conditioned attention composes these candidates as

ad​d′=softmaxd′​\(⟨Wq​z^d,Wk​zd′⟩ha\+η​ln⁡\(ωd​d′\+ϵ\)\),z¯d=∑d′∈𝒩k​\(d\)ad​d′​zd′\.\\begin\{gathered\}a\_\{dd^\{\\prime\}\}=\\mathrm\{softmax\}\_\{d^\{\\prime\}\}\\\!\\Bigl\(\\tfrac\{\\langle W\_\{q\}\\hat\{z\}\_\{d\},\\;W\_\{k\}z\_\{d^\{\\prime\}\}\\rangle\}\{\\sqrt\{h\_\{a\}\}\}\+\\eta\\ln\(\\omega\_\{dd^\{\\prime\}\}\{\+\}\\epsilon\)\\Bigr\),\\\\\[6\.0pt\] \\bar\{z\}\_\{d\}=\\textstyle\\sum\_\{d^\{\\prime\}\\in\\mathcal\{N\}\_\{k\}\(d\)\}a\_\{dd^\{\\prime\}\}\\,z\_\{d^\{\\prime\}\}\.\\end\{gathered\}\(6\)Here,WqW\_\{q\}andWkW\_\{k\}are learnable projections into anhah\_\{a\}\-dimensional attention space,η≥0\\eta\\geq 0controls the contribution of retrieval confidence, andϵ\>0\\epsilon\>0ensures numerical stability\. The coefficientad​d′a\_\{dd^\{\\prime\}\}is the final target\-conditioned contribution of candidate domaind′d^\{\\prime\}, whereasωd​d′\\omega\_\{dd^\{\\prime\}\}reflects its confidence before composition\. Thus, retrieval determines which source domains are admissible transfer references, while attention determines how their complementary information is combined\. The fusion module is trained to reconstruct the held\-out representationzdz\_\{d\}fromz¯d\\bar\{z\}\_\{d\}while preserving its neighborhood relations in the source\-domain graph, thereby learning a transferable composition rule rather than memorizing domain identities\.

#### Topology\-Preserving Soft\-Token Injection\.

The fused representationz¯d\\bar\{z\}\_\{d\}lies in an external domain space and must be converted into a condition that can participate directly in language\-model computation\. We use a projectorΠψ\\Pi\_\{\\psi\}, parameterized byψ\\psi, to producenTn\_\{T\}soft tokens:

Td\\displaystyle T\_\{d\}=Πψ​\(z¯d\)=reshape⁡\[W2​φ​\(W1​z¯d\)\]∈ℝnT×h,\\displaystyle=\\Pi\_\{\\psi\}\(\\bar\{z\}\_\{d\}\)=\\operatorname\{reshape\}\\left\[W\_\{2\}\\varphi\(W\_\{1\}\\bar\{z\}\_\{d\}\)\\right\]\\in\\mathbb\{R\}^\{n\_\{T\}\\times h\},\(7\)ℒtopo\\displaystyle\\mathcal\{L\}\_\{\\mathrm\{topo\}\}=1−CKA⁡\(Z¯,T¯\)\+λattn​‖SA−Sz‖F2\.\\displaystyle=1\-\\operatorname\{CKA\}\(\\overline\{Z\},\\overline\{T\}\)\+\\lambda\_\{\\mathrm\{attn\}\}\\left\\\|S^\{A\}\-S^\{z\}\\right\\\|\_\{F\}^\{2\}\.Here,W1W\_\{1\}andW2W\_\{2\}are the projector parameters,φ\\varphiis its nonlinear activation,nTn\_\{T\}is the number of domain tokens, andhhis the hidden\-state dimensionality of the language model\. The matricesZ¯\\overline\{Z\}andT¯\\overline\{T\}respectively collect the fused domain representations in a training batch and the pooled representations of their corresponding token sequences\. The first term uses centered kernel alignment \(CKA\) to preserve the global relational geometry between these two spaces\. Each entry ofSzS^\{z\}is the pairwise cosine similarity between two fused domain representations, while each entry ofSAS^\{A\}is the corresponding symmetrized cosine similarity after their soft tokens pass through the frozen query and key projections of the language model\. The coefficientλattn≥0\\lambda\_\{\\mathrm\{attn\}\}\\geq 0controls this attention\-space alignment, and∥⋅∥F\\\|\\cdot\\\|\_\{F\}denotes the Frobenius norm\. The second term therefore preserves source\-domain relations at the actual interface through which the language model consumes the injected tokens\.

The fusion module and projector are jointly optimized with the episodic query\-generation loss, together with the held\-out reconstruction and relation\-alignment objectives\. During each meta\-training episode,TdT\_\{d\}is injected into both the support and query inputs of the pseudo\-target task and remains fixed throughout within\-task adaptation, ensuring that Meta\-LoRA is trained and evaluated under the same domain condition\. For an unseen target domaindtd\_\{t\}, the same retrieval and composition mechanism producesz¯t\\bar\{z\}\_\{t\}, which is projected intoTt=Πψ​\(z¯t\)T\_\{t\}=\\Pi\_\{\\psi\}\(\\bar\{z\}\_\{t\}\)and injected together with the textual user prompt\. Unlike standard retrieve\-and\-project pipelines, our design preserves source\-domain relation structure throughout retrieval, target\-conditioned composition, and projection into the language model’s attention space\. Feature construction and graph estimation are detailed in Appendix D\.

#### User\-Level Textual Conditioning\.

A source\-domain trained summarizer converts user’s cross\-domain historyℋu\\mathcal\{H\}\_\{u\}into a textual preference promptPuP\_\{u\}\. Trained with held\-out personalization utility and cross\-view consistency, it emphasizes persistent preferences while suppressing source\-specific or transient patterns\. We prependPuP\_\{u\}to support and query inputs during meta\-training, and compute it once before target adaptation without using current responses\. The textual prompt provides interpretable user context, complementing the graph\-derived domain tokensTdT\_\{d\}\. Source\-only summarizer training is detailed in Appendix E\.

### Unified Zero\- and Few\-Shot Inference

Given a target support set𝒮t=\{\(xi,yi\)\}i=1K\\mathcal\{S\}\_\{t\}=\\\{\(x\_\{i\},y\_\{i\}\)\\\}\_\{i=1\}^\{K\}, the target\-domain description and the available support inputs\{xi\}i=1K\\\{x\_\{i\}\\\}\_\{i=1\}^\{K\}are encoded in the domain representation space to construct the uncertainty\-aware target\-evidence representationz^t\\widehat\{z\}\_\{t\}\. Equation \([6](https://arxiv.org/html/2608.12389#Sx3.E6)\) retrieves and composes relevant source\-domain representations intoz¯t\\bar\{z\}\_\{t\}, which is projected by Eq\. \([7](https://arxiv.org/html/2608.12389#Sx3.E7)\) intoTtT\_\{t\}\. In parallel, user’s cross\-domain historyℋu\\mathcal\{H\}\_\{u\}is summarized into the textual preference promptPuP\_\{u\}\. The current support responses\{yi\}i=1K\\\{y\_\{i\}\\\}\_\{i=1\}^\{K\}are used only to compute the support\-based adaptation loss and task statistics\. The prompt and domain tokens are computed once per target episode and held fixed across all inner\-loop updates\.

WhenKK=0, the model generates directly from\[Pu;Tt;xt\]\[P\_\{u\};T\_\{t\};x\_\{t\}\]usingθ0\\theta\_\{0\}\. WhenKK\>0, the support examples are conditioned on the same prompt and domain tokens,H~​\(𝒮t;θ0\)\\widetilde\{H\}\(\\mathcal\{S\}\_\{t\};\\theta\_\{0\}\)and\{F^l\}\\\{\\widehat\{F\}\_\{l\}\\\}are estimated once atθ0\\theta\_\{0\}, andMMupdates ofℒadapt\\mathcal\{L\}\_\{\\mathrm\{adapt\}\}produce the task\-adapted parametersθ∗\\theta^\{\\ast\}, which are used to generate the final response from\[Pu;Tt;xt\]\[P\_\{u\};T\_\{t\};x\_\{t\}\]\. This shared conditioning interface supports zero\-shot generation and few\-shot evidence\-calibrated refinement within the same framework\.

## Experiments

Table 1:Cross\-domain 5\-shot results on HiCUPID\.ΔWR↓=WRin−Avg\(WRtarget\)\\Delta\\mathrm\{WR\}\\downarrow=\\mathrm\{WR\}\_\{\\mathrm\{in\}\}\-\\operatorname\{Avg\}\(\\mathrm\{WR\}\_\{\\mathrm\{target\}\}\)measures cross\-domain degradation\.#### Experimental Setup\.

We evaluate on HiCUPID\(Moket al\.[2025a](https://arxiv.org/html/2608.12389#bib.bib110)\)and LaMP\-QA\(Salemi and Zamani[2025](https://arxiv.org/html/2608.12389#bib.bib44)\), using disjoint source domains for meta\-training and held\-out domains for evaluation\. We use LLaMA\-3\-8B\-Instruct as the default backbone and Qwen3\-8B for cross\-model evaluation\. The backbone is frozen\. Source\-domain training optimizes the method\-specific components\. Before target adaptation, all components except the LoRA parameters are frozen\. Hyperparameters are selected only on source\-domain validation episodes, with identical support splits and decoding settings across methods\. Complete protocols are in Appendix G\.

#### Metrics and Baselines\.

Win Rate \(WR\), our primary metric, is the percentage of test instances for which the pairwise evaluator prefers the generated personalized response over the corresponding reference response\. CCS aggregates utility, honesty, and personal fit, while PG \(Personal Gap\) measures the semantic distinction from a generic response\. We defineΔ​WR\\Delta\\mathrm\{WR\}as source\-domain WR minus mean target\-domain WR\. HiCUPID uses its official evaluator, whereas the LaMP\-QA evaluator is validated against blinded human judgments\. Baselines cover retrieval/text conditioning \(LaMP\(Salemiet al\.[2024b](https://arxiv.org/html/2608.12389#bib.bib49)\)and GPG\(Zhang[2024](https://arxiv.org/html/2608.12389#bib.bib48)\)\), continuous prompts \(SPT\(Huanget al\.[2024](https://arxiv.org/html/2608.12389#bib.bib54)\)and DEP\(Qiuet al\.[2025b](https://arxiv.org/html/2608.12389#bib.bib46)\)\), parameter adaptation \(IDL\(Chenget al\.[2024](https://arxiv.org/html/2608.12389#bib.bib56)\), OPPU\(Tanet al\.[2024b](https://arxiv.org/html/2608.12389#bib.bib50)\), PROPER\(Zhanget al\.[2025a](https://arxiv.org/html/2608.12389#bib.bib52)\), and CoPL\(Choiet al\.[2025b](https://arxiv.org/html/2608.12389#bib.bib45)\)\), inference\-time personalization \(BiPO\(Caoet al\.[2024a](https://arxiv.org/html/2608.12389#bib.bib57)\), CHAMELEON\(Zhanget al\.[2025c](https://arxiv.org/html/2608.12389#bib.bib58)\), OPAD\(Zhuet al\.[2025](https://arxiv.org/html/2608.12389#bib.bib104)\), and Proactive\(Zhanget al\.[2025b](https://arxiv.org/html/2608.12389#bib.bib105)\)\), and Model\-Agnostic Meta\-Learning \(MAML\(Finnet al\.[2017a](https://arxiv.org/html/2608.12389#bib.bib55)\)\)\. Where applicable, baselines use the same target support sets\. All methods use matched backbones, decoding settings, and evaluation instances\. Reproducibility, limitations, and other details are in Appendix I\.

Table 2:Performance on LaMP\-QA under the cross\-domain 5\-shot personalization setting\. WR is reported as mean±\\pmstd\.Table 3:Performance on theQwen3\-8Bbackbone\.Table 4:Performance in theUnseen Userscenario\.
#### Main Results\.

Table[1](https://arxiv.org/html/2608.12389#Sx4.T1)shows that our method achieves the strongest 5\-shot cross\-domain personalization performance on HiCUPID\. It obtains the best WR and CCS on all five held\-out target domains\. More importantly, it reducesΔ​WR\\Delta\\mathrm\{WR\}to5\.985\.98, corresponding to a47\.9%47\.9\\%relative reduction in cross\-domain degradation\. These results indicate that the improvement is not merely due to stronger source\-domain fitting, but reflects substantially more stable transfer of personalized behavior to unseen domains\.

The advantage also generalizes across datasets and backbone architectures\. On LaMP\-QA, Table[2](https://arxiv.org/html/2608.12389#Sx4.T2)shows that our method improves WR over PROPER by17\.3%17\.3\\%on Parenting and14\.7%14\.7\\%on Interpersonal, while reducingΔ​WR\\Delta\\mathrm\{WR\}from12\.7512\.75to5\.455\.45\. On Qwen3\-8B, Table[3](https://arxiv.org/html/2608.12389#Sx4.T3)reports a cross\-domain WR of70\.970\.9, a21\.2%21\.2\\%relative improvement over PROPER, together with the smallest degradation of6\.76\.7\. The consistent gains demonstrate that the framework is not tied to a particular task format or model architecture\.

The unseen\-user setting is more challenging because the target user has no personal interaction history, leaving the model without direct user evidence\. As shown in Table[4](https://arxiv.org/html/2608.12389#Sx4.T4), our method achieves a WR of43\.543\.5, more than twice that of the strongest baseline\. In this extreme cold\-start regime, the graph\-guided domain\-prior construction supplies structured target\-domain conditions from related source domains, helping the model avoid generic responses\.

Table 5:Ablation results averaged over five target domains\.
#### Ablation Study\.

Table[5](https://arxiv.org/html/2608.12389#Sx4.T5)provides cumulative attribution and component\-wise ablations\. Starting from textual conditioning, adding domain soft tokens improves 5\-shot WR from59\.359\.3to63\.463\.4, vanilla Meta\-LoRA further raises it to68\.468\.4, and evidence\-calibrated adaptation reaches73\.273\.2\. The corresponding 0\-shot gain from55\.855\.8to59\.559\.5shows that domain conditioning also improves transfer before target updates\. Among its components, removingK−1K^\{\-1\}scaling causes the largest degradation\. Removing graph reliability or topology preservation also degrades both regimes, confirming that the improvement does not arise from inserting arbitrary soft tokens\. Because each ablated variant is meta\-trained from scratch, changes to the inner\-loop objective can also alter the learned initializationθ0\\theta\_\{0\}and thus affect zero\-shot performance\. Exact ablation definitions and additional attribution are in Appendix H\.

![Refer to caption](https://arxiv.org/html/2608.12389v1/x3.png)Figure 3:Sample efficiency across support sizes\.
#### Sample Efficiency and Adaptation Analysis\.

Figure[3](https://arxiv.org/html/2608.12389#Sx4.F3)compares SFT, MAML, and our method across different support sizes\. Our method achieves the highest WR throughoutK∈\{0,1,3,5,10\}K\\in\\\{0,1,3,5,10\\\}\. Its WR increases from54\.754\.7in the zero\-shot setting to66\.966\.9with a single target\-domain interaction, and reaches approximately90%90\\%of its 10\-shot performance with only three interactions\. These results indicate that episodic meta\-learning provides an initialization capable of rapidly exploiting sparse target\-user evidence, while controlled adaptation mitigates unstable updates\.

#### Case Study\.

Table[6](https://arxiv.org/html/2608.12389#Sx4.T6)presents an illustrative example\. Given the same query and source\-domain history, MAML captures a coarse domain\-relevant interest, whereas our method integrates multiple history\-supported preferences into a coherent response\.

