PersonaEdit: Representative Sample Selection for Personalized Model Editing

You-Mei Huang; Chung-Chi Chen; An-Zi Yen

cs.CLAug 28, 2026

PersonaEdit: Representative Sample Selection for Personalized Model Editing

Authors: You-Mei Huang, Chung-Chi Chen, An-Zi Yen

Organizations: Department of Computer Science, National Yang Ming Chiao Tung University, Taiwan · Human-Agent Ally Lab (HAA Lab), National Institute of Informatics (NII), Japan

Abstract

Personalization has attracted growing interest in LLM applications, yet existing retrieval-based approaches depend heavily on retrieval quality and degrade in long-term interactions. Model editing, which directly modifies internal model parameters to incorporate new knowledge, has demonstrated effective knowledge modification capabilities in factual knowledge editing tasks and may provide a potential solution for personalization. However, scaling model editing to personalization is non-trivial. Editing large amounts of user data increases computational cost and causes interference among edits, motivating the need for effective sample selection. To address this issue, we propose, PersonaEdit, a hidden representation clustering strategy that selects representative editing samples through proportional stratified sampling. Experiments show that model editing is effective for personalization, and that our selection strategy preserves most of the performance while substantially reducing the number of required editing samples. Beyond standalone editing, we find that combining model editing with retrieval-based prompt augmentation further improves personalization, as edited knowledge and retrieved context provide complementary information. These results demonstrate the potential of model editing as an efficient and scalable approach for LLM personalization.

Figures & tables

Method	Accuracy	Gain	# Gain
Vanilla	33.38%	–	–
Few-shot ${}_{\text{BM25}}$	43.00%	9.67%	30
Few-shot ${}_{\text{contriever}}$	42.84%	9.49%	30
Persona	41.44%	8.05%	30
All info ${}_{\text{BM25}}$	44.60%	11.28%	30
All info ${}_{\text{contriever}}$	44.51%	11.15%	30

Table 1: Evaluation results on the OpinionQA dataset. Bold and underlined values indicate the best and second-best performance, respectively. “Gain” denotes the average accuracy improvement per respondent compared to the Vanilla baseline, and “# Gain” indicates the number of respondents with performance gains.

Prompt	Vanilla	PersonaEdit	Gain	# Gain
w/o prompt	33.38%	44.67%	11.31%	30
FS ${}_{\text{(BM25)}}$	43.00%	45.35%	2.72%	25
Persona	41.44%	45.73%	3.96%	25
All info ${}_{\text{(BM25)}}$	44.60%	47.00%	2.51%	25

Table 2: Comparative analysis of Vanilla and PersonaEdit across different prompt augmentation strategies on the OpinionQA dataset. Bold and underlined values indicate the best and second-best performance within columns. FS denotes few-shot prompting. “Gain” denotes the average accuracy improvement per respondent brought by model editing, and “# Gain” indicates the number of respondents who achieved gains.

Prompt	PersonaEdit	GPT-5.4 mini	GPT-5.4
w/o prompt	44.67%	43.89%	37.63%
Few-shot ${}_{\text{(BM25)}}$	45.35%	47.64%	48.39%
Persona	45.73%	47.01%	49.67%
All info ${}_{\text{(BM25)}}$	47.00%	49.72%	51.51%
Total Cost	–	$9.63	$57.58

Table 3: Comparison of personalization performance between PersonaEdit and closed-source GPT-5.4 mini and GPT-5.4 models.

Method	Accuracy	MMAE ( $\downarrow$ )	CEM ( $\uparrow$ )
Vanilla	35.48%	1.0728	0.5356
Few-shot ${}_{\text{(BM25)}}$	41.47%	1.0842	0.5370
Few-shot ${}_{\text{(Contriever)}}$	40.33%	1.0802	0.5381
Persona	38.52%	1.0952	0.5377
All info ${}_{\text{(BM25)}}$	42.50%	1.1067	0.5350
All info ${}_{\text{(Contriever)}}$	40.97%	1.0986	0.5366

Table 4: Evaluation results on the OpinionQA dataset focusing on questions with ordinal options. Bold and underlined values indicate the best and second-best performance within each column, respectively.

\uparrow

indicates higher values are better, while

\downarrow

indicates lower values are better.

Model	Method	Eff.	Para.	Flu.	Loc.
LLaMA	Vanilla	39.65	41.02	641.15	62.85
	ROME	39.28	40.83	634.62	62.51
	MEMIT	89.67	64.68	631.45	62.95
	AlphaEdit	89.47	66.33	630.93	62.88
GPT-J	Vanilla	27.18	23.92	621.41	45.45
	ROME	28.32	23.68	618.69	45.53

Table 5: Comparison of model editing methods across different base models. Locality is reported as the average performance across GLUE and MMLU benchmarks. Eff., Para., Flu., and Loc. denote efficacy, paraphrase, fluency, and locality, respectively. Bold and underlined values indicate the best and second-best performance within each model group, respectively (excluding the Vanilla baseline for Fluency ranking).

