PaperlayerView exact revision on arXiv

Source-linked AI summary

PersonaEdit: Representative Sample Selection for Personalized Model Editing

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

arXiv:2608.27816v1cs.CL

TL;DR

Personalization methods can depend on retrieval quality, limited context windows, or externally maintained information, while editing many user preferences is costly and can cause interference. PersonaEdit clusters hidden representations and samples proportionally across clusters to select representative edits. Experiments show that this preserves most personalization performance with fewer samples, complements retrieval augmentation, and remains limited by a narrow evaluation setting and deployment challenges.

  • Problem

    Existing retrieval-based personalization degrades in long-term interactions, while scaling model editing to many preferences increases computational cost and can cause interference.

  • Method

    PersonaEdit clusters hidden representations and uses proportional stratified sampling to select representative editing samples.

  • Results

    Experiments show that hidden representation-based selection reduces required editing samples while preserving most personalization performance, and retrieval augmentation complements model editing.

  • Takeaways & Limitations

    Model editing shows potential as an efficient and scalable approach for LLM personalization when combined with representative sample selection and retrieval-based prompt augmentation.

  • Takeaways & Limitations

    Evaluation primarily uses 30 OpinionQA users on social and political topics in a discrete option-prediction setting, limiting coverage of open-ended personalization scenarios.

Abstract

from arXiv · show

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.

1 Introduction

PersonaEdit applies model editing to personalization and selects representative editing samples using hidden representations. The approach improves sample efficiency while preserving personalization capability, and user characteristics influence effectiveness.

  • Existing retrieval-based personalization depends on retrieval quality and prompt engineering, is constrained by context windows, and degrades during long-term interactions.
  • Model editing directly modifies parameters to incorporate user-specific knowledge and preferences, potentially enabling more persistent personalized behavior.
  • Personalization performance varies across users and is influenced by stance consistency and preference discrepancies between users and models.
  • Hidden representation clustering selects representative editing samples, reducing the required sample count while preserving personalization performance.

2 Related Work

Prior personalization methods rely on prompts, retrieval, or externally maintained latent information, while model editing incorporates knowledge into model parameters. PersonaEdit addresses the scaling challenges of editing many personalized preferences through representative sample selection.

  • LLM Personalization: Prompt-based methods are training-free but depend strongly on prompt design quality and in-context reasoning capabilities.
  • LLM Personalization: Retrieval-based methods depend on retrieval quality, face context-window constraints, and degrade in long-term interactions.
  • LLM Personalization: Latent personalization methods require personalized information to be provided or maintained externally rather than incorporated directly into the model.
  • Model Editing: Editing many personalized preferences increases computational overhead and may worsen catastrophic forgetting and reduce scalability.
  • Model Editing: PersonaEdit uses hidden-representation clustering to select compact representative preferences before editing, unlike prior clustering for preference augmentation.

3 PERSONAEDIT

PersonaEdit clusters samples by hidden representations and uses proportional stratified sampling to construct a compact editing set. Samples nearest cluster centroids are selected to preserve representation-space coverage while reducing sampling bias.

  • Motivation: Editing effects can transfer to related inputs, while similar subjects may create interference across editing outcomes.
  • Hidden Representation Extraction: The procedure extracts an intermediate-layer hidden representation for each sample’s subject token, aligned with the layer used for subsequent editing.
  • Clustering via K-Means: K-means partitions hidden vectors into K disjoint clusters by minimizing intra-cluster variance, grouping prompts with similar hidden representations.
  • Proportional Stratified Sampling: Proportional stratified sampling allocates selected samples across clusters to maintain representation-space coverage and reduce sampling bias.
  • Proportional Stratified Sampling: Within each cluster, samples nearest the centroid are selected to form the final editing set.

4 Experiments

Experiments on OpinionQA evaluate model editing, representative sample selection, prompt augmentation, and comparisons with retrieval-based, prompt-based, and closed-source personalization methods. Model editing performs competitively, improves with augmentation, and shows stronger ordinal-metric results, while FERMI comparisons are limited by model, output-format, data-scale, and implementation differences.

  • Main Personalization Results: 47.00% accuracy is achieved by combining model editing with All-info, outperforming all baseline approaches; model editing gains occur for all 30 respondents.Accuracy is determined by matching the option in the model output.
  • Effect of Prompt Augmentation: Model editing consistently outperforms the non-edited model across prompt settings, with statistically significant improvements and gains for 25 of 30 respondents.This evaluates combinations of editing with retrieval-based and prompt-based augmentation strategies.
  • Comparison with Frontier Closed-Source Models: GPT-5.4 and GPT-5.4 mini reach 51.51% and 49.72% accuracy under All-info(BM25), compared with 47.00% for PersonaEdit.The comparison notes that prompting-based personalization requires repeatedly providing user information during inference or risks persona forgetting in multiterm interactions.
  • Ordinal Evaluation: Model editing significantly outperforms other methods on accuracy and MMAE, while Model Editing with Few-shot(BM25) significantly outperforms others on CEM.MMAE measures distance between predicted and ground-truth labels, while CEM accounts for informational closeness between ordinal classes.

5 Discussion

The discussion evaluates model editing, sample-selection strategies, clustering settings, and user factors affecting personalized performance. Hidden-representation clustering improves sample efficiency while preserving personalization, but editing outcomes also depend on data volume and user characteristics.