Table 6:Illustrative cross\-domain case\. The cues are not provided as oracle inputs and are shown only for analysis\.
#### Extreme Domain\-Shift Robustness\.

Using an independent external encoder, we represent each domain by theℓ2\\ell\_\{2\}\-normalized mean embeddingμd\\mu\_\{d\}of an equal\-sized held\-out set of interactions\. We define target–source proximity asst=maxds∈𝒟s⁡cos⁡\(μdt,μds\)s\_\{t\}=\\max\_\{d\_\{s\}\\in\\mathcal\{D\}\_\{s\}\}\\cos\(\\mu\_\{d\_\{t\}\},\\mu\_\{d\_\{s\}\}\)and domain shift as1−st1\-s\_\{t\}\. LaMP\-QA Philosophy and HiCUPID Food have the lowest target–source proximities \(0\.156/0\.3020\.156/0\.302\)\. Our method improves WR/CCS over the strongest baseline by15\.4%/13\.9%15\.4\\%/13\.9\\%and9\.6%/21\.6%9\.6\\%/21\.6\\%, respectively, showing that its gains persist under the largest observed shifts\. Experimental details and supplementary analyses \(latency, metric reliability, topology preservation, etc\.\) are provided in Appendix H\.

## Conclusion

We study cross\-domain zero\- and few\-shot LLM personalization under sparse target\-domain evidence\. Our framework combines evidence\-calibrated Meta\-LoRA, which regulates adaptation through PAC\-Bayes\-guided anchoring, with graph\-guided domain injection that preserves domain relations through soft tokens while retaining interpretable textual user prompts\. Experiments show consistent gains across domains, backbones, support sizes, and severe cold\-start settings\. Ablation and structural analyses further validate the contributions of adaptive regularization and topology\-preserving domain transfer\.

## References

- The statistical analysis of compositional data\.Chapman and Hall\.Cited by:[§D\.2](https://arxiv.org/html/2608.12389#A4.SS2.p1.2)\.
- N\. Balepur, V\. Padmakumar, F\. Yang, S\. Feng, R\. Rudinger, and J\. Boyd\-Graber \(2025\)Whose boat does it float? improving personalization in preference tuning via inferred user personas\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Vienna, Austria,pp\. 3371–3393\.External Links:[Link](https://aclanthology.org/2025.acl-long.168/),[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.168)Cited by:[LLM Personalization\.](https://arxiv.org/html/2608.12389#Sx2.SS0.SSS0.Px1.p1.1)\.
- H\. Bu, C\. Jung, M\. Kang, and J\. Kim \(2025\)Personalized LLM decoding via contrasting personal preference\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,Suzhou, China,pp\. 33958–33978\.External Links:[Link](https://aclanthology.org/2025.emnlp-main.1723/),[Document](https://dx.doi.org/10.18653/v1/2025.emnlp-main.1723)Cited by:[LLM Personalization\.](https://arxiv.org/html/2608.12389#Sx2.SS0.SSS0.Px1.p1.1)\.
- Y\. Cao, T\. Zhang, B\. Cao, Z\. Yin, L\. Lin, F\. Ma, and J\. Chen \(2024a\)Personalized steering of large language models: versatile steering vectors through bi\-directional preference optimization\.Advances in Neural Information Processing Systems37,pp\. 49519–49551\.Cited by:[Introduction](https://arxiv.org/html/2608.12389#Sx1.p3.1),[Metrics and Baselines\.](https://arxiv.org/html/2608.12389#Sx4.SSx3.SSS0.Px2.p1.1)\.
- Y\. Cao, T\. Zhang, B\. Cao, Z\. Yin, L\. Lin, F\. Ma, and J\. Chen \(2024b\)Personalized steering of large language models: versatile steering vectors through bi\-directional preference optimization\.InAdvances in Neural Information Processing Systems,Vol\.37,pp\. 49519–49551\.External Links:[Document](https://dx.doi.org/10.52202/079017-1567),[Link](https://proceedings.neurips.cc/paper_files/paper/2024/hash/58cbe393b4254da8966780a40d023c0b-Abstract-Conference.html)Cited by:[10th item](https://arxiv.org/html/2608.12389#A7.I1.i10.p1.1),[LLM Personalization\.](https://arxiv.org/html/2608.12389#Sx2.SS0.SSS0.Px1.p1.1)\.
- R\. Chen, X\. Zhang, M\. Luo, W\. Chai, and Z\. Liu \(2024\)Pad: personalized alignment of llms at decoding\-time\.arXiv preprint arXiv:2410\.04070\.Cited by:[Introduction](https://arxiv.org/html/2608.12389#Sx1.p3.1)\.
- C\. Cheng, Q\. Tu, S\. Shang, C\. Mao, Z\. Yu, W\. Wu, and R\. Yan \(2024\)" In dialogues we learn": towards personalized dialogue without pre\-defined profiles through in\-dialogue learning\.arXiv preprint arXiv:2403\.03102\.Cited by:[6th item](https://arxiv.org/html/2608.12389#A7.I1.i6.p1.1),[Metrics and Baselines\.](https://arxiv.org/html/2608.12389#Sx4.SSx3.SSS0.Px2.p1.1)\.
- Y\. Choi, S\. Cho, M\. Lee, M\. Park, Y\. Ko, J\. Ok, and D\. Kim \(2025a\)CoPL: collaborative preference learning for personalizing LLMs\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,Suzhou, China,pp\. 12875–12893\.External Links:[Link](https://aclanthology.org/2025.emnlp-main.650/),[Document](https://dx.doi.org/10.18653/v1/2025.emnlp-main.650)Cited by:[9th item](https://arxiv.org/html/2608.12389#A7.I1.i9.p1.1),[Introduction](https://arxiv.org/html/2608.12389#Sx1.p2.1),[LLM Personalization\.](https://arxiv.org/html/2608.12389#Sx2.SS0.SSS0.Px1.p1.1)\.
- Y\. Choi, S\. Cho, M\. Lee, M\. Park, Y\. Ko, J\. Ok, and D\. Kim \(2025b\)CoPL: collaborative preference learning for personalizing llms\.arXiv preprint arXiv:2503\.01658\.Cited by:[Introduction](https://arxiv.org/html/2608.12389#Sx1.p3.1),[Metrics and Baselines\.](https://arxiv.org/html/2608.12389#Sx4.SSx3.SSS0.Px2.p1.1)\.
- M\. Cuturi and A\. Doucet \(2014\)Fast computation of wasserstein barycenters\.InProceedings of the 31st International Conference on Machine Learning,Cited by:[§D\.4](https://arxiv.org/html/2608.12389#A4.SS4.p1.6)\.
- M\. Cuturi \(2013\)Sinkhorn distances: lightspeed computation of optimal transport\.InAdvances in Neural Information Processing Systems,Cited by:[§D\.4](https://arxiv.org/html/2608.12389#A4.SS4.p1.6)\.
- C\. Finn, P\. Abbeel, and S\. Levine \(2017a\)Model\-agnostic meta\-learning for fast adaptation of deep networks\.InInternational conference on machine learning,pp\. 1126–1135\.Cited by:[Metrics and Baselines\.](https://arxiv.org/html/2608.12389#Sx4.SSx3.SSS0.Px2.p1.1)\.
- C\. Finn, P\. Abbeel, and S\. Levine \(2017b\)Model\-agnostic meta\-learning for fast adaptation of deep networks\.InProceedings of the 34th International Conference on Machine Learning,D\. Precup and Y\. W\. Teh \(Eds\.\),Proceedings of Machine Learning Research, Vol\.70,pp\. 1126–1135\.External Links:[Link](https://proceedings.mlr.press/v70/finn17a.html)Cited by:[§B\.3](https://arxiv.org/html/2608.12389#A2.SS3.p1.3),[14th item](https://arxiv.org/html/2608.12389#A7.I1.i14.p1.1),[Meta\-Learning and Controlled Adaptation\.](https://arxiv.org/html/2608.12389#Sx2.SS0.SSS0.Px2.p1.1)\.
- C\. Gao, Y\. Huang, J\. Yao, X\. Wu, and J\. Feng \(2026\)Beyond static profiles: capturing the fluidity of user preferences in diverse scenarios\.InFindings of the Association for Computational Linguistics: ACL 2026,San Diego, California, United States,pp\. 20618–20634\.External Links:[Link](https://aclanthology.org/2026.findings-acl.1033/),[Document](https://dx.doi.org/10.18653/v1/2026.findings-acl.1033)Cited by:[Introduction](https://arxiv.org/html/2608.12389#Sx1.p1.1),[Introduction](https://arxiv.org/html/2608.12389#Sx1.p2.1)\.
- E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, W\. Chen,et al\.\(2022\)Lora: low\-rank adaptation of large language models\.\.ICLR1\(2\),pp\. 3\.Cited by:[§B\.1](https://arxiv.org/html/2608.12389#A2.SS1.p2.4),[Meta\-Learning and Controlled Adaptation\.](https://arxiv.org/html/2608.12389#Sx2.SS0.SSS0.Px2.p1.1)\.
- Q\. Huang, X\. Liu, T\. Ko, B\. Wu, W\. Wang, Y\. Zhang, and L\. Tang \(2024\)Selective prompting tuning for personalized conversations with llms\.arXiv preprint arXiv:2406\.18187\.Cited by:[4th item](https://arxiv.org/html/2608.12389#A7.I1.i4.p1.1),[Metrics and Baselines\.](https://arxiv.org/html/2608.12389#Sx4.SSx3.SSS0.Px2.p1.1)\.
- M\. Kim, K\. Lee, S\. Joo, H\. Lee, T\. Thonet, and K\. Jung \(2025\)Drift: decoding\-time personalized alignments with implicit user preferences\.arXiv preprint arXiv:2502\.14289\.Cited by:[Introduction](https://arxiv.org/html/2608.12389#Sx1.p3.1)\.
- J\. Kirkpatrick, R\. Pascanu, N\. Rabinowitz, J\. Veness, G\. Desjardins, A\. A\. Rusu, K\. Milan, J\. Quan, T\. Ramalho, A\. Grabska\-Barwinska,et al\.\(2017\)Overcoming catastrophic forgetting in neural networks\.Proceedings of the National Academy of Sciences114\(13\),pp\. 3521–3526\.Cited by:[§C\.6](https://arxiv.org/html/2608.12389#A3.SS6.p1.8)\.
- S\. Kornblith, M\. Norouzi, H\. Lee, and G\. Hinton \(2019\)Similarity of neural network representations revisited\.InProceedings of the 36th International Conference on Machine Learning,Cited by:[§D\.7](https://arxiv.org/html/2608.12389#A4.SS7.p1.3)\.
- R\. Krishnan and O\. Tickoo \(2020\)Improving model calibration with accuracy versus uncertainty optimization\.Advances in Neural Information Processing Systems33,pp\. 18237–18248\.Cited by:[§C\.5](https://arxiv.org/html/2608.12389#A3.SS5.p1.6),[Evidence\- and Structure\-Aware Adaptation Objective\.](https://arxiv.org/html/2608.12389#Sx3.SSx1.SSS0.Px2.p1.4)\.
- B\. Lester, R\. Al\-Rfou, and N\. Constant \(2021\)The power of scale for parameter\-efficient prompt tuning\.InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing,Online and Punta Cana, Dominican Republic,pp\. 3045–3059\.External Links:[Link](https://aclanthology.org/2021.emnlp-main.243/),[Document](https://dx.doi.org/10.18653/v1/2021.emnlp-main.243)Cited by:[4th item](https://arxiv.org/html/2608.12389#A7.I1.i4.p1.1),[LLM Personalization\.](https://arxiv.org/html/2608.12389#Sx2.SS0.SSS0.Px1.p1.1)\.
- X\. L\. Li and P\. Liang \(2021\)Prefix\-tuning: optimizing continuous prompts for generation\.InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing \(Volume 1: Long Papers\),Online,pp\. 4582–4597\.External Links:[Link](https://aclanthology.org/2021.acl-long.353/),[Document](https://dx.doi.org/10.18653/v1/2021.acl-long.353)Cited by:[LLM Personalization\.](https://arxiv.org/html/2608.12389#Sx2.SS0.SSS0.Px1.p1.1)\.
- C\. Ling, X\. Zhao, J\. Lu, C\. Deng, C\. Zheng, J\. Wang, T\. Chowdhury, Y\. Li, H\. Cui, X\. Zhang,et al\.\(2025\)Domain specialization as the key to make large language models disruptive: a comprehensive survey\.ACM Computing Surveys58\(3\),pp\. 1–39\.Cited by:[Introduction](https://arxiv.org/html/2608.12389#Sx1.p1.1),[Introduction](https://arxiv.org/html/2608.12389#Sx1.p2.1)\.
- 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\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Vienna, Austria,pp\. 9373–9385\.External Links:[Link](https://aclanthology.org/2025.acl-long.461/),[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.461)Cited by:[Introduction](https://arxiv.org/html/2608.12389#Sx1.p1.1),[LLM Personalization\.](https://arxiv.org/html/2608.12389#Sx2.SS0.SSS0.Px1.p1.1)\.
- A\. Maurer \(2004\)A note on the pac bayesian theorem\.arXiv preprint cs/0411099\.Cited by:[§C\.1](https://arxiv.org/html/2608.12389#A3.SS1.p1.2),[§C\.2](https://arxiv.org/html/2608.12389#A3.SS2.SSS0.Px3.p1.2),[Meta\-Learned Prior and PAC\-Bayes\-Guided Constraint\.](https://arxiv.org/html/2608.12389#Sx3.SSx1.SSS0.Px1.p2.10)\.
- D\. A\. McAllester \(1999\)Some PAC\-Bayesian theorems\.Machine Learning37\(3\),pp\. 355–363\.External Links:[Document](https://dx.doi.org/10.1023/A%3A1007618624809),[Link](https://doi.org/10.1023/A:1007618624809)Cited by:[§C\.1](https://arxiv.org/html/2608.12389#A3.SS1.p1.2),[Meta\-Learning and Controlled Adaptation\.](https://arxiv.org/html/2608.12389#Sx2.SS0.SSS0.Px2.p1.1)\.
- J\. Mok, I\. Kim, S\. Park, and S\. Yoon \(2025a\)Exploring the potential of LLMs as personalized assistants: dataset, evaluation, and analysis\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Vienna, Austria,pp\. 10212–10239\.External Links:[Link](https://aclanthology.org/2025.acl-long.504/),[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.504)Cited by:[§G\.1](https://arxiv.org/html/2608.12389#A7.SS1.p1.1),[§G\.3](https://arxiv.org/html/2608.12389#A7.SS3.SSS0.Px3.p1.4),[§G\.4](https://arxiv.org/html/2608.12389#A7.SS4.p1.1),[Experimental Setup\.](https://arxiv.org/html/2608.12389#Sx4.SSx3.SSS0.Px1.p1.1)\.
- J\. Mok, I\. Kim, S\. Park, and S\. Yoon \(2025b\)Exploring the potential of llms as personalized assistants: dataset, evaluation, and analysis\.arXiv preprint arXiv:2506\.01262\.Cited by:[Introduction](https://arxiv.org/html/2608.12389#Sx1.p1.1)\.
- H\. Nam, Y\. Wan, M\. Liu, J\. Lian, P\. Ahnn, and N\. Jaques \(2025\)Learning to summarize user information for personalized reinforcement learning from human feedback\.arXiv preprint arXiv:2507\.13579\.Cited by:[Introduction](https://arxiv.org/html/2608.12389#Sx1.p3.1)\.
- Y\. Qiu, T\. Shi, X\. Zhao, F\. Zhu, Y\. Zhang, and F\. Feng \(2025a\)Latent inter\-user difference modeling for LLM personalization\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,Suzhou, China,pp\. 10599–10617\.External Links:[Link](https://aclanthology.org/2025.emnlp-main.536/),[Document](https://dx.doi.org/10.18653/v1/2025.emnlp-main.536)Cited by:[5th item](https://arxiv.org/html/2608.12389#A7.I1.i5.p1.1),[LLM Personalization\.](https://arxiv.org/html/2608.12389#Sx2.SS0.SSS0.Px1.p1.1)\.
- Y\. Qiu, T\. Shi, X\. Zhao, F\. Zhu, Y\. Zhang, and F\. Feng \(2025b\)Latent inter\-user difference modeling for llm personalization\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 10610–10628\.Cited by:[Metrics and Baselines\.](https://arxiv.org/html/2608.12389#Sx4.SSx3.SSS0.Px2.p1.1)\.
- Qwen Team \(2025\)Qwen3 technical report\.arXiv preprint arXiv:2505\.09388\.Cited by:[§G\.4](https://arxiv.org/html/2608.12389#A7.SS4.p2.1),[Table 10](https://arxiv.org/html/2608.12389#A7.T10.49.51.1.3.1.1),[§H\.2](https://arxiv.org/html/2608.12389#A8.SS2.p1.1)\.
- N\. Reimers and I\. Gurevych \(2019\)Sentence\-BERT: sentence embeddings using siamese BERT\-networks\.InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing,Cited by:[§G\.3](https://arxiv.org/html/2608.12389#A7.SS3.SSS0.Px2.p1.4),[§H\.8](https://arxiv.org/html/2608.12389#A8.SS8.p1.2)\.
- C\. Riou, P\. Alquier, and B\. Chérief\-Abdellatif \(2023\)Bayes meets bernstein at the meta level: an analysis of fast rates in meta\-learning with pac\-bayes\.arXiv preprint arXiv:2302\.11709\.Cited by:[Appendix C](https://arxiv.org/html/2608.12389#A3.p1.1)\.
- A\. Salemi, S\. Mysore, M\. Bendersky, and H\. Zamani \(2024a\)LaMP: when large language models meet personalization\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Bangkok, Thailand,pp\. 7370–7392\.External Links:[Link](https://aclanthology.org/2024.acl-long.399/),[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.399)Cited by:[2nd item](https://arxiv.org/html/2608.12389#A7.I1.i2.p1.1),[LLM Personalization\.](https://arxiv.org/html/2608.12389#Sx2.SS0.SSS0.Px1.p1.1)\.
- A\. Salemi, S\. Mysore, M\. Bendersky, and H\. Zamani \(2024b\)Lamp: when large language models meet personalization\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 7370–7392\.Cited by:[Introduction](https://arxiv.org/html/2608.12389#Sx1.p3.1),[Metrics and Baselines\.](https://arxiv.org/html/2608.12389#Sx4.SSx3.SSS0.Px2.p1.1)\.
- A\. Salemi and H\. Zamani \(2025\)LaMP\-qa: a benchmark for personalized long\-form question answering\.arXiv preprint arXiv:2506\.00137\.Cited by:[§G\.1](https://arxiv.org/html/2608.12389#A7.SS1.p1.1),[§G\.4](https://arxiv.org/html/2608.12389#A7.SS4.p1.1),[Introduction](https://arxiv.org/html/2608.12389#Sx1.p1.1),[Experimental Setup\.](https://arxiv.org/html/2608.12389#Sx4.SSx3.SSS0.Px1.p1.1)\.
- M\. Seeger \(2002\)PAC\-bayesian generalisation error bounds for gaussian process classification\.Journal of machine learning research3\(Oct\),pp\. 233–269\.Cited by:[§C\.1](https://arxiv.org/html/2608.12389#A3.SS1.p1.2)\.
- Z\. Shao, P\. Wang, Q\. Zhu, R\. Xu, J\. Song, X\. Bi, H\. Zhang, M\. Zhang, Y\. Li, Y\. Wu,et al\.\(2024\)Deepseekmath: pushing the limits of mathematical reasoning in open language models\.arXiv preprint arXiv:2402\.03300\.Cited by:[§E\.3](https://arxiv.org/html/2608.12389#A5.SS3.p1.4)\.
- A\. Singh, S\. Hsu, K\. Hsu, E\. Mitchell, S\. Ermon, T\. Hashimoto, A\. Sharma, and C\. Finn \(2025\)Fspo: few\-shot preference optimization of synthetic preference data in llms elicits effective personalization to real users\.arXiv preprint arXiv:2502\.19312\.Cited by:[Introduction](https://arxiv.org/html/2608.12389#Sx1.p3.1)\.
- Z\. Tan, Q\. Zeng, Y\. Tian, Z\. Liu, B\. Yin, and M\. Jiang \(2024a\)Democratizing large language models via personalized parameter\-efficient fine\-tuning\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Miami, Florida, USA,pp\. 6476–6491\.External Links:[Link](https://aclanthology.org/2024.emnlp-main.372/),[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-main.372)Cited by:[7th item](https://arxiv.org/html/2608.12389#A7.I1.i7.p1.1),[Introduction](https://arxiv.org/html/2608.12389#Sx1.p2.1),[LLM Personalization\.](https://arxiv.org/html/2608.12389#Sx2.SS0.SSS0.Px1.p1.1),[Meta\-Learning and Controlled Adaptation\.](https://arxiv.org/html/2608.12389#Sx2.SS0.SSS0.Px2.p1.1)\.
- Z\. Tan, Q\. Zeng, Y\. Tian, Z\. Liu, B\. Yin, and M\. Jiang \(2024b\)Democratizing large language models via personalized parameter\-efficient fine\-tuning\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,pp\. 6476–6491\.Cited by:[Metrics and Baselines\.](https://arxiv.org/html/2608.12389#Sx4.SSx3.SSS0.Px2.p1.1)\.
- Z\. Tan, Z\. Zhang, H\. Wen, Z\. Li, R\. Zhang, P\. Chen, F\. Mo, Z\. Liu, Q\. Zeng, Q\. Yin,et al\.\(2025\)Instant personalized large language model adaptation via hypernetwork\.arXiv preprint arXiv:2510\.16282\.Cited by:[Introduction](https://arxiv.org/html/2608.12389#Sx1.p3.1)\.
- Z\. Wang, Z\. Li, Z\. Jiang, D\. Tu, and W\. Shi \(2024\)Crafting personalized agents through retrieval\-augmented generation on editable memory graphs\.arXiv preprint arXiv:2409\.19401\.Cited by:[Introduction](https://arxiv.org/html/2608.12389#Sx1.p3.1)\.
- J\. Zhang \(2024\)Guided profile generation improves personalization with llms\.arXiv preprint arXiv:2409\.13093\.Cited by:[3rd item](https://arxiv.org/html/2608.12389#A7.I1.i3.p1.1),[Introduction](https://arxiv.org/html/2608.12389#Sx1.p3.1),[Metrics and Baselines\.](https://arxiv.org/html/2608.12389#Sx4.SSx3.SSS0.Px2.p1.1)\.
- L\. Zhang, J\. Wu, D\. Zhou, and Y\. He \(2025a\)PROPER: a progressive learning framework for personalized large language models with group\-level adaptation\.arXiv preprint arXiv:2503\.01303\.Cited by:[8th item](https://arxiv.org/html/2608.12389#A7.I1.i8.p1.1),[Introduction](https://arxiv.org/html/2608.12389#Sx1.p3.1),[Metrics and Baselines\.](https://arxiv.org/html/2608.12389#Sx4.SSx3.SSS0.Px2.p1.1)\.
- W\. Zhang, P\. Zhang, B\. Zhang, X\. Wang, and D\. Wang \(2023\)A collaborative transfer learning framework for cross\-domain recommendation\.InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining,pp\. 5576–5585\.Cited by:[Introduction](https://arxiv.org/html/2608.12389#Sx1.p2.1)\.
- X\. Zhang, Y\. Wang, R\. Chen, Z\. Wang, R\. Hou, and Z\. Liu \(2025b\)Towards proactive personalization through profile customization for individual users in dialogues\.arXiv preprint arXiv:2512\.15302\.Cited by:[13rd item](https://arxiv.org/html/2608.12389#A7.I1.i13.p1.1),[Metrics and Baselines\.](https://arxiv.org/html/2608.12389#Sx4.SSx3.SSS0.Px2.p1.1)\.
- Y\. Zhang, D\. Adila, C\. Shin, and F\. Sala \(2025c\)Personalize your llm: fake it then align it\.arXiv preprint arXiv:2503\.01048\.Cited by:[11st item](https://arxiv.org/html/2608.12389#A7.I1.i11.p1.1),[Metrics and Baselines\.](https://arxiv.org/html/2608.12389#Sx4.SSx3.SSS0.Px2.p1.1)\.
- L\. Zheng, W\. Chiang, Y\. Sheng, S\. Zhuang, Z\. Wu, Y\. Zhuang, Z\. Lin, Z\. Li, D\. Li, E\. P\. Xing,et al\.\(2023\)Judging LLM\-as\-a\-judge with MT\-Bench and chatbot arena\.Advances in Neural Information Processing Systems36\.Cited by:[§G\.3](https://arxiv.org/html/2608.12389#A7.SS3.SSS0.Px1.p1.4)\.
- M\. Zhu, Y\. Liu, L\. Zhang, J\. Guo, and Z\. Mao \(2025\)On\-the\-fly preference alignment via principle\-guided decoding\.arXiv preprint arXiv:2502\.14204\.Cited by:[12nd item](https://arxiv.org/html/2608.12389#A7.I1.i12.p1.1),[Introduction](https://arxiv.org/html/2608.12389#Sx1.p3.1),[Metrics and Baselines\.](https://arxiv.org/html/2608.12389#Sx4.SSx3.SSS0.Px2.p1.1)\.