Evaluation Set	Mean $\Delta$ Acc.	Std	Min $\Delta$	Max $\Delta$	# Improved	# Degraded	# Unchanged
Edit set	+49.58	10.09	+30.83	+61.67	10	0	0
Same-topic hold-out	+1.60	11.07	-11.76	+22.58	3	4	3
Other-topic hold-out	+3.54	2.85	-1.32	+7.79	9	1	0

Table 6: Editing influence across different hold-out settings under topic-based sample selection. Mean, standard deviation, minimum, and maximum

\Delta

Acc. are computed across 10 respondents. # Improved, # Degraded, and # Unchanged indicate the number of respondents (N=10).

Figure 1: Semantic similarity visualization illustrating the transfer effects of model editing. Red triangles denote the edited samples used for model editing, while black markers represent other samples from the same semantic topic category. Circles indicate samples whose predictions improved after editing, whereas crosses denote samples whose predictions became incorrect after editing.

Figure 2: Hidden representation visualization illustrating the transfer effects of model editing. Red triangles denote the edited samples used for model editing. Circles indicate samples whose predictions improved after editing, whereas crosses denote samples whose predictions became incorrect after editing. Different colors represent clusters constructed from hidden representation similarity.

Figure 3: Test accuracy improvement under different clustering hyperparameters $K$ and editing data sizes.

Group	Before	After	Improvement
Low Accuracy	0.280	0.401	+0.121
Mid Accuracy	0.335	0.447	+0.113
High Accuracy	0.387	0.492	+0.105

Table 7: Average editing performance before and after editing across different accuracy groups.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Component	Runtime
AlphaEdit projection (one-time)	$\sim$ 25 hours
Editing 200 samples	$\sim$ 2 hours

Appendix

Table 8: Runtime of the model editing pipeline on an NVIDIA H100 GPU. The AlphaEdit projection computation is performed once and reused across editing experiments.

Model	Method	SST	MRPC	CoLA	RTE	NLI	MMLU
LLaMA 3.1 8B	Vanilla	95.99 ±0.00	69.81 ±0.00	69.38 ±0.00	22.77 ±0.00	61.62 ±0.00	57.52 ±0.00
	ROME	96.13 ±1.41	69.52 ±1.52	68.75 ±0.62	22.82 ±1.61	62.44 ±1.49	55.42 ±1.58
	MEMIT	95.86 ±0.34	66.10 ±2.56	76.37 ±2.80	21.18 ±0.92	61.16 ±0.92	57.04 ±1.99
	AlphaEdit	95.79 ±0.84	64.42 ±1.30	78.29 ±2.09	21.77 ±0.97	61.21 ±1.66	55.78 ±1.78
GPT-J-6B	Vanilla	75.00 ±0.00	42.07 ±0.00	38.56 ±0.00	48.33 ±0.00	50.00 ±0.00	18.75 ±0.00
	ROME	75.08 ±0.28	43.04 ±0.86	38.56 ±0.00	47.48 ±0.00	49.30 ±0.46	19.74 ±0.55

Appendix

Table 9: Detailed evaluation of locality performance across individual text classification and language understanding benchmarks. Bold and underlined values indicate the best and second-best performance within each model group, respectively.

User	Edit $\Delta$	Same-Topic $\Delta$	Cross-Topic $\Delta$
User 1	+0.617	+0.114	+0.042
User 2	+0.375	-0.125	+0.062
User 3	+0.567	+0.226	+0.020
User 4	+0.458	-0.067	+0.009
User 5	+0.417	-0.095	-0.016
User 6	+0.508	-0.040	+0.013

Appendix

Table 10: Per-user editing performance changes on edited, same-topic, and cross-topic samples before and after editing.

Compared Method	Baseline Mean	Method Mean	Improvement	Wilcoxon p-value
PersonaEdit	0.4464	0.4467	+0.0003	0.940
PersonaEdit + persona	0.4464	0.4537	+0.0073	0.280
PersonaEdit + Few-shot ${}_{\text{BM25}}$	0.4464	0.4575	+0.0111	0.058
PersonaEdit + All info ${}_{\text{BM25}}$	0.4464	0.4702	+0.0238	0.00013

Appendix

Table 11: User-level personalization accuracy across methods using BM25 retrieval. Baseline Mean and Method Mean represent the mean user-level accuracy averaged across respondents. All p-values are computed against the All info

{}_{\text{BM25}}

baseline using paired Wilcoxon signed-rank tests.