  • Model Editing Effectiveness: Across LLaMA 3.1 8B Instruct and GPT-J-6B, MEMIT and AlphaEdit generally outperform ROME on efficacy, paraphrase, and fluency, while locality differences remain small.AlphaEdit significantly improves paraphrase robustness on LLaMA 3.1 8B Instruct and several metrics on GPT-J-6B.
  • Sample Selection Strategies: Human-annotated topics do not reliably characterize editing generalization: same-topic edits produce both improvements and degradations across respondents.Semantic similarity also lacks a consistent monotonic relationship with editing influence, including substantial effects for some low-similarity samples.
  • Sample Selection Strategies: Hidden representation similarity better characterizes editing transfer behavior, motivating clustering-based selection of representative editing samples.The strategy groups samples by similarity in the model’s internal representations rather than by human-annotated topics.
  • Sample Selection Strategies: 0.3823 vs. 0.3756: clustering-based selection achieves higher test accuracy than random sampling with N=300 and M=20.The improvement is statistically significant across respondents.
  • Clustering and Sample Size Effects: Approximately 5%–7% gains result as editing samples increase, but increasing samples from 120 to 200 adds only approximately 1%–2%, while changing K from 8 to 18 causes mostly 1%–2% fluctuations.These results indicate that editing-data volume matters more than the clustering hyperparameter K, with moderate representative samples preserving most performance.
  • User Factors: Model editing benefits all user groups, with the largest improvement for the Low Accuracy group and smaller gains for the High Accuracy group.Response consistency correlates positively with personalization improvement on training data, with ρ = 0.4218 and p = 0.0202.

6 Conclusion

The paper finds that hidden representation similarity is a more effective basis for selecting editing samples than human-annotated topics. PersonaEdit reduces the required editing samples while preserving most personalization performance.

  • Conclusion: Hidden representation similarity provides a more effective basis for editing sample selection than human-annotated semantic topics on OpinionQA.The conclusion links this finding to editing generalization and personalization performance.
  • Conclusion: The proposed clustering strategy reduces required editing samples while preserving most personalization performance.The study frames this as evidence for efficient sample selection in personalization-oriented model editing.
  • Conclusion: The current study is limited to a single benchmark and restricted editing settings.The authors identify broader personalization scenarios as a direction for future work.

Limitations

The paper’s limitations concern evaluation scope, incomplete comparisons with alternative selection strategies, model coverage, and deployment assumptions.

  • Evaluation Scope: Evaluation primarily uses OpinionQA, 30 users, social and political topics, and discrete option prediction rather than open-ended personalization settings.The authors note that long-form dialogue, continual interaction, and dynamic preference shifts are not fully captured.
  • Comparison Scope: The study does not extensively compare hidden representation clustering with other sample-selection strategies, including external text-embedding methods.Broader comparisons would better assess the advantages of model-internal representations.
  • Model Scope: Experiments use a limited set of open-weight language models, so generalizability across architectures and training settings remains limited.The authors explicitly caution that editing behavior may vary across model architectures and training settings.
  • Deployment Scope: Per-user editing is evaluated independently under controlled conditions without a concrete deployment architecture.Practical deployment may involve update frequency, rollback, privacy, and cumulative-interference challenges when maintaining separate edited models or deltas.

C Locality Results

Locality is evaluated across six language-understanding and text-classification benchmarks for both studied base models.

  • Locality Evaluation: Locality results cover SST, MRPC, CoLA, RTE, NLI, and MMLU for LLaMA 3.1 8B Instruct and GPT-J-6B.The table supplements the main-text locality summary with per-task performance.

D Same-Topic and Cross-Topic Editing Results

Table 10 evaluates editing on edited, same-topic, and cross-topic samples, with results reported separately for each user.

  • Evaluation covers edited samples, same-topic samples, and cross-topic samples.Same-topic questions share the edited question’s topic category; cross-topic questions come from different topic categories.
  • Same-topic samples test generalization within the edited question’s topic category.
  • Cross-topic samples test generalization beyond the edited question’s topic category.

E Significance Test Results

The paper uses paired Wilcoxon signed-rank tests to compare personalization methods with retrieval baselines. Combined model editing, persona information, and retrieval-based few-shot examples produce the strongest reported gains.

  • Paired Wilcoxon signed-rank tests evaluate user-level personalization across option accuracy, ordinal accuracy, macro-averaged MAE, and CEM.
  • Comparisons test personalization strategies including model editing, persona augmentation, and retrieval-based few-shot prompting against corresponding retrieval baselines.
  • Combined model editing, persona information, and retrieval-based few-shot examples achieve the strongest improvements across most evaluation metrics.
  • The combined setting shows statistically significant gains in ordinal accuracy and macro-averaged MAE over both BM25 and Contriever retrieval baselines.
  • Model editing significantly improves user-level option accuracy across all compared retrieval-based personalization settings.The largest improvement occurs on the Vanilla setting, while retrieval- and persona-enhanced settings also benefit.

F Editing Results with Respect to k and Dataset Size

The appendix reports editing and locality evaluations across clustering and sample-scale settings, including accuracy, efficacy, generalization, fluency, and individual locality.

  • Additional comparisons report user-level ordinal accuracy, macro-averaged MAE, CEM, and option accuracy significance tests against retrieval baselines.
  • Hyperparameter analysis varies clustering K and edited-sample scales of 20, 120, and 200 in terms of accuracy.
  • AlphaEdit is evaluated under varied k and s configurations on LLaMA 3.1 8B Instruct across efficacy, generalization, fluency, and individual locality tasks.
Loading 2608.27816v1…