## Appendix

## Appendix ASupplementary Contents

Appendix B\.Task definition, source episodes, information boundaries, and notation\.

Appendix C\.PAC\-Bayes assumptions, derivations, deterministic surrogate, and computational cost\.

Appendix D\.Multi\-view domain features, reliability graph, Wasserstein fusion, retrieval, and topology\-preserving projection\.

Appendix E\.Source\-only user\-preference summarizer training and target\-time use\.

Appendix F\.Complete source\-training and zero\-/few\-shot inference algorithms\.

Appendix G\.Datasets, baselines, metrics, evaluators, hyperparameters, and statistical protocol\.

Appendix H\.Ablations, cross\-backbone results, sample efficiency, optimization stability, topology, attention attribution, domain shift, latency, evaluator agreement, sensitivity, and case study\.

Appendix I\.Reproducibility package, limitations, privacy, and safety\.

## Appendix BTask Definition and Information Boundaries

### B\.1Cross\-Domain Personalization Task

Let𝒟s\\mathcal\{D\}\_\{s\}be the set of source domains available during training and letdt∉𝒟sd\_\{t\}\\notin\\mathcal\{D\}\_\{s\}be a held\-out target domain\. For a target useruu,ℋu\\mathcal\{H\}\_\{u\}contains only pre\-target interactions from source domains and may be empty in the unseen\-user evaluation\. The optional target support set is𝒮t=\{\(xi,yi\)\}i=1K\\mathcal\{S\}\_\{t\}=\\\{\(x\_\{i\},y\_\{i\}\)\\\}\_\{i=1\}^\{K\}, whereK=0K=0denotes zero\-shot inference andK\>0K\>0denotes few\-shot adaptation\. The test queryxtx\_\{t\}and its responseyty\_\{t\}are disjoint from𝒮t\\mathcal\{S\}\_\{t\}\. The target\-domain description is public task metadata rather than user supervision\.

The framework constructs a textual user promptPuP\_\{u\}, a target\-domain soft\-token sequenceTtT\_\{t\}, and, whenK\>0K\>0, adapted low\-rank parameters\(Huet al\.[2022](https://arxiv.org/html/2608.12389#bib.bib39)\)θt∗\\theta\_\{t\}^\{\\ast\}\. The generated response is

y^t=\{G​\(\[Pu;Tt;xt\];θ0\),K=0,G​\(\[Pu;Tt;xt\];θt∗\),K\>0,\\widehat\{y\}\_\{t\}=\\begin\{cases\}G\(\[P\_\{u\};T\_\{t\};x\_\{t\}\];\\theta\_\{0\}\),&K=0,\\\\ G\(\[P\_\{u\};T\_\{t\};x\_\{t\}\];\\theta\_\{t\}^\{\\ast\}\),&K\>0,\\end\{cases\}\(8\)where the backbone language\-model parameters are frozen and only LoRA parameters are changed during target adaptation\.

### B\.2Functional Factorization and Claim Boundary

The user/domain decomposition is functional rather than an attribute\-wise hard partition\. The textual prompt summarizes tendencies that are stable across the observed history, whereas the domain tokens specify how those tendencies should be realized under the current domain’s terminology, discourse structure, and pragmatic conventions\. Consequently, an attribute such as tone or language is not assigned permanently to one channel\. For example, a preference for empathetic explanations can be retained inPuP\_\{u\}, while its wording, degree of subjectivity, and evidence style are modulated byTtT\_\{t\}in a technical or social domain\. Both channels condition the same generator and are optimized by held\-out source\-task performance\. We therefore claim complementary conditioning roles, not identifiable or statistically independent latent factors\.

### B\.3Source\-Domain Episodic Tasks

Each source taskτ\\tauis a user–domain pair\. Following episodic gradient\-based meta\-learning\(Finnet al\.[2017b](https://arxiv.org/html/2608.12389#bib.bib118)\), its interactions are partitioned into disjoint support and query sets,𝒮τ\\mathcal\{S\}\_\{\\tau\}and𝒬τ\\mathcal\{Q\}\_\{\\tau\}\. The support set simulates target\-time adaptation, whereas the query set estimates performance on a new interaction from the same user–domain task\. No query response is used by the inner loop\. Support sizes are sampled from the same range evaluated at target time so that the meta\-objective learns an initialization and regularization geometry across different evidence regimes\.

### B\.4Information\-Access Matrix

Table[7](https://arxiv.org/html/2608.12389#A2.T7)makes the access boundary explicit\. In particular, target support responses supervise only the few\-shot LoRA update and its support\-conditioned entropy and Fisher statistics\. They are not used to construct either personalization prior\. This separation prevents the prior bank, user prompt, and target\-domain tokens from encoding the answers that are later used as adaptation supervision\.

Table 7:Information available to each component\. “Source only” means that the component is learned before the target support set is observed\.
### B\.5Offline and Online States

The backbone, preference summarizer, source\-domain bank, graph encoder, retrieval and composition modules, soft\-token projector, meta\-initialization, and learned group variances are fixed before a target episode begins\. At target time,PuP\_\{u\}andTtT\_\{t\}are computed once and held fixed across all inner\-loop steps\. The entropy and diagonal Fisher proxy are also computed once atθ0\\theta\_\{0\}and treated as stop\-gradient task statistics\. Only the target copy of the LoRA parameters is updated, and that copy is discarded after the episode\. Consequently, one target user’s adaptation cannot modify the shared model or another user’s state\.

### B\.6Notation

Table 8:Symbols used repeatedly in the supplement\.

## Appendix CPAC\-Bayes\-Regularized Meta\-LoRA

PAC\-Bayes meta\-learning permits source tasks to determine a task\-transfer prior before a new task is observed\(Riouet al\.[2023](https://arxiv.org/html/2608.12389#bib.bib99)\)\. Our construction specializes this principle to LoRA personalization: the episodically learned initialization is the prior mean, depth\-group variances encode source\-learned adaptation freedom, and target support size controls the deterministic anchor derived below\.