Compared Method	Baseline Mean	Method Mean	Improvement	Wilcoxon p-value
PersonaEdit	0.4451	0.4467	+0.0016	0.657566
PersonaEdit + persona	0.4451	0.4537	+0.0086	0.036921
PersonaEdit + Few-shot ${}_{\text{BM25}}$	0.4451	0.4575	+0.0124	0.017897
PersonaEdit + All info ${}_{\text{BM25}}$	0.4451	0.4702	+0.0251	0.00000514

Appendix

Table 12: User-level personalization accuracy across methods using Contriever retrieval. Baseline Mean and Method Mean represent the mean user-level accuracy averaged across respondents. All p-values are computed against the All info

{}_{\text{contriever}}

baseline using paired Wilcoxon signed-rank tests.

Compared Method	Baseline Mean	Method Mean	Improvement	Wilcoxon p-value
All info ${}_{\text{BM25}}$ vs PersonaEdit	0.4254	0.4288	+0.0034	0.626346
All info ${}_{\text{BM25}}$ vs PersonaEdit + persona	0.4254	0.4271	+0.0017	0.597808
All info ${}_{\text{BM25}}$ vs PersonaEdit + Few-shot ${}_{\text{BM25}}$	0.4254	0.4368	+0.0114	0.145999
All info ${}_{\text{BM25}}$ vs PersonaEdit + All info ${}_{\text{BM25}}$	0.4254	0.4485	+0.0231	0.000555

Appendix

Table 13: User-level ordinal accuracy comparison across methods. Baseline Mean and Method Mean represent the mean user-level accuracy averaged across respondents. All p-values are computed using paired Wilcoxon signed-rank tests against the corresponding retrieval baseline.

Compared Method	Baseline Mean	Method Mean	Improvement	Wilcoxon p-value
All info ${}_{\text{BM25}}$ vs PersonaEdit	0.2936	0.2831	+0.0105	0.073244
All info ${}_{\text{BM25}}$ vs PersonaEdit + persona	0.2936	0.2802	+0.0133	0.004338
All info ${}_{\text{BM25}}$ vs PersonaEdit + Few-shot ${}_{\text{BM25}}$	0.2936	0.2834	+0.0102	0.007612
All info ${}_{\text{BM25}}$ vs PersonaEdit + All info ${}_{\text{BM25}}$	0.2936	0.2764	+0.0172	$4.97\times 10^{-5}$

Appendix

Table 14: User-level macro-averaged MAE comparison across methods. Baseline Mean and Method Mean represent the mean user-level accuracy averaged across respondents. All p-values are computed using paired Wilcoxon signed-rank tests against the corresponding retrieval baseline.

Compared Method	Baseline Mean	Method Mean	Improvement	Wilcoxon p-value
All info ${}_{\text{BM25}}$ vs PersonaEdit	0.5350	0.5383	+0.0033	0.013663
All info ${}_{\text{BM25}}$ vs PersonaEdit + persona	0.5350	0.5380	+0.0030	0.027741
All info ${}_{\text{BM25}}$ vs PersonaEdit + Few-shot ${}_{\text{BM25}}$	0.5350	0.5385	+0.0035	0.000111
All info ${}_{\text{BM25}}$ vs PersonaEdit + All info ${}_{\text{BM25}}$	0.5350	0.5371	+0.0021	0.004665

Appendix

Table 15: User-level CEM comparison across methods. Baseline Mean and Method Mean represent the mean user-level accuracy averaged across respondents. All p-values are computed using paired Wilcoxon signed-rank tests against the corresponding retrieval baseline.

Compared Method	Baseline Mean	Method Mean	Improvement	Wilcoxon p-value
Vanilla vs PersonaEdit	0.3337	0.4467	+0.1131	$1.73\times 10^{-6}$
Few-shot ${}_{\text{BM25}}$ vs PersonaEdit + Few-shot ${}_{\text{BM25}}$	0.4304	0.4575	+0.0272	$2.35\times 10^{-6}$
Persona vs PersonaEdit + persona	0.4141	0.4537	+0.0396	$5.14\times 10^{-6}$
All info ${}_{\text{BM25}}$ vs PersonaEdit + All info ${}_{\text{BM25}}$	0.4464	0.4702	+0.0238	0.00013

Appendix

Table 16: User-level option accuracy comparison between retrieval-based personalization baselines and their corresponding model editing variants. All p-values are computed using paired Wilcoxon signed-rank tests.

K = 8
	20		120		200
Split	Pre-edit	Edited	Pre-edit	Edited	Pre-edit	Edited
Train	30.67%	85.67%	34.67%	88.75%	34.22%	89.45%
Test	33.37%	37.37%	33.37%	41.75%	33.37%	44.47%
K = 13
Train	28.83%	92.83%	33.11%	88.50%	34.57%	89.07%

Appendix

Table 17: Analysis of the clustering hyperparameter K and the scale of edited samples (20, 120, and 200) in terms of accuracy.