### C\.1Bounded Analysis Loss

The training loss is the token\-averaged negative log\-likelihood \(NLL\), which is nonnegative but unbounded\. The PAC\-Bayes argument instead requires a bounded loss\(McAllester[1999](https://arxiv.org/html/2608.12389#bib.bib119); Maurer[2004](https://arxiv.org/html/2608.12389#bib.bib47); Seeger[2002](https://arxiv.org/html/2608.12389#bib.bib102)\)\. Letℓnll​\(z;θ\)\\ell\_\{\\mathrm\{nll\}\}\(z;\\theta\)be the token\-averaged NLL for interactionzz, and define the analysis\-only loss

ℓb​\(z;θ\)=1−exp⁡\(−ℓnll​\(z;θ\)/c\)∈\[0,1\],\\ell\_\{b\}\(z;\\theta\)=1\-\\exp\\\!\\left\(\-\\ell\_\{\\mathrm\{nll\}\}\(z;\\theta\)/c\\right\)\\in\[0,1\],\(9\)wherec\>0c\>0is fixed using source\-domain validation data before any target episode\. This mapping is monotone, so it preserves the ordering of example losses, but it is not used by the adaptation optimizer\. All reported model updates continue to minimize the original NLL\.

### C\.2PAC\-Bayes Statement and Assumptions

#### Assumptions\.

Conditioned on the complete source\-training procedure, the priorP0P\_\{0\}is independent of the target support set\. TheK≥1K\\geq 1target support interactions are sampled independently from the target user–domain distribution, and the bounded loss in Eq\. \([9](https://arxiv.org/html/2608.12389#A3.E9)\) is fixed before observing them\. The posterior may depend on the support set\.

#### Proposition 1\.

For any target\-support\-independent priorP0P\_\{0\}, with probability at least1−δ1\-\\deltaover a support set of sizeKK, every posteriorQQsatisfies

R​\(Q\)≤R^𝒮​\(Q\)\+KL⁡\(Q∥P0\)\+ln⁡\(\(K\+1\)/δ\)2​K\.R\(Q\)\\leq\\widehat\{R\}\_\{\\mathcal\{S\}\}\(Q\)\+\\sqrt\{\\frac\{\\operatorname\{KL\}\(Q\\\|P\_\{0\}\)\+\\ln\(\(K\+1\)/\\delta\)\}\{2K\}\}\.\(10\)

#### Derivation\.

A standard PAC\-Bayes change\-of\-measure argument for bounded losses yields the square\-root form used in the main paper\(Maurer[2004](https://arxiv.org/html/2608.12389#bib.bib47)\)

R​\(Q\)−R^𝒮​\(Q\)≤KL⁡\(Q∥P0\)\+ln⁡\(\(K\+1\)/δ\)2​K,R\(Q\)\-\\widehat\{R\}\_\{\\mathcal\{S\}\}\(Q\)\\leq\\sqrt\{\\frac\{\\operatorname\{KL\}\(Q\\\|P\_\{0\}\)\+\\ln\(\(K\+1\)/\\delta\)\}\{2K\}\},\(11\)simultaneously for allQQ\. Equation \([10](https://arxiv.org/html/2608.12389#A3.E10)\) is therefore a restatement of Eq\. \([11](https://arxiv.org/html/2608.12389#A3.E11)\); no stronger guarantee is introduced in the supplement\.

### C\.3Closed\-Form Gaussian KL

For layerll, the prior and posterior have identical isotropic covariance:

Pl=𝒩​\(θ0,l,σg​\(l\)2​I\),Ql=𝒩​\(θl,σg​\(l\)2​I\)\.P\_\{l\}=\\mathcal\{N\}\(\\theta\_\{0,l\},\\sigma\_\{g\(l\)\}^\{2\}I\),\\qquad Q\_\{l\}=\\mathcal\{N\}\(\\theta\_\{l\},\\sigma\_\{g\(l\)\}^\{2\}I\)\.\(12\)The KL between twodld\_\{l\}\-dimensional Gaussians is

KL⁡\(Ql∥Pl\)\\displaystyle\\operatorname\{KL\}\(Q\_\{l\}\\\|P\_\{l\}\)=12\[tr\(Σl−1Σl\)−dl\+lndetΣldetΣl\\displaystyle=\\frac\{1\}\{2\}\\left\[\\operatorname\{tr\}\(\\Sigma\_\{l\}^\{\-1\}\\Sigma\_\{l\}\)\-d\_\{l\}\+\\ln\\frac\{\\det\\Sigma\_\{l\}\}\{\\det\\Sigma\_\{l\}\}\\right\.\+\(θ0,l−θl\)⊤Σl−1\(θ0,l−θl\)\]\\displaystyle\\hskip 68\.2866pt\\left\.\+\(\\theta\_\{0,l\}\-\\theta\_\{l\}\)^\{\\top\}\\Sigma\_\{l\}^\{\-1\}\(\\theta\_\{0,l\}\-\\theta\_\{l\}\)\\right\]=‖θl−θ0,l‖222​σg​\(l\)2\.\\displaystyle=\\frac\{\\\|\\theta\_\{l\}\-\\theta\_\{0,l\}\\\|\_\{2\}^\{2\}\}\{2\\sigma\_\{g\(l\)\}^\{2\}\}\.\(13\)Independence across layers makes the joint KL additive:

KL⁡\(Qθ∥P0\)=∑l=1L‖θl−θ0,l‖222​σg​\(l\)2\.\\operatorname\{KL\}\(Q\_\{\\theta\}\\\|P\_\{0\}\)=\\sum\_\{l=1\}^\{L\}\\frac\{\\\|\\theta\_\{l\}\-\\theta\_\{0,l\}\\\|\_\{2\}^\{2\}\}\{2\\sigma\_\{g\(l\)\}^\{2\}\}\.\(14\)No covariance, trace, determinant, or dimensionality term remains because the prior and posterior share the same covariance\.

### C\.4From the Square Root to aK−1K^\{\-1\}\-Scaled Surrogate

LetAθ=KL⁡\(Qθ∥P0\)\+ln⁡\(\(K\+1\)/δ\)A\_\{\\theta\}=\\operatorname\{KL\}\(Q\_\{\\theta\}\\\|P\_\{0\}\)\+\\ln\(\(K\+1\)/\\delta\)\. For anyβ\>0\\beta\>0andx≥0x\\geq 0,\(x−β\)2≥0\(x\-\\beta\)^\{2\}\\geq 0impliesx≤x2/\(2​β\)\+β/2x\\leq x^\{2\}/\(2\\beta\)\+\\beta/2\. Substitutingx=Aθ/\(2​K\)x=\\sqrt\{A\_\{\\theta\}/\(2K\)\}gives

Aθ2​K≤Aθ4​β​K\+β2\.\\sqrt\{\\frac\{A\_\{\\theta\}\}\{2K\}\}\\leq\\frac\{A\_\{\\theta\}\}\{4\\beta K\}\+\\frac\{\\beta\}\{2\}\.\(15\)For fixedKK,δ\\delta, andβ\\beta, the logarithmic term andβ/2\\beta/2do not depend onθ\\theta\. The remaining parameter\-dependent contribution is

14​β​K​KL⁡\(Qθ∥P0\),\\frac\{1\}\{4\\beta K\}\\operatorname\{KL\}\(Q\_\{\\theta\}\\\|P\_\{0\}\),\(16\)which motivates an anchor whose weight is proportional toK−1K^\{\-1\}\. The constant1/\(4​β\)1/\(4\\beta\)is absorbed into the source\-validated regularization coefficient\.

#### Scope of the guarantee\.

Equation \([10](https://arxiv.org/html/2608.12389#A3.E10)\) applies to the randomized predictor obtained by sampling fromQθQ\_\{\\theta\}\. Deployment uses the posterior meanθ\\theta\. We therefore use the bound to select the prior center, quadratic complexity geometry, and support\-size dependence of a deterministic surrogate\. We do not claim that Eq\. \([10](https://arxiv.org/html/2608.12389#A3.E10)\) is itself a bound on deterministic mean\-parameter generation\.

#### Status and empirical coverage of the theoretical claims\.

Proposition 1 is the only formal generalization statement made by this work; it is a specialization of a standard PAC\-Bayes result rather than a new PAC\-Bayes theorem\. Its assumptions, change\-of\-measure step, Pinsker reduction, Gaussian KL calculation, and linearization are given above\. The subsequent entropy calibration, Fisher reweighting, and depth\-group variances define a deterministic optimization surrogate and are not presented as additional generalization guarantees\. Their observable implications are tested separately: theK−1K^\{\-1\}and entropy ablations test evidence\-dependent anchoring, the Fisher and shared\-variance ablations test the proposed geometry, and the displacement analysis tests whether sparse\-support updates remain closer toθ0\\theta\_\{0\}\. Thus, the theoretical statement is fully derived, while the method\-specific consequences are evaluated empirically without conflating empirical support with proof of a deterministic risk bound\.

### C\.5Predictive\-Entropy Calibration

Letℐ​\(𝒮\)\\mathcal\{I\}\(\\mathcal\{S\}\)be the set of valid target\-token positions in the support responses under teacher forcing, and letpθ0​\(v∣cj\)p\_\{\\theta\_\{0\}\}\(v\\mid c\_\{j\}\)be the next\-token distribution at positionjjgiven its conditioned prefixcj=\[Pu;Tt;xi;yi,<j\]c\_\{j\}=\[P\_\{u\};T\_\{t\};x\_\{i\};y\_\{i,<j\}\]\. With vocabulary𝒱\\mathcal\{V\}, the normalized predictive entropy is

H~​\(𝒮;θ0\)\\displaystyle\\widetilde\{H\}\(\\mathcal\{S\};\\theta\_\{0\}\)=1\|ℐ​\(𝒮\)\|​∑j∈ℐ​\(𝒮\)H\(pθ0\(⋅∣cj\)\)ln⁡\|𝒱\|,\\displaystyle=\\frac\{1\}\{\|\\mathcal\{I\}\(\\mathcal\{S\}\)\|\}\\sum\_\{j\\in\\mathcal\{I\}\(\\mathcal\{S\}\)\}\\frac\{H\(p\_\{\\theta\_\{0\}\}\(\\cdot\\mid c\_\{j\}\)\)\}\{\\ln\|\\mathcal\{V\}\|\},\(17\)H\(pθ0\(⋅∣cj\)\)\\displaystyle H\(p\_\{\\theta\_\{0\}\}\(\\cdot\\mid c\_\{j\}\)\)=−∑v∈𝒱pθ0​\(v∣cj\)​ln⁡pθ0​\(v∣cj\)\.\\displaystyle=\-\\sum\_\{v\\in\\mathcal\{V\}\}p\_\{\\theta\_\{0\}\}\(v\\mid c\_\{j\}\)\\ln p\_\{\\theta\_\{0\}\}\(v\\mid c\_\{j\}\)\.ThusH~∈\[0,1\]\\widetilde\{H\}\\in\[0,1\]\. Entropy is a model\-relative uncertainty statistic, not an intrinsic measure of data quality, consistent with uncertainty\-aware calibration\(Krishnan and Tickoo[2020](https://arxiv.org/html/2608.12389#bib.bib83)\)\. The evidence weight is

γ​\(𝒮\)=1\+α​H~​\(𝒮;θ0\)K,\\gamma\(\\mathcal\{S\}\)=\\frac\{1\+\\alpha\\widetilde\{H\}\(\\mathcal\{S\};\\theta\_\{0\}\)\}\{K\},\(18\)whereα≥0\\alpha\\geq 0is selected on source\-domain validation episodes\. The unit term retains the PAC\-Bayes\-motivatedK−1K^\{\-1\}dependence, and the entropy term increases anchoring when the meta\-initialized predictor is diffuse on the observed responses\.

### C\.6Diagonal Empirical Fisher Proxy

For support interactionii, letℓi​\(θ0\)\\ell\_\{i\}\(\\theta\_\{0\}\)be its token\-averaged NLL and definegi,l=∇θlℓi​\(θ0\)g\_\{i,l\}=\\nabla\_\{\\theta\_\{l\}\}\\ell\_\{i\}\(\\theta\_\{0\}\)\. The diagonal squared\-gradient estimate is

F^l=Diag⁡\(1K​∑i=1Kgi,l⊙2\)\.\\widehat\{F\}\_\{l\}=\\operatorname\{Diag\}\\\!\\left\(\\frac\{1\}\{K\}\\sum\_\{i=1\}^\{K\}g\_\{i,l\}^\{\\odot 2\}\\right\)\.\(19\)This is a tractable empirical Fisher proxy rather than the exact population Fisher\. Squared\-gradient curvature proxies are widely used to preserve parameters that are locally important to predictive behavior\(Kirkpatricket al\.[2017](https://arxiv.org/html/2608.12389#bib.bib122)\)\. Its role here is operational: coordinates with larger support\-loss sensitivity receive larger penalties\. It is estimated once atθ0\\theta\_\{0\}, detached from the meta\-gradient graph, and held fixed during theMMinner\-loop steps\. BecauseF^l\\widehat\{F\}\_\{l\}depends on𝒮\\mathcal\{S\}, it is not part ofP0P\_\{0\}and cannot be used to justify prior independence\. It only reweights the deterministic surrogate after the PAC\-Bayes geometry has been derived\.

### C\.7Source\-Learned Depth\-Group Variances

To ensure positivity and avoid degenerate variances, each unconstrained source\-trained scalarρg\\rho\_\{g\}is mapped to

σg2=σmin2\+\(σmax2−σmin2\)​sigmoid⁡\(ρg\),\\sigma\_\{g\}^\{2\}=\\sigma\_\{\\min\}^\{2\}\+\\left\(\\sigma\_\{\\max\}^\{2\}\-\\sigma\_\{\\min\}^\{2\}\\right\)\\operatorname\{sigmoid\}\(\\rho\_\{g\}\),\(20\)where the bounds and depth partition are selected using source validation only\. All LoRA layers within a depth group share one variance\. The variances receive outer\-loop gradients through the adapted query loss, but they are not updated inside a task and are frozen before target adaptation\. A smaller learned variance makes displacement in that depth group more costly; a larger variance permits greater task\-specific movement\.

### C\.8Deterministic Adaptation and Meta\-Objective

Combining Eqs\. \([18](https://arxiv.org/html/2608.12389#A3.E18)\) and \([19](https://arxiv.org/html/2608.12389#A3.E19)\), the target\-time objective is

ℒadapt​\(θ;𝒮\)=ℒ^𝒮​\(θ\)\+γ​\(𝒮\)​∑l=1L\(θl−θ0,l\)⊤​F^l​\(θl−θ0,l\)2​σg​\(l\)2\.\\mathcal\{L\}\_\{\\mathrm\{adapt\}\}\(\\theta;\\mathcal\{S\}\)=\\widehat\{\\mathcal\{L\}\}\_\{\\mathcal\{S\}\}\(\\theta\)\+\\gamma\(\\mathcal\{S\}\)\\sum\_\{l=1\}^\{L\}\\frac\{\(\\theta\_\{l\}\-\\theta\_\{0,l\}\)^\{\\top\}\\widehat\{F\}\_\{l\}\(\\theta\_\{l\}\-\\theta\_\{0,l\}\)\}\{2\\sigma\_\{g\(l\)\}^\{2\}\}\.\(21\)For source taskτ\\tau, initializeθτ\(0\)=θ0\\theta\_\{\\tau\}^\{\(0\)\}=\\theta\_\{0\}and update

θτ\(m\+1\)\\displaystyle\\theta\_\{\\tau\}^\{\(m\+1\)\}=θτ\(m\)−ηin​∇θℒadapt​\(θτ\(m\);𝒮τ\),\\displaystyle=\\theta\_\{\\tau\}^\{\(m\)\}\-\\eta\_\{\\mathrm\{in\}\}\\nabla\_\{\\theta\}\\mathcal\{L\}\_\{\\mathrm\{adapt\}\}\\\!\\left\(\\theta\_\{\\tau\}^\{\(m\)\};\\mathcal\{S\}\_\{\\tau\}\\right\),\(22\)m=0,…,M−1\.\\displaystyle\\hskip 79\.66771ptm=0,\\ldots,M\-1\.The outer objective is

ℒmeta\\displaystyle\\mathcal\{L\}\_\{\\mathrm\{meta\}\}=𝔼τ​\[ℒ^𝒬τ​\(θτ\(M\);Pτ,Td​\(τ\)\)\]\\displaystyle=\\mathbb\{E\}\_\{\\tau\}\\left\[\\widehat\{\\mathcal\{L\}\}\_\{\\mathcal\{Q\}\_\{\\tau\}\}\\\!\\left\(\\theta\_\{\\tau\}^\{\(M\)\};P\_\{\\tau\},T\_\{d\(\\tau\)\}\\right\)\\right\]\(23\)\+λcomp​ℒcomp\+λtopo​ℒtopo\.\\displaystyle\\quad\+\\lambda\_\{\\mathrm\{comp\}\}\\mathcal\{L\}\_\{\\mathrm\{comp\}\}\+\\lambda\_\{\\mathrm\{topo\}\}\\mathcal\{L\}\_\{\\mathrm\{topo\}\}\.The latter two terms train source\-only domain composition and the topology\-preserving projector\. The query\-generation term propagates through the inner updates toθ0\\theta\_\{0\}and the variance parameters\. The preference summarizer is optimized separately with Eq\. \([40](https://arxiv.org/html/2608.12389#A5.E40)\) before episodic training and is then fixed\. The implementation uses the same declared first\- or second\-order differentiation setting for every meta\-learning baseline\.

### C\.9Computational and Storage Cost

The entropy calculation reuses the support forward pass\. The Fisher proxy requires per\-example LoRA gradients once atθ0\\theta\_\{0\}and stores one diagonal value per adapted parameter\. Thereafter, each regularization evaluation is an element\-wise multiply and reduction\. ForpLoRAp\_\{\\mathrm\{LoRA\}\}adapted parameters, the additional storage isO​\(pLoRA\)O\(p\_\{\\mathrm\{LoRA\}\}\)and the additional arithmetic per inner step isO​\(pLoRA\)O\(p\_\{\\mathrm\{LoRA\}\}\)\. No full Fisher matrix, Hessian, or target\-specific backbone copy is stored\.

## Appendix DReliable Domain Prior Construction

### D\.1Three Observable Views

The domain representation is deliberately based on observable aggregate evidence rather than manually assigned semantic attributes\.

- •Semantic\-content view\.Topic and entity distributions, intent frequencies, and fixed\-encoder keyword representations summarize what is discussed\.
- •Interaction\-pragmatics view\.Question type, instruction versus discussion ratios, terminology density, discourse markers, and organization cues summarize how interactions are conventionally expressed\.
- •Response\-structure view\.Token and sentence length, paragraph count, list frequency, formatting patterns, and related aggregate statistics summarize the form expected of responses\.

The views are not assumed to be semantically disjoint\. They are operational groupings that expose complementary statistics and permit view\-specific geometry and reliability treatment\.

### D\.2Geometry\-Compatible Preprocessing

For a simplex\-valued featurepi\(v\)p\_\{i\}^\{\(v\)\}, reliability shrinkage is first applied:

pi,sh\(v\)=ri\(v\)​pi\(v\)\+\(1−ri\(v\)\)​p¯\(v\)\.p\_\{i,\\mathrm\{sh\}\}^\{\(v\)\}=r\_\{i\}^\{\(v\)\}p\_\{i\}^\{\(v\)\}\+\(1\-r\_\{i\}^\{\(v\)\}\)\\bar\{p\}^\{\(v\)\}\.\(24\)Afterϵ\\epsilonsmoothing, the centered log\-ratio transform maps it to Aitchison Euclidean coordinates\(Aitchison[1986](https://arxiv.org/html/2608.12389#bib.bib123)\):

clr⁡\(p\)=ln⁡p−1dim\(p\)​𝟏𝟏⊤​ln⁡p\.\\operatorname\{clr\}\(p\)=\\ln p\-\\frac\{1\}\{\\dim\(p\)\}\\bm\{1\}\\bm\{1\}^\{\\top\}\\ln p\.\(25\)Proportion\-valued features use a clipped logit followed by median/MAD standardization\. Positive scalar features useln⁡\(1\+x\)\\ln\(1\+x\)followed by the same robust standardization\. Semantic vectors areℓ2\\ell\_\{2\}normalized before similarity calculation\. Each processed view is mapped by a view\-specific encoder into a shared dimensionality, but view statistics are not concatenated before their geometries are normalized\.

### D\.3Reliability Estimation and Graph Construction

For domainddand viewvv, letnd\(v\)n\_\{d\}^\{\(v\)\}be the effective sample size and letud\(v\)u\_\{d\}^\{\(v\)\}summarize estimation variance after bounded robust preprocessing\. We use the error\-radius proxy

bd\(v\)\\displaystyle b\_\{d\}^\{\(v\)\}=2​ud\(v\)​ln⁡\(1/δr\)nd\(v\)\+3​ln⁡\(1/δr\)nd\(v\),\\displaystyle=\\sqrt\{\\frac\{2u\_\{d\}^\{\(v\)\}\\ln\(1/\\delta\_\{r\}\)\}\{n\_\{d\}^\{\(v\)\}\}\}\+\\frac\{3\\ln\(1/\\delta\_\{r\}\)\}\{n\_\{d\}^\{\(v\)\}\},\(26\)rd\(v\)\\displaystyle r\_\{d\}^\{\(v\)\}=exp⁡\(−bd\(v\)/τr\)\.\\displaystyle=\\exp\(\-b\_\{d\}^\{\(v\)\}/\\tau\_\{r\}\)\.This score increases with effective sample size and decreases with uncertainty\. The constantsδr\\delta\_\{r\}andτr\\tau\_\{r\}are fixed using source validation\.

Letdi​j\(v\)d\_\{ij\}^\{\(v\)\}be a view\-compatible distance, such as Jensen–Shannon distance for distributions or cosine distance for normalized semantic vectors\. The reliability\-adjusted edge strength is

πi​j\(v\)=sigmoid⁡\(−di​j\(v\)/τv\+ln⁡\(ri\(v\)​rj\(v\)\+ϵ\)\)\.\\pi\_\{ij\}^\{\(v\)\}=\\operatorname\{sigmoid\}\\left\(\-d\_\{ij\}^\{\(v\)\}/\\tau\_\{v\}\+\\ln\(r\_\{i\}^\{\(v\)\}r\_\{j\}^\{\(v\)\}\+\\epsilon\)\\right\)\.\(27\)Edges below a source\-validated threshold are removed\. For the surviving neighborhood𝒩v​\(i\)\\mathcal\{N\}\_\{v\}\(i\), normalized aggregation gives

mi\(v\)=hi\(v\)\+∑j∈𝒩v​\(i\)πi​j\(v\)​Wv​hj\(v\)1\+∑j∈𝒩v​\(i\)πi​j\(v\)\.m\_\{i\}^\{\(v\)\}=\\frac\{h\_\{i\}^\{\(v\)\}\+\\sum\_\{j\\in\\mathcal\{N\}\_\{v\}\(i\)\}\\pi\_\{ij\}^\{\(v\)\}W\_\{v\}h\_\{j\}^\{\(v\)\}\}\{1\+\\sum\_\{j\\in\\mathcal\{N\}\_\{v\}\(i\)\}\\pi\_\{ij\}^\{\(v\)\}\}\.\(28\)The self term preserves the domain’s own evidence, while reliability\-weighted propagation lets low\-resource domains borrow information without treating every neighbor as equally trustworthy\.

### D\.4Reliability\-Weighted Wasserstein Fusion

Each propagated view representation is mapped to a probability measureμi\(v\)\\mu\_\{i\}^\{\(v\)\}on a shared learned support\. The view weights are

λi\(v\)=\(ri\(v\)\)κ∑v′\(ri\(v′\)\)κ\.\\lambda\_\{i\}^\{\(v\)\}=\\frac\{\(r\_\{i\}^\{\(v\)\}\)^\{\\kappa\}\}\{\\sum\_\{v^\{\\prime\}\}\(r\_\{i\}^\{\(v^\{\\prime\}\)\}\)^\{\\kappa\}\}\.\(29\)The fused measure is the entropy\-regularized Wasserstein barycenter\(Cuturi[2013](https://arxiv.org/html/2608.12389#bib.bib124); Cuturi and Doucet[2014](https://arxiv.org/html/2608.12389#bib.bib125)\)

μi∗=arg⁡minμ∈Δ​∑vλi\(v\)​Wε​\(μ,μi\(v\);C\),\\mu\_\{i\}^\{\\ast\}=\\arg\\min\_\{\\mu\\in\\Delta\}\\sum\_\{v\}\\lambda\_\{i\}^\{\(v\)\}W\_\{\\varepsilon\}\(\\mu,\\mu\_\{i\}^\{\(v\)\};C\),\(30\)whereCCis the ground\-cost matrix induced by the shared support andWεW\_\{\\varepsilon\}is the Sinkhorn\-regularized transport cost\. A learned readout mapsclr⁡\(μi∗\)\\operatorname\{clr\}\(\\mu\_\{i\}^\{\\ast\}\)tozi∈ℝhzz\_\{i\}\\in\\mathbb\{R\}^\{h\_\{z\}\}\. Reliability affects both graph propagation and cross\-view fusion, but its values are computed exclusively from source\-domain observations\.

### D\.5Leave\-One\-Domain\-Out Composition

During a source\-domain composition episode, a domainddis selected as a pseudo\-target\. Its node and incident edges are masked before graph propagation, andzdz\_\{d\}is removed from the candidate bank\. Its full source representation remains available only as a reconstruction target\. The target evidence encoder constructsz^d\\widehat\{z\}\_\{d\}from the domain description and a simulated sparse subset of interaction inputs\. Candidate retrieval and target\-conditioned composition producez¯d\\bar\{z\}\_\{d\}\. The source\-only training objective is

ℒcomp=‖z¯d−zd‖22\+λnbr​‖s​\(z¯d,ℬ−d\)−s​\(zd,ℬ−d\)‖22,\\mathcal\{L\}\_\{\\mathrm\{comp\}\}=\\\|\\,\\bar\{z\}\_\{d\}\-z\_\{d\}\\,\\\|\_\{2\}^\{2\}\+\\lambda\_\{\\mathrm\{nbr\}\}\\\|\\,s\(\\bar\{z\}\_\{d\},\\mathcal\{B\}\_\{\-d\}\)\-s\(z\_\{d\},\\mathcal\{B\}\_\{\-d\}\)\\,\\\|\_\{2\}^\{2\},\(31\)wheres​\(q,ℬ−d\)s\(q,\\mathcal\{B\}\_\{\-d\}\)is the vector of cosine relations betweenqqand the remaining source bank\. The second term prevents a reconstruction with low pointwise error but incorrect neighbors\. Masking both the candidate node and its graph edges prevents trivial identity copying\.

### D\.6Target Evidence and Uncertainty\-Aware Retrieval

Letedesce\_\{\\mathrm\{desc\}\}be the normalized encoding of the target\-domain description and letesupe\_\{\\mathrm\{sup\}\}be the normalized mean encoding of the support inputs whenK\>0K\>0\. Source domainddhas a corresponding text anchorcdc\_\{d\}and learned bank representationzdz\_\{d\}\. For evidence streamv∈\{desc,sup\}v\\in\\\{\\mathrm\{desc\},\\mathrm\{sup\}\\\}, define

wd\(v\)\\displaystyle w\_\{d\}^\{\(v\)\}=exp⁡\(κv​cos⁡\(ev,cd\)\)∑d′∈𝒟sexp⁡\(κv​cos⁡\(ev,cd′\)\),\\displaystyle=\\frac\{\\exp\(\\kappa\_\{v\}\\cos\(e\_\{v\},c\_\{d\}\)\)\}\{\\sum\_\{d^\{\\prime\}\\in\\mathcal\{D\}\_\{s\}\}\\exp\(\\kappa\_\{v\}\\cos\(e\_\{v\},c\_\{d^\{\\prime\}\}\)\)\},\(32\)ρv\\displaystyle\\rho\_\{v\}=1−−∑dwd\(v\)​ln⁡wd\(v\)ln⁡\|𝒟s\|\.\\displaystyle=1\-\\frac\{\-\\sum\_\{d\}w\_\{d\}^\{\(v\)\}\\ln w\_\{d\}^\{\(v\)\}\}\{\\ln\|\\mathcal\{D\}\_\{s\}\|\}\.The reliabilityρv\\rho\_\{v\}is high when the evidence identifies a concentrated source neighborhood and low when it is nearly uninformative\. The target\-evidence representation is

z^t=ρdesc​∑dwd\(desc\)​zd\+𝕀​\[K\>0\]​ρsup​∑dwd\(sup\)​zdρdesc\+𝕀​\[K\>0\]​ρsup\+ϵ\.\\widehat\{z\}\_\{t\}=\\frac\{\\rho\_\{\\mathrm\{desc\}\}\\sum\_\{d\}w\_\{d\}^\{\(\\mathrm\{desc\}\)\}z\_\{d\}\+\\mathbb\{I\}\[K\>0\]\\rho\_\{\\mathrm\{sup\}\}\\sum\_\{d\}w\_\{d\}^\{\(\\mathrm\{sup\}\)\}z\_\{d\}\}\{\\rho\_\{\\mathrm\{desc\}\}\+\\mathbb\{I\}\[K\>0\]\\rho\_\{\\mathrm\{sup\}\}\+\\epsilon\}\.\(33\)Only support inputs enter Eq\. \([33](https://arxiv.org/html/2608.12389#A4.E33)\); support responses do not\. In zero\-shot inference the description stream alone is used\.

Letrdr\_\{d\}be the aggregate source reliability of candidateddand letut=1−max⁡\(ρdesc,𝕀​\[K\>0\]​ρsup\)u\_\{t\}=1\-\\max\(\\rho\_\{\\mathrm\{desc\}\},\\mathbb\{I\}\[K\>0\]\\rho\_\{\\mathrm\{sup\}\}\)\. Retrieval confidence is

ωt​d∝rd​exp⁡\(cos⁡\(z^t,zd\)τ0\+τu​ut\)\.\\omega\_\{td\}\\propto r\_\{d\}\\exp\\\!\\left\(\\frac\{\\cos\(\\widehat\{z\}\_\{t\},z\_\{d\}\)\}\{\\tau\_\{0\}\+\\tau\_\{u\}u\_\{t\}\}\\right\)\.\(34\)High target uncertainty increases the effective temperature and yields a less brittle candidate distribution\. The topkkcandidates form𝒩k​\(t\)\\mathcal\{N\}\_\{k\}\(t\)\. Their final target\-conditioned contributions are those in Eq\. \(8\) of the main paper: retrieval determines admissible references, while attention may correct their relative weights for the current target\.