Setting	Method	Efficacy ( $\uparrow$ )	Paraphrase ( $\uparrow$ )	Fluency ( $\uparrow$ )	Locality ( $\uparrow$ )
Setting	Method	Efficacy ( $\uparrow$ )	Paraphrase ( $\uparrow$ )	Fluency ( $\uparrow$ )	SST	MRPC	CoLA	RTE	NLI	MMLU
$k=8,s=20$	Vanilla	40.17 ±13.57	42.33 ±12.43	638.76 ±0.00	95.99 ±0.00	69.81 ±0.00	69.38 ±0.00	22.77 ±0.00	61.62 ±0.00	57.52 ±0.00
$k=8,s=20$	AlphaEdit	90.33 ±7.52	62.00 ±11.15	630.02 ±2.95	96.13 ±0.43	65.51 ±1.91	68.83 ±1.02	20.72 ±0.66	62.03 ±0.69	55.57 ±1.41
$k=8,s=120$	Vanilla	39.56 ±6.95	42.50 ±7.39	640.94 ±0.00	95.99 ±0.00	69.81 ±0.00	69.38 ±0.00	22.77 ±0.00	61.62 ±0.00	57.52 ±0.00
$k=8,s=120$	AlphaEdit	89.56 ±2.96	64.25 ±5.91	632.24 ±1.08	96.29 ±0.69	63.54 ±1.81	78.13 ±2.35	20.97 ±0.91	60.38 ±1.27	54.54 ±2.06
$k=8,s=200$	Vanilla	39.88 ±5.94	40.32 ±6.67	640.98 ±0.00	95.99 ±0.00	69.81 ±0.00	69.38 ±0.00	22.77 ±0.00	61.62 ±0.00	57.52 ±0.00

Appendix

Table 18: Editing results of AlphaEdit under various

k

and

s

hyperparameter configurations on the LLaMA 3.1 8B Instruct model across efficacy, generalization, fluency, and individual locality tasks.

LLM Personalization

LLM personalization adapts language models to users’ preferences, histories, expertise, and changing goals while balancing usefulness, privacy, and control.

See all 130 papers

Unlocking Latent Personalization in LLMs
Sep 27, 2026Wei Chen, Guanghui Zhu, Zhongliang Cai +1LLM Personalization Personalized Language Models
State Key Laboratory for Novel Software Technology, Nanjing University
Beyond Retrieval: Learning Compact User Representations for Scalable LLM Personalization
Jun 3, 2026Heng Cao, Fan Zhang, Jian Yao +8LLM Personalization Prompt Tuning
*Microsoft · †The Hong Kong Polytechnic University · ⋄Shanghai International Studies University +1
Iterative Critique-Refine Framework for Enhancing LLM Personalization
Oct 28, 2025Durga Prasad Maram, Dhruvin Gandhi, Zonghai Yao +5LLM Personalization Personalized Language Models
University of Massachusetts Amherst · Adobe Research · University of Oregon +1

Sep 27, 2026J/K move · Enter open · S save

Unlocking Latent Personalization in LLMs

Wei Chen, Guanghui Zhu, Zhongliang Cai, Yihua Huang

State Key Laboratory for Novel Software Technology, Nanjing University

LLM Personalization Personalized Language Models User Preference Modeling

Open paper AskPDF arXiv

Large language models (LLMs) are increasingly expected to adapt to individual users, yet effective personalization remains challenging when only limited user-specific samples are available. In this work, we take an alternative perspective: pretrained LLMs may already possess latent capacity for personalization, and a few user samples may therefore suffice to guide the model toward user-aligned behavior with minimal user-specific adaptation. From this perspective, we propose LatentPersonal, a framework that formulates personalization as navigation in a shared latent adaptation space. LatentPersonal infers a compact latent representation from a few user samples to guide user-specific model adaptation, regularized with a variational information bottleneck to encourage compact preference representations. We instantiate LatentPersonal with LoRA, leveraging its low-rank parameterization as a natural low-dimensional adaptation space for personalization. By simply inserting a user-specific guidance vector between the shared low-rank factors, the model can navigate toward personalized adaptations through lightweight inference of this compact representation, without updating the shared LoRA parameters. Experiments across multiple personalization datasets demonstrate that LatentPersonal substantially reduces user-specific adaptation overhead while achieving effective personalization from only a few user-specific interactions, with particularly strong performance in the one-shot regime.

Abstract

LLM Personalization

Explore similar work

Unlocking Latent Personalization in LLMs

Beyond Retrieval: Learning Compact User Representations for Scalable LLM Personalization

Iterative Critique-Refine Framework for Enhancing LLM Personalization