### D\.7Topology\-Preserving Projection

For a batch ofBBcomposed domain representations, letZ¯∈ℝB×hz\\overline\{Z\}\\in\\mathbb\{R\}^\{B\\times h\_\{z\}\}stack the centered representations and letT¯∈ℝB×h\\overline\{T\}\\in\\mathbb\{R\}^\{B\\times h\}stack centered mean\-pooled soft\-token representations\. Their linear CKA\(Kornblithet al\.[2019](https://arxiv.org/html/2608.12389#bib.bib126)\)is

CKA⁡\(Z¯,T¯\)=‖Z¯⊤​T¯‖F2‖Z¯⊤​Z¯‖F​‖T¯⊤​T¯‖F\.\\operatorname\{CKA\}\(\\overline\{Z\},\\overline\{T\}\)=\\frac\{\\\|\\overline\{Z\}^\{\\top\}\\overline\{T\}\\\|\_\{F\}^\{2\}\}\{\\\|\\overline\{Z\}^\{\\top\}\\overline\{Z\}\\\|\_\{F\}\\\|\\overline\{T\}^\{\\top\}\\overline\{T\}\\\|\_\{F\}\}\.\(35\)For composed domainsiiandjj,Si​jzS^\{z\}\_\{ij\}is their cosine similarity\. Letqiq\_\{i\}andkik\_\{i\}be the mean\-pooled outputs obtained after the domain tokens pass through the frozen query and key projections of the language model\. The symmetrized attention\-interface relation is

Si​jA=12​\[cos⁡\(qi,kj\)\+cos⁡\(qj,ki\)\]\.S^\{A\}\_\{ij\}=\\tfrac\{1\}\{2\}\\left\[\\cos\(q\_\{i\},k\_\{j\}\)\+\\cos\(q\_\{j\},k\_\{i\}\)\\right\]\.\(36\)The topology loss is

ℒtopo=1−CKA⁡\(Z¯,T¯\)\+λattn​‖SA−Sz‖F2\.\\mathcal\{L\}\_\{\\mathrm\{topo\}\}=1\-\\operatorname\{CKA\}\(\\overline\{Z\},\\overline\{T\}\)\+\\lambda\_\{\\mathrm\{attn\}\}\\\|S^\{A\}\-S^\{z\}\\\|\_\{F\}^\{2\}\.\(37\)The first term preserves global relational geometry across representation spaces; the second preserves relations after the frozen projections through which the model consumes the injected tokens\. This is relation preservation, not a claim that the two spaces are pointwise identical or topologically homeomorphic\.

### D\.8Injection and Source\-Only Training

The projector uses a bottleneck MLP followed by a reshape:

Td=reshape⁡\(W2​φ​\(W1​z¯d\)\)∈ℝnT×h\.T\_\{d\}=\\operatorname\{reshape\}\\\!\\left\(W\_\{2\}\\varphi\(W\_\{1\}\\bar\{z\}\_\{d\}\)\\right\)\\in\\mathbb\{R\}^\{n\_\{T\}\\times h\}\.\(38\)The tokens are inserted after the system/user\-preference prefix and before the interaction query\. In every source episode, the sameTdT\_\{d\}conditions support and query examples and is held fixed while the task LoRA parameters adapt\. The composition module and projector receive gradients from reconstruction, topology alignment, and episodic query generation\. At target time they are frozen\.

## Appendix ESource\-Trained User Preference Summarizer

### E\.1Leave\-One\-Domain\-Out Training

The summarizerfωf\_\{\\omega\}is trained only on source users and source domains\. For a source user, one source domain is held out as a pseudo\-target\. The summarizer observes history from the remaining domains and generates a textual preference prompt\. Held\-out interactions are used to score whether the prompt improves personalization beyond the evidence from which it was constructed\. This protocol trains the prompt to preserve tendencies that transfer to an unobserved domain instead of copying local topics from the input history\.

### E\.2Multi\-View History Perturbations

Two history views for the same user are constructed by independently sampling interaction subsets, masking segments, permuting interaction order, and varying the balance among questions, responses, and metadata fields\. The perturbations remove reliable access to any single surface template\. Consistency is imposed on core preference statements supported by both views, while view\-specific details are allowed to differ\. This avoids forcing an identical prompt when the two views contain genuinely different evidence\.

### E\.3Source\-Only Preference Optimization

The implementation uses group\-relative policy optimization\(Shaoet al\.[2024](https://arxiv.org/html/2608.12389#bib.bib107)\)to train the discrete prompt generator\. For one history conditioncc, sampleJJcandidate prompts\{Pj\}j=1J\\\{P\_\{j\}\\\}\_\{j=1\}^\{J\}\. LetRjR\_\{j\}be the source\-only reward and define the group\-normalized advantage

Aj=Rj−R¯sR\+ϵ\.A\_\{j\}=\\frac\{R\_\{j\}\-\\overline\{R\}\}\{s\_\{R\}\+\\epsilon\}\.\(39\)The policy objective is

ℒsum\\displaystyle\\mathcal\{L\}\_\{\\mathrm\{sum\}\}=−1J​∑j=1JAj​∑r=1\|Pj\|ln⁡πω​\(Pj,r∣c,Pj,<r\)\\displaystyle=\-\\frac\{1\}\{J\}\\sum\_\{j=1\}^\{J\}A\_\{j\}\\sum\_\{r=1\}^\{\|P\_\{j\}\|\}\\ln\\pi\_\{\\omega\}\(P\_\{j,r\}\\mid c,P\_\{j,<r\}\)\(40\)\+λKL​KL⁡\(πω∥πref\)\.\\displaystyle\\quad\+\\lambda\_\{\\mathrm\{KL\}\}\\operatorname\{KL\}\(\\pi\_\{\\omega\}\\\|\\pi\_\{\\mathrm\{ref\}\}\)\.
The reward contains four auditable source\-only terms\. First, a utility term measures the improvement in the log\-likelihood margin between a held\-out personalized responsey\+y^\{\+\}and a generic alternativey−y^\{\-\}:

Δ​m​\(P\)\\displaystyle\\Delta m\(P\)=\[ln⁡p​\(y\+∣x,P\)−ln⁡p​\(y−∣x,P\)\]\\displaystyle=\\left\[\\ln p\(y^\{\+\}\\mid x,P\)\-\\ln p\(y^\{\-\}\\mid x,P\)\\right\]\(41\)−\[ln⁡p​\(y\+∣x\)−ln⁡p​\(y−∣x\)\]\.\\displaystyle\\quad\-\\left\[\\ln p\(y^\{\+\}\\mid x\)\-\\ln p\(y^\{\-\}\\mid x\)\\right\]\.Second, a domain\-leakage term rewards high entropy under a source\-trained domain classifier, discouraging prompts that reveal an input domain from surface topics\. Third, a cross\-view stability term rewards semantic agreement between core statements extracted from two views of the same user, gated by the overlap of their evidence\. Fourth, format and length terms keep the prompt concise and parseable\. The total reward is

R=wu​rutil\+wd​rdom\+ws​rstab\+wf​rfmt\+wℓ​rlen,R=w\_\{u\}r\_\{\\mathrm\{util\}\}\+w\_\{d\}r\_\{\\mathrm\{dom\}\}\+w\_\{s\}r\_\{\\mathrm\{stab\}\}\+w\_\{f\}r\_\{\\mathrm\{fmt\}\}\+w\_\{\\ell\}r\_\{\\mathrm\{len\}\},\(42\)with all weights selected on source validation\. The domain\-leakage and stability weights are increased only after the utility and formatting rewards stabilize, which reduces collapse to generic, high\-entropy summaries\.

### E\.4Prompt Format and Target\-Time Use

The output is a short natural\-language profile organized around repeatedly supported preferences: desired level of detail, reasoning and evidence\-use tendencies, presentation and interaction preferences, and explicit negative constraints\. Domain\-specific entities are included only when repeated evidence supports them as a transferable preference rather than an incidental source\-domain topic\. The prompt contains no user identifier\. Ifℋu\\mathcal\{H\}\_\{u\}is empty, a fixed neutral prompt states that no stable user preference is available\. At target time the summarizer is frozen,Pu=fω​\(ℋu\)P\_\{u\}=f\_\{\\omega\}\(\\mathcal\{H\}\_\{u\}\)is computed once, and the target support and test responses are unavailable to it\.

## Appendix FEnd\-to\-End Training and Inference

### F\.1Source\-Domain Training

Algorithm[1](https://arxiv.org/html/2608.12389#alg1)gives the complete source\-only training sequence and makes the pseudo\-target masking and information boundaries explicit\.

Algorithm 1Source\-domain episodic training0:Source interactions, source domains

𝒟s\\mathcal\{D\}\_\{s\}, frozen backbone

1:Train the preference summarizer on source\-only leave\-one\-domain\-out prompt episodes\.

2:Construct the three source\-domain views, reliability scores, view\-specific graphs, and source bank

ℬ\\mathcal\{B\}\.

3:foreach meta\-training iterationdo

4:Sample a source domain

ddas a pseudo\-target and remove its node and incident edges from the reference graph\.

5:Construct pseudo\-target evidence from its description and a sparse subset of interaction inputs\.

6:Retrieve and compose candidates from

ℬ∖\{zd\}\\mathcal\{B\}\\setminus\\\{z\_\{d\}\\\}; project the composition to

TdT\_\{d\}\.

7:Sample a user–domain task

τ\\tauwith disjoint

𝒮τ\\mathcal\{S\}\_\{\\tau\}and

𝒬τ\\mathcal\{Q\}\_\{\\tau\}and compute its fixed prompt

PτP\_\{\\tau\}\.

8:Compute

H~​\(𝒮τ;θ0\)\\widetilde\{H\}\(\\mathcal\{S\}\_\{\\tau\};\\theta\_\{0\}\)and

\{F^l\}\\\{\\widehat\{F\}\_\{l\}\\\}once at

θ0\\theta\_\{0\}; stop gradients through these statistics\.

9:Starting from

θ0\\theta\_\{0\}, take

MMsteps on Eq\. \([21](https://arxiv.org/html/2608.12389#A3.E21)\) using

𝒮τ\\mathcal\{S\}\_\{\\tau\}\.

10:Evaluate the adapted parameters on

𝒬τ\\mathcal\{Q\}\_\{\\tau\}under the same

PτP\_\{\\tau\}and

TdT\_\{d\}\.

11:Update

θ0\\theta\_\{0\}, variance parameters, graph/composition modules, and projector with the outer objective\.

12:endfor

13:Freeze all shared components before target evaluation\.

### F\.2Unified Zero\- and Few\-Shot Inference

Algorithm[2](https://arxiv.org/html/2608.12389#alg2)gives the common zero\- and few\-shot inference path\.

Algorithm 2Target\-domain inference0:Pre\-target history

ℋu\\mathcal\{H\}\_\{u\}, target description, query

xtx\_\{t\}, optional support

𝒮t\\mathcal\{S\}\_\{t\}, frozen shared modules

1:

Pu←fω​\(ℋu\)P\_\{u\}\\leftarrow f\_\{\\omega\}\(\\mathcal\{H\}\_\{u\}\)\.

2:Construct

z^t\\widehat\{z\}\_\{t\}from the target description and support inputs only\.

3:Retrieve

𝒩k​\(t\)\\mathcal\{N\}\_\{k\}\(t\)from

ℬ\\mathcal\{B\}, compose

z¯t\\bar\{z\}\_\{t\}, and set

Tt←Πψ​\(z¯t\)T\_\{t\}\\leftarrow\\Pi\_\{\\psi\}\(\\bar\{z\}\_\{t\}\)\.

4:

θ←θ0\\theta\\leftarrow\\theta\_\{0\}\.

5:if

K\>0K\>0then

6:Condition every support example on the same

\[Pu;Tt\]\[P\_\{u\};T\_\{t\}\]\.

7:Compute

H~​\(𝒮t;θ0\)\\widetilde\{H\}\(\\mathcal\{S\}\_\{t\};\\theta\_\{0\}\)and

\{F^l\}\\\{\\widehat\{F\}\_\{l\}\\\}once and detach them\.

8:for

m=0,…,M−1m=0,\\ldots,M\-1do

9:

θ←θ−ηtest​∇θℒadapt​\(θ;𝒮t\)\\theta\\leftarrow\\theta\-\\eta\_\{\\mathrm\{test\}\}\\nabla\_\{\\theta\}\\mathcal\{L\}\_\{\\mathrm\{adapt\}\}\(\\theta;\\mathcal\{S\}\_\{t\}\)\.

10:endfor

11:endif

12:Generate and return

G​\(\[Pu;Tt;xt\];θ\)G\(\[P\_\{u\};T\_\{t\};x\_\{t\}\];\\theta\)\.

### F\.3Behavior at the Two Endpoints

WhenK=0K=0, no target gradient, entropy statistic, or Fisher statistic is computed; prediction usesθ0\\theta\_\{0\}with user and domain conditioning\. WhenK\>0K\>0, the target responses enter only the LoRA adaptation path\. IncreasingKKweakens the explicitK−1K^\{\-1\}anchor, while the entropy and Fisher terms retain model\-relative and direction\-relative calibration\. The method therefore implements one conditioning interface with two adaptation modes rather than two separately trained systems\.

## Appendix GExperimental Protocol

### G\.1Benchmarks and Domain Splits

HiCUPID\(Moket al\.[2025a](https://arxiv.org/html/2608.12389#bib.bib110)\)evaluates personalized assistant responses whose relevant user attributes must be inferred from interaction history\. The ten source domains are Major, Self\-improvement, Social Media, Family, Forms of Living, Technology, Environment, Electronics, Religion, and Art\. Politics, Music, Finance, Beauty, and Food are held out as target domains\. LaMP\-QA\(Salemi and Zamani[2025](https://arxiv.org/html/2608.12389#bib.bib44)\)evaluates personalized long\-form question answering using hidden user\-expected aspects\. The held\-out target set used in the complete cross\-domain analysis is Parenting, Interpersonal, Politics, Law, and Philosophy; the main paper shows the representative domains that fit within its space budget\. We select these two public benchmarks because they provide complementary tests of the same claim: HiCUPID supplies controlled personalized\-assistant domains and an official evaluator, whereas LaMP\-QA provides naturally occurring long\-form questions, diverse domain conventions, and non\-unique valid answers\. Their combination tests whether the method transfers across both conversational domains and task/evaluation formats rather than exploiting one benchmark’s construction\. The exact domain and interaction assignments are emitted by the deterministic split utility in the code archive\.

For every dataset, source training, source validation, and target evaluation are separated at the domain level\. Hyperparameter selection uses only source validation episodes\. Within a target user–domain task, support and test interactions are disjoint\. The same support indices are reused across methods and random seeds\.

Figure[4](https://arxiv.org/html/2608.12389#A7.F4)visualizes the aggregate inter\-domain similarities used only to audit the chosen source/target partition\. The target domains span both near and distant relations to the source set, so the evaluation is not restricted to trivial semantic neighbors\. These visualizations are descriptive; they are not supplied to any compared method\.

![Refer to caption](https://arxiv.org/html/2608.12389v1/figure/hi_heatmap.png)\(a\)HiCUPID\.
![Refer to caption](https://arxiv.org/html/2608.12389v1/figure/la_heatmap.png)\(b\)LaMP\-QA\.

Figure 4:Inter\-domain similarity audit for the two benchmark partitions\. Darker cells indicate larger cosine similarity between aggregate held\-out interaction representations\.
### G\.2Baseline Implementations

- •SFT/LoRAadapts the source\-trained model on the target support set without episodic initialization\.
- •LaMP\(Salemiet al\.[2024a](https://arxiv.org/html/2608.12389#bib.bib108)\)retrieves user\-history evidence and prepends it to the generator input\.
- •GPG\(Zhang[2024](https://arxiv.org/html/2608.12389#bib.bib48)\)generates a textual profile from user history and conditions generation on that profile\.
- •SPT\(Huanget al\.[2024](https://arxiv.org/html/2608.12389#bib.bib54); Lesteret al\.[2021](https://arxiv.org/html/2608.12389#bib.bib120)\)learns continuous prompt vectors while keeping the backbone frozen\.
- •DEP\(Qiuet al\.[2025a](https://arxiv.org/html/2608.12389#bib.bib113)\)uses a latent preference representation for continuous conditioning\.
- •IDL\(Chenget al\.[2024](https://arxiv.org/html/2608.12389#bib.bib56)\)composes shared low\-rank components for user\-level adaptation\.
- •OPPU\(Tanet al\.[2024a](https://arxiv.org/html/2608.12389#bib.bib109)\)optimizes a user\-specific LoRA module with its original objective on exactly the same target support set\. Its source\-trained backbone is frozen\.
- •PROPER\(Zhanget al\.[2025a](https://arxiv.org/html/2608.12389#bib.bib52)\)performs its published progressive/group\-level parameter adaptation using the same support examples\.
- •CoPL\(Choiet al\.[2025a](https://arxiv.org/html/2608.12389#bib.bib112)\)applies its published contrastive personalization objective under the same source/target split\.
- •BiPO\(Caoet al\.[2024b](https://arxiv.org/html/2608.12389#bib.bib117)\)performs bidirectional preference steering at inference\.
- •CHAMELEON\(Zhanget al\.[2025c](https://arxiv.org/html/2608.12389#bib.bib58)\)edits internal representations using its published personalization procedure\.
- •OPAD\(Zhuet al\.[2025](https://arxiv.org/html/2608.12389#bib.bib104)\)performs on\-the\-fly preference alignment with its original inference objective\.
- •Proactive\(Zhanget al\.[2025b](https://arxiv.org/html/2608.12389#bib.bib105)\)uses proactive clarification or preference elicitation before response generation\.
- •MAML\(Finnet al\.[2017b](https://arxiv.org/html/2608.12389#bib.bib118)\)is Model\-Agnostic Meta\-Learning over the same LoRA parameter subset used by the other parameter\-adaptation methods\.
- •Vanilla Meta\-LoRAuses the same source episodes, LoRA parameterization, number of inner steps, and conditioning inputs as our method, but its inner loop minimizes only support NLL; it has noK−1K^\{\-1\}anchor, entropy calibration, learned depth\-group variances, or Fisher weighting\.

Each baseline uses its original objective and receives no target interaction unavailable to another method in the same shot setting\. Decoding parameters, maximum context, support partitions, test instances, and pairwise evaluation order are shared\. When a method cannot natively use a component of our framework, that component is not added unless the experiment explicitly reports a controlled plug\-in analysis\.

Table 9:Final baseline settings\. Entries list deviations from the shared backbone, data splits, context limit, and decoding protocol in Table[10](https://arxiv.org/html/2608.12389#A7.T10); unlisted options retain the cited implementation’s defaults\.
### G\.3Metrics

#### Win Rate\.

For test instanceii, a fixed pairwise evaluator compares the generated personalized response with the corresponding benchmark reference under randomized response order\. Pairwise, position\-randomized LLM evaluation is preferred to an unblinded absolute rating because it reduces scale\-calibration and presentation biases\(Zhenget al\.[2023](https://arxiv.org/html/2608.12389#bib.bib128)\)\. Letwi=1w\_\{i\}=1for a win,wi=0w\_\{i\}=0for a loss, andwi=1/2w\_\{i\}=1/2for a persistent tie under the fixed tie policy\. Then

WR=100​1N​∑i=1Nwi\.\\mathrm\{WR\}=100\\,\\frac\{1\}\{N\}\\sum\_\{i=1\}^\{N\}w\_\{i\}\.\(43\)HiCUPID uses its official evaluator\. LaMP\-QA uses the fixed hidden\-aspect pairwise rubric described below\.

#### Personal Gap\.

Lete​\(⋅\)e\(\\cdot\)be a frozen Sentence\-BERT encoder\(Reimers and Gurevych[2019](https://arxiv.org/html/2608.12389#bib.bib127)\),yi\+y\_\{i\}^\{\+\}the personalized target,yi0y\_\{i\}^\{0\}the generic alternative, andy^i\\widehat\{y\}\_\{i\}the generated response\. The judge\-independent specificity margin is

PG=1N​∑i=1N\[cos⁡\(e​\(y^i\),e​\(yi\+\)\)−cos⁡\(e​\(y^i\),e​\(yi0\)\)\]\.\\mathrm\{PG\}=\\frac\{1\}\{N\}\\sum\_\{i=1\}^\{N\}\\left\[\\cos\(e\(\\widehat\{y\}\_\{i\}\),e\(y\_\{i\}^\{\+\}\)\)\-\\cos\(e\(\\widehat\{y\}\_\{i\}\),e\(y\_\{i\}^\{0\}\)\)\\right\]\.\(44\)A positive value means that the response is closer to the personalized target than to a generic alternative\.

#### Fine\-grained quality and CCS\.

A fixed method\-blinded rubric, following the multidimensional evaluation practice of HiCUPID\(Moket al\.[2025a](https://arxiv.org/html/2608.12389#bib.bib110)\), scores UtilityUiU\_\{i\}, HonestyHiH\_\{i\}, and Personal FitPiP\_\{i\}on\{0,1,2\}\\\{0,1,2\\\}\. The Calibrated Composite Score is

CCS=1N​∑i=1N\(0\.3​Ui\+0\.3​Hi\+0\.4​Pi\)\.\\mathrm\{CCS\}=\\frac\{1\}\{N\}\\sum\_\{i=1\}^\{N\}\\left\(0\.3U\_\{i\}\+0\.3H\_\{i\}\+0\.4P\_\{i\}\\right\)\.\(45\)The larger weight on Personal Fit reflects the personalization objective while retaining explicit utility and factuality terms\. The same prompt, weights, and judge checkpoint are fixed before comparing methods\.

#### Cross\-domain degradation\.

IfWRsrc\\mathrm\{WR\}\_\{\\mathrm\{src\}\}is the mean source\-domain WR andWR¯tgt\\overline\{\\mathrm\{WR\}\}\_\{\\mathrm\{tgt\}\}is the unweighted mean over held\-out target domains, then

Δ​WR=WRsrc−WR¯tgt\.\\Delta\\mathrm\{WR\}=\\mathrm\{WR\}\_\{\\mathrm\{src\}\}\-\\overline\{\\mathrm\{WR\}\}\_\{\\mathrm\{tgt\}\}\.\(46\)All arithmetic is performed from unrounded domain\-level values; displayed values are rounded only after aggregation\.

### G\.4Evaluator and Human\-Agreement Protocol

The HiCUPID WR calculation follows its official evaluator\(Moket al\.[2025a](https://arxiv.org/html/2608.12389#bib.bib110)\)\. LaMP\-QA lacks a single gold answer, so evaluation follows its hidden\-aspect protocol\(Salemi and Zamani[2025](https://arxiv.org/html/2608.12389#bib.bib44)\)and asks the frozen evaluator to compare two responses against the annotated expected aspects\. Method names are removed and answer order is randomized\. PG is computed without an LLM judge\.

For the reliability study, 100 test instances are sampled independently from each benchmark\. Human assessors receive the same query, permissible user evidence, and rubric but no method identity\. Scores from the fixed Qwen3\-32B evaluator\(Qwen Team[2025](https://arxiv.org/html/2608.12389#bib.bib129)\)are correlated with human scores and an independent GPT\-4o evaluation using Pearson’srr\. The resulting agreement is reported in Appendix[H\.10](https://arxiv.org/html/2608.12389#A8.SS10); it evaluates the measurement procedure rather than the superiority of any judge model\.

### G\.5Implementation and Hyperparameters

Table[10](https://arxiv.org/html/2608.12389#A7.T10)records the final shared settings, and Table[11](https://arxiv.org/html/2608.12389#A7.T11)reports the complete grid for every tunable coefficient, capacity, and learning rate in the proposed modules\. Architecture\-imposed quantities, deterministic decoding, hardware, and cited baseline defaults are fixed rather than optimized\.

Grid search uses nested source\-only validation\. Source domains are rotated as pseudo\-targets; each is removed from the reference bank, and its query examples remain disjoint from support\. Every candidate in a stage receives identical episode manifests, training budget, and seeds\. We select common LoRA/meta\-optimization settings first, followed by the evidence\-calibrated objective, domain retrieval and reliability, soft\-token projection, and summarizer optimization\. The primary criterion is equal\-domain\-weighted mean WR across source\-validation pseudo\-targets\. If candidates are within0\.20\.2WR points, CCS is the tie\-breaker; a remaining tie favors the smaller model or weaker regularizer\. The selected value is frozen before the next stage\. The complete configuration is then retrained on source data and used unchanged for every held\-out target domain and shot setting\. Target labels and target\-domain development scores never enter search, early stopping, or checkpoint choice\.

Table 10:Final implementation settings\. Values governing the new modules are selected exclusively on source\-validation episodes\.Table 11:Development search spaces and selected values\. Each brace\-delimited set is the complete set tried for that parameter; fixed choices are reported in Table[10](https://arxiv.org/html/2608.12389#A7.T10)\.#### Randomness control\.

For run seeds∈\{42,…,51\}s\\in\\\{42,\\ldots,51\\\}, the same value initializes Python’srandom, NumPy, PyTorch CPU, and every CUDA device, and is passed to episode sampling, support\-set construction, data\-loader shuffling, and model initialization\. We setPYTHONHASHSEED=ss, disable cuDNN benchmarking, enable deterministic cuDNN kernels, and request deterministic PyTorch algorithms with warning mode for operations without a deterministic CUDA implementation\. Data\-loader workers use seeds\+worker\_ids\+\\texttt\{worker\\\_id\}, and the generator supplied to each loader is seeded byss\. Split manifests and nested support indices are generated once and reused by all methods\. Greedy decoding removes sampling randomness at generation time; the response\-order permutation used by the pairwise evaluator is separately stored for every instance and seed\.

### G\.6Statistical Reporting and Fairness Controls

WR is reported as mean and standard deviation over ten independent seeds\. Every method is evaluated on identical target support/test indices\. CCS and PG are first computed per test instance and then aggregated using the same indices\. Paired comparisons use a two\-sided Wilcoxon signed\-rank test on matched seed\-level scores and a10,00010\{,\}000\-resample paired bootstrap interval over matched test instances; Holm correction is applied when one analysis covers multiple domains\. Relative improvements use the strongest competing method in the same cell as the denominator\. Ablations that alter source training are retrained from scratch rather than disabled only at inference\. The code archive implements paired tests, bootstrap intervals, deterministic response ordering, and nested support sampling\.

## Appendix HExtended Analyses

### H\.1Exact Ablation Definitions

- •Textual conditioning onlyretainsPuP\_\{u\}but removes domain soft tokens and target parameter adaptation\.
- •\+ domain soft tokensadds the complete graph\-guided domain\-prior path but still usesθ0\\theta\_\{0\}without target LoRA updates\.
- •\+ vanilla Meta\-LoRAadds standard support\-loss\-only episodic LoRA adaptation\.
- •Full modelreplaces the vanilla inner objective with Eq\. \([21](https://arxiv.org/html/2608.12389#A3.E21)\)\.
- •WithoutK−1K^\{\-1\}scalingreplaces1/K1/Kby a source\-validated constant while retaining entropy and Fisher terms\.
- •Without entropy calibrationsetsα=0\\alpha=0but retains1/K1/K\.
- •Shared group variancereplaces all learnedσg2\\sigma\_\{g\}^\{2\}values by one source\-learned scalar\.
- •Without Fisher weightingreplacesF^l\\widehat\{F\}\_\{l\}by the identity matrix, reducing the anchor to a depth\-weighted Euclidean distance\.
- •Without graph reliabilityremoves reliability from edge construction, aggregation, and barycenter weighting while preserving the same view features\.
- •Without topology preservationsets the CKA and attention\-alignment weights to zero while retaining the projector architecture and token count\.

These definitions separate cumulative contribution from component ablation\. In the main\-paper cumulative block, domain tokens raise zero\-shot WR from55\.855\.8to59\.059\.0; in the 5\-shot regime, adding domain tokens, vanilla Meta\-LoRA, and the complete evidence\-calibrated objective raises WR from59\.359\.3to63\.463\.4,68\.468\.4, and73\.273\.2, respectively\. Thus the gains cannot be attributed only to a stronger text\-conditioned base model\. Because a changed inner objective also changes the learnedθ0\\theta\_\{0\}, zero\-shot performance may change even though zero\-shot inference itself performs no target update\.

### H\.2Cross\-Backbone Generalization on Qwen3

Table[12](https://arxiv.org/html/2608.12389#A8.T12)reports the complete cross\-model experiment underlying the Qwen3 result summarized in the main paper\. Every method uses the same Qwen3\-8B checkpoint\(Qwen Team[2025](https://arxiv.org/html/2608.12389#bib.bib129)\), target support examples, greedy decoding configuration, and evaluation instances; backbone\-dependent adapters and the soft\-token projector are reinitialized and source\-trained rather than copied from the LLaMA experiment\. The complete method attains70\.970\.9WR and the smallest cross\-domain degradation, while improving all three fine\-grained dimensions\. This matched replacement isolates portability across model families from differences in data or evaluation\.

Table 12:Cross\-domain personalization on the Qwen3\-8B backbone\. All values are from the matched Qwen3 experiment; WR is mean±\\pmstandard deviation over ten seeds\.
### H\.3Per\-Domain Personalization Specificity

Table[13](https://arxiv.org/html/2608.12389#A8.T13)expands the HiCUPID Personal Gap result by domain\. The complete method has the largest judge\-independent specificity margin in every target domain; its target\-domain mean is0\.0910\.091, matching the 5\-shot aggregate in the main\-paper ablation table after rounding\. The particularly large Music and Beauty margins show that the WR gain is accompanied by stronger distinction from generic answers rather than only improved fluency\.

Table 13:HiCUPID 5\-shot Personal Gap by domain\. Higher values indicate greater semantic proximity to the personalized target relative to the generic alternative\.
### H\.4Sample Efficiency and Parameter Displacement

Sample efficiency is evaluated atK∈\{0,1,3,5,10\}K\\in\\\{0,1,3,5,10\\\}using nested support sets: theK=1K=1example is contained in theK=3K=3set, and so forth\. This prevents differences between shot levels from being driven by unrelated sampled examples\. Alongside WR, we measure normalized parameter displacement

DK=\(∑l‖θK,l∗−θ0,l‖22\)1/2\(∑l‖θ0,l‖22\)1/2\+ϵ\.D\_\{K\}=\\frac\{\\left\(\\sum\_\{l\}\\\|\\theta\_\{K,l\}^\{\\ast\}\-\\theta\_\{0,l\}\\\|\_\{2\}^\{2\}\\right\)^\{1/2\}\}\{\\left\(\\sum\_\{l\}\\\|\\theta\_\{0,l\}\\\|\_\{2\}^\{2\}\\right\)^\{1/2\}\+\\epsilon\}\.\(47\)The intended behavior is not uniformly smallDKD\_\{K\}\. We test whether the controlled objective yields smaller displacement than vanilla Meta\-LoRA when evidence is sparse or uncertain and whether this gap narrows as support evidence increases\.

Figure[5](https://arxiv.org/html/2608.12389#A8.F5)reports the performance trajectory used in the main paper\. The proposed method increases from54\.754\.7WR atK=0K=0to66\.966\.9with one interaction and reaches approximately90%90\\%of its 10\-shot performance byK=3K=3\. The nested\-support construction makes each increase attributable to additional evidence rather than a different draw of support examples\.

![Refer to caption](https://arxiv.org/html/2608.12389v1/x4.png)Figure 5:Win Rate acrossK∈\{0,1,3,5,10\}K\\in\\\{0,1,3,5,10\\\}\. Every larger support set contains the smaller one for the same user, domain, and seed\.
### H\.5Optimization Stability

Figure[6](https://arxiv.org/html/2608.12389#A8.F6)compares the source meta\-training trajectories of vanilla Meta\-LoRA and the complete evidence\-calibrated objective under identical batches and optimization settings\. After warmup, the complete objective attains a14\.1%14\.1\\%lower terminal loss and a14\.8%14\.8\\%lower rolling loss standard deviation\. This is an optimization diagnostic, not a separate generalization guarantee: it shows that the structured anchor changes the observed training dynamics in the intended direction, while the target\-domain ablations in the main paper establish its predictive contribution\.

![Refer to caption](https://arxiv.org/html/2608.12389v1/figure/pac-bayes.png)Figure 6:Matched source\-training dynamics\. \(A\) Total loss\. \(B\) Rolling loss standard deviation\. The vertical marker denotes the end of warmup\.
### H\.6Topology Metrics

Three metrics test complementary properties of the projection\. CKA measures global relation alignment\. Neighborhood Recall@kkis

NR​@⁡k=1\|𝒟\|​∑d∈𝒟\|𝒩kz​\(d\)∩𝒩kT​\(d\)\|k,\\operatorname\{NR@\}k=\\frac\{1\}\{\|\\mathcal\{D\}\|\}\\sum\_\{d\\in\\mathcal\{D\}\}\\frac\{\|\\mathcal\{N\}\_\{k\}^\{z\}\(d\)\\cap\\mathcal\{N\}\_\{k\}^\{T\}\(d\)\|\}\{k\},\(48\)where the two neighborhoods are constructed in the domain and projected\-token spaces\. Attention Relation Error is the normalized Frobenius error‖SA−Sz‖F2/\|𝒟\|2\\\|S^\{A\}\-S^\{z\}\\\|\_\{F\}^\{2\}/\|\\mathcal\{D\}\|^\{2\}\. Reporting all three prevents a high global alignment score from obscuring poor local retrieval structure or poor fidelity at the actual attention interface\.

Table 14:Endpoint effect of topology preservation under otherwise identical zero\-shot settings\.Component\-wise analysis shows complementary effects: CKA alignment primarily improves global relational consistency, whereas attention\-space alignment more directly reduces error after the tokens pass through the frozen query and key projections\. Combining both yields the highest CKA and Neighborhood Recall@55among the tested variants, reduces attention relation error from0\.310\.31to0\.070\.07, and improves zero\-shot WR from57\.657\.6to59\.559\.5\. The generation gain indicates that relation preservation is functionally useful rather than only improving geometric agreement\.

Figure[7](https://arxiv.org/html/2608.12389#A8.F7)provides the corresponding qualitative audit\. Neighborhood connectivity and block structure are retained after projection; localized deviations reflect the projective bottleneck and generation supervision rather than a pointwise identity map\.

![Refer to caption](https://arxiv.org/html/2608.12389v1/figure/soft_token_1_1.png)\(a\)Two\-dimensional layouts and mutual\-neighbor edges\.
![Refer to caption](https://arxiv.org/html/2608.12389v1/figure/soft_token_1_2.png)\(b\)Domain–domain cosine\-similarity matrices\.

Figure 7:Relations before and after topology\-preserving soft\-token projection\. The analysis complements CKA, Neighborhood Recall@55, and attention\-relation error with a direct visual comparison\.
### H\.7Attention Attribution to Domain Soft Tokens

To test whether the injected token is actually read, let𝒮T\\mathcal\{S\}\_\{T\}be the soft\-token positions and𝒜\\mathcal\{A\}all available context positions\. At layerlland decoding steptt, we measureal,t=∑j∈𝒮TAttnl,t→j/∑j∈𝒜Attnl,t→ja\_\{l,t\}=\\sum\_\{j\\in\\mathcal\{S\}\_\{T\}\}\\operatorname\{Attn\}\_\{l,t\\rightarrow j\}/\\sum\_\{j\\in\\mathcal\{A\}\}\\operatorname\{Attn\}\_\{l,t\\rightarrow j\}\. We compare the correct target\-domain token, a token from the least similar source domain, and an equal\-length non\-informative placeholder\. As shown in Figure[8](https://arxiv.org/html/2608.12389#A8.F8), the mean budgets are6\.5%6\.5\\%,3\.3%3\.3\\%, and1\.0%1\.0\\%, respectively\. The correct token receives a98%98\\%larger budget than the wrong\-domain token \(bootstrap95%95\\%interval:92%92\\%–104%104\\%\), showing semantic selectivity rather than a position\-only effect\.

![Refer to caption](https://arxiv.org/html/2608.12389v1/figure/soft_token_2.png)Figure 8:Attention assigned to correct, wrong\-domain, and non\-informative soft\-token conditions across layers, decoding positions, and the aggregate comparison\.
### H\.8Robustness to Source–Target Shift

The frozensentence\-transformers/all\-MiniLM\-L6\-v2encoder\(Reimers and Gurevych[2019](https://arxiv.org/html/2608.12389#bib.bib127)\), which is independent of every compared personalization method, embeds an equal\-sized held\-out set of interaction inputs from each domain\. Letμd\\mu\_\{d\}be theℓ2\\ell\_\{2\}\-normalized mean embedding\. Target–source proximity and shift are

s​\(dt,𝒟s\)\\displaystyle s\(d\_\{t\},\\mathcal\{D\}\_\{s\}\)=maxds∈𝒟s⁡cos⁡\(μdt,μds\),\\displaystyle=\\max\_\{d\_\{s\}\\in\\mathcal\{D\}\_\{s\}\}\\cos\(\\mu\_\{d\_\{t\}\},\\mu\_\{d\_\{s\}\}\),\(49\)shift⁡\(dt\)\\displaystyle\\operatorname\{shift\}\(d\_\{t\}\)=1−s​\(dt,𝒟s\)\.\\displaystyle=1\-s\(d\_\{t\},\\mathcal\{D\}\_\{s\}\)\.Equal sample sizes prevent domain frequency from changing the estimator’s variance\. The external encoder and held\-out interactions are not used by the domain graph or retrieval module\.

The most shifted evaluated targets are LaMP\-QA Philosophy \(s=0\.156s=0\.156\) and HiCUPID Food \(s=0\.302s=0\.302\)\. Relative to the strongest competing method in each setting, our methodimproves WR/CCS by15\.4%/13\.9%15\.4\\%/13\.9\\%on Philosophy and9\.6%/21\.6%9\.6\\%/21\.6\\%on Food\. Across the five HiCUPID held\-out domains, the mean cell\-wise relative WR gain is15\.5%15\.5\\%, compared with4\.1%4\.1\\%on the source\-domain average\. The larger improvement under the evaluated shifts supports the intended transfer behavior, but these observations are not a guarantee for arbitrary domains outside the support of the source data\.

### H\.9Latency Breakdown

Table[15](https://arxiv.org/html/2608.12389#A8.T15)reports matched end\-to\-end latency, including support\-statistic computation and all inner\-loop updates for the few\-shot systems\.

Table 15:Mean end\-to\-end latency per response under identical hardware, batch size, sequence length, and decoding settings\.The full method adds0\.220\.22seconds over vanilla Meta\-LoRA in this setting and therefore introduces no substantial deployment\-latency increase under the matched configuration\. Absolute latency remains hardware\- and decoding\-dependent\.

### H\.10Evaluator Agreement Results

Table[16](https://arxiv.org/html/2608.12389#A8.T16)compares the fixed Qwen3\-based scores against two method\-independent judgment sources on the same blinded samples\.

Table 16:Agreement of the Qwen3\-based metrics with method\-blinded human and independent GPT\-4o judgments\.All four Pearson correlations are high, indicating that the reported metrics track human\-perceived personalization quality and are robust to evaluator choice\. This evidence supports the validity of the measurement protocol; it does not imply that automatic evaluation can replace human assessment in every deployment setting\.

### H\.11Hyperparameter Sensitivity Protocol

Sensitivity is evaluated without target\-domain tuning\. Tables[17](https://arxiv.org/html/2608.12389#A8.T17)and[18](https://arxiv.org/html/2608.12389#A8.T18)report the source\-validation sweeps used to select the entropy coefficient and the initialization of the bounded, outer\-loop\-learned group variances\. The low–high–low initialization is not a manually fixed final schedule: the three values initialize trainable log\-variances, which are subsequently updated only on source episodes\. Moderate perturbations preserve a clear advantage, whereas a uniform or inverted initialization is substantially worse\. Likewise, both removing entropy modulation and over\-scaling it reduce performance, withα=1\\alpha=1providing the best joint WR/CCS result\. The absolute source\-validation scores are not held\-out target\-test estimates and are reported only to document selection and sensitivity\.

Table 17:Source\-validation sensitivity to early/middle/late variance initialization\. The variances remain trainable on source episodes\.Table 18:Source\-validation sensitivity to predictive\-entropy modulation\.Retrieval additionally evaluates candidate capsk∈\{1,3,5,8\}k\\in\\\{1,3,5,8\\\}and source\-selected edge thresholds around the default\. Topology alignment variesλattn\\lambda\_\{\\mathrm\{attn\}\}and the number of soft tokens while keeping the total source training budget fixed\. We report WR and CCS together with the number of retained neighbors and learned variance values, because performance alone cannot distinguish a robust region from a degenerate sparse graph\.

The graph\-threshold sweep in Figure[9](https://arxiv.org/html/2608.12389#A8.F9)isolates the retrieval precision–coverage trade\-off\. CCS peaks at the source\-selected threshold0\.650\.65\(1\.45±0\.031\.45\\pm 0\.03\), where an average of3\.23\.2source neighbors is retained\. Lower thresholds admit unreliable relations; thresholds above0\.80\.8frequently leave a singleton or empty transferable neighborhood\. Performance is stable in the0\.600\.60–0\.700\.70region, indicating that the reported result is not tied to a knife\-edge choice\.

![Refer to caption](https://arxiv.org/html/2608.12389v1/figure/soft_token_3.png)Figure 9:Source\-validation sensitivity to the graph edge threshold\. The left panel reports CCS; the right panel tracks the retained\-neighborhood behavior and resistance to irrelevant\-domain injection\.
### H\.12Case Study

Table[19](https://arxiv.org/html/2608.12389#A8.T19)contrasts a meta\-learning baseline with the complete framework on one held\-out Food\-domain query; the preference cues are reconstructed from history and shown only for interpretation\.

Table 19:Illustrative Food\-domain case\. Preference cues are shown only for analysis and are not supplied as oracle input\.The two baselines recover the broad urban\-design topic but omit the jointly supported history, detail\-orientation, and bubble\-tea cues\. The complete method coordinates those cues in one domain\-appropriate suggestion\. This is qualitative evidence only; aggregate conclusions are based on the quantitative test sets\.

### H\.13Result Scope and Traceability

The main\-paper tables are the authoritative report of target\-domain predictive performance\. Appendix H adds the full Qwen3 cross\-backbone table and analyses of components in the submitted method: nested\-shot efficiency, source\-training stability, topology and attention diagnostics, source–target shift, latency, evaluator agreement, source\-validation sensitivity, and the qualitative case\. The code archive fixes the final backbone identifiers, domain splits, seeds, metric formulas, and component configurations, providing one traceable specification for every reported analysis\.

## Appendix IReproducibility, Limitations, and Responsible Use

### I\.1Reproducibility Package

The accompanying materials contain this technical appendix and a separate code archive\. Table[20](https://arxiv.org/html/2608.12389#A9.T20)maps its contents to the paper\. The archive uses relative paths and includes an English README plus a short evaluation guide with executable entry points for preprocessing, source\-only prior construction, episodic training, target adaptation, evaluation, statistical testing, and core\-equation verification\.

Table 20:Manifest of the accompanying code archive\. The listed files are contained in the released artifact\.The archive additionally contains:

- •deterministic utilities that create source/target and nested\-support manifests from public benchmark files;
- •the exact environment, library versions, hardware record, final configuration, and complete development search space;
- •source\-training, target\-adaptation, and evaluation entry points;
- •evaluator prompts, response\-order randomization code, and tie policy;
- •the ten declared random seeds and implementations of paired confidence intervals and corrected tests;
- •a synthetic smoke test that verifies the central equations without redistributing benchmark text or model weights\.

Original code is licensed under Apache\-2\.0\. Dataset files, pretrained checkpoints, and third\-party baselines are not relicensed or redistributed; each remains governed by its public release terms\. The code archive uses relative paths and excludes local filesystem paths, private repository references, and machine\-specific metadata\. The included tests and smoke example run independently of benchmark text and pretrained model downloads\.

### I\.2Checklist\-to\-Artifact Cross\-Reference

The conceptual method and complete inference path requested by the reproducibility checklist are provided in Appendices[C](https://arxiv.org/html/2608.12389#A3)through[F](https://arxiv.org/html/2608.12389#A6)\. The status and scope of the theoretical statements are delineated in Appendix[C](https://arxiv.org/html/2608.12389#A3); Proposition 1 is a standard randomized\-predictor PAC\-Bayes result, whereas the entropy\-, Fisher\-, and depth\-calibrated deterministic objective is an empirically evaluated surrogate rather than an additional guarantee\. Dataset motivation and splits, baseline settings, metric definitions, development ranges and selection criterion, randomness controls, computing infrastructure, run counts, variation, and statistical tests are recorded in Appendix[G](https://arxiv.org/html/2608.12389#A7)\. The archive paths supporting preprocessing, training, evaluation, analysis, licensing, and code\-to\-paper comments are listed in Table[20](https://arxiv.org/html/2608.12389#A9.T20)\.

### I\.3Limitations

The PAC\-Bayes statement concerns a randomized predictor under i\.i\.d\. bounded\-loss assumptions; deployment uses a deterministic surrogate\. Predictive entropy can be miscalibrated, and a squared\-gradient Fisher proxy is only a local sensitivity estimate\. The domain prior can fail when descriptions are ambiguous or every source domain is irrelevant\. A textual preference prompt may omit subtle or changing preferences, while continuous domain tokens are not directly human\-readable\. With no pre\-target history, the method falls back to a neutral textual prompt and domain conditioning, so it cannot infer genuinely user\-specific tendencies until user evidence becomes available\. These limitations constrain the claims and motivate calibration, abstention, and longitudinal evaluation\.

### I\.4Privacy and Data Governance

The experiments use anonymized benchmark signals\. The method does not require cross\-user graph propagation in its current form\. Deployment should be opt\-in and should provide a way to inspect, edit, or delete the textual user prompt and associated target adapters\. Raw histories should be retained only as required by the application, and cached prompts or adapters should be access\-controlled and scoped to one user\. The method does not constitute a formal privacy guarantee; differential privacy, membership\-inference resistance, and secure deletion remain separate engineering and research requirements\.

### I\.5Safety and Misuse

Personalization may amplify an incorrect inferred preference, create filter bubbles, or overfit sensitive attributes\. The generation policy must continue to obey task\-level safety constraints even when a user prompt or domain token suggests otherwise\. Applications should distinguish user\-editable preferences from protected or sensitive inferences, use conservative defaults when evidence is weak, and permit non\-personalized fallback\. Human agreement with an evaluator supports metric validity but does not establish safety in high\-stakes domains\.

相似文章

MuLoRA:用于持续学习的谱均衡低秩适配方法

arXiv cs.LG

MuLoRA 提出了一种面向持续学习的谱均衡低秩适配方法。该方法识别出 LoRA 更新中的「谱可塑性坍缩」现象,并通过历史白化和基于动量的正交化加以应对,在类别增量基准的 16 项指标中有 15 项取得了最佳平均准确率。