Source-linked AI summary
PLUME: Parameter-Efficient Personalization of Large Language Models via Low-Rank User Modulation in Shared Subspaces
Xinyu Li, Hao Zhou, Jianfeng Zhu, Julina Maharjan, Ruixin Guo, Feodor Dragan, Ruoming Jin
TL;DR
Personalized LLMs need user-level adaptation without the storage and overfitting costs of separate full adapters. PLUME learns a shared task subspace and modulates it with compact user-specific components, matching or exceeding personalized LoRA variants while reducing per-user parameters by over 95%.
Problem
Per-user personalization improves user-level adaptation but creates storage overhead at scale and risks overfitting when each user has limited or variable-quality data.
Method
PLUME learns global task-specific LoRA representations, modulates their shared subspace with USM, shares personalized parameters across layers through PCLS, and adds rank-1 residuals.
Results
PLUME consistently matches or exceeds existing personalized LoRA variants across five personalized text-generation benchmarks while reducing per-user parameters by over 95%.
Takeaways & Limitations
PLUME provides a compact user representation for personalized LLMs while retaining personalization performance and expressive capacity.
Takeaways & Limitations
OPPU-style per-user adaptation is constrained by storage and memory costs at large user counts, and independent full-LoRA optimization can overfit or become unstable.
Abstract
from arXiv · showhide
Personalizing large language models (LLMs) is essential for delivering AI assistance that aligns with individual users' styles, intents, and preferences. While per-user fine-tuning can substantially enhance personalization quality, it introduces significant parameter and storage overhead, limiting scalability to large user populations. We propose PLUME (Personalized Low-Rank Adaptation through User Modulation and Shared Subspace), a lightweight framework that achieves efficient and expressive per-user adaptation by leveraging a shared task-specific subspace. Specifically, PLUME first learns a global task subspace from aggregated user data. Personalization is then achieved by training only a lightweight small square matrix within this subspace, enabling each user to obtain a tailored model while keeping shared components fixed. Cross-layer shared parameters and rank-1 residual terms are further introduced to significantly reduce redundancy while maintaining expressiveness. Experiments on multiple personalized text generation benchmarks demonstrate that PLUME achieves comparable or superior performance to strong baselines, while reducing per-user parameters by over 95%. These results establish shared-subspace modulation with minimal residuals as a scalable and semantically grounded approach to LLM personalization.
1 Introduction
Personalized LLMs aim to tailor outputs to individual users, but existing prompt- and fine-tuning-based approaches face signal, context, personalization, storage, and overfitting challenges. PLUME addresses this by combining a shared task subspace with lightweight user-specific modules.
- Personalized LLMs tailor outputs to users’ styles, preferences, or contexts, unlike one-size-fits-all models.
- Prompt-based personalization is simple and interpretable but depends heavily on explicit, high-quality user signals and limited context length.
- Universal fine-tuning can blur distinctions between users and underperform on fine-grained personalization.
- Per-user models improve personalization but create storage overhead at scale and risk overfitting because individual writing data is limited.
- PLUME learns a shared task-specific subspace, then uses user-conditioned modulation, cross-layer sharing, and rank-1 residuals for efficient personalization.
- PLUME frames the central problem as compressing per-user PEFT parameters while preserving personalization effectiveness and model expressiveness.
- PLUME reduces per-user parameters to less than 5% of standard LoRA while preserving personalization performance and expressive capacity.
2 Problem and Preliminaries
The paper formulates personalized text generation around user-specific adapters, retrieved histories, and a shared backbone, then contrasts this setup with per-user PEFT. Vanilla OPPU stores and optimizes a separate LoRA module for every user, creating scalability and stability problems.
- Personalized text generation seeks a unique adapter for each user, conditioned on the current query and contextual information.
- Each user is represented by a history of query–response pairs that can support personalized context construction.
- At inference, top-k similar historical query–response pairs are retrieved for the user’s test query before generating a response.
- The model combines a shared backbone with a lightweight user adapter, while training predicts response tokens from prior tokens, the query, and retrieved context.
- Auxiliary information from other users can be encoded into global task-specific LoRA representations shared by all users.
- OPPU assigns an independent PEFT module to each user and constructs personalized models by combining user-specific and globally trained adaptations.
- OPPU’s separate LoRA modules impose high storage and memory costs at scale, while full per-user optimization can cause overfitting or unstable personalization.
3 Method
PLUME reduces personalized adaptation redundancy by modulating a shared task subspace, sharing personalization across layers, and adding lightweight residuals for local flexibility.
- PLUME frames personalization as compressing per-user PEFT parameters while preserving personalization effectiveness and model expressiveness.
- User-Conditioned Subspace Mixing reweights shared task-subspace directions, allowing users to exploit common task knowledge differently.
- When individual residual rank is approximately 4, PLUME matches OPPU with rank 64; larger ranks add no gains and can cause redundancy and overfitting.
- OPPU adapters show inter-layer CKA values often exceeding 0.9, indicating redundant user-specific directions across depth and motivating shared subspaces.
- PLUME separates cross-layer personalization into a shared PCLS subspace and rank-1 layer-wise residuals for fine-grained local adaptation.
- Per-user complexity is O(Lr_g^2 + r_shd + Ld), compared with O(Ld r_OPPU) for OPPU’s independent adapters.
4 Experiments
Experiments evaluate PLUME across five personalized generation tasks, testing performance, efficiency, component contributions, and rank sensitivity. PLUME maintains or improves quality while substantially reducing per-user parameters.
- Experiments cover five short- and long-form personalized generation tasks from the LaMP and LongLaMP benchmarks.
- PLUME is compared with non-personalized and personalized PEFT baselines under consistent LoRA settings on LLaMA2-7B and Mistral-7B.
- Personalized methods outperform non-personalized methods across metrics, while PLUME matches or exceeds strong personalized baselines with fewer parameters.
- Compared with OPPU’s 168M parameters per user, PLUME and PLUME-s use 6.8% and 3.1%, respectively, achieving over 15×–30× compression.
- Removing USM, PCLS, or Resid consistently lowers ROUGE-L and METEOR, with module impact ordered Resid > PCLS > USM.
- Residual-rank sensitivity peaks around rank 4 for Abstract Generation, whereas Scholarly Title performance is already strong at rank 1 and declines slightly thereafter.
- Increasing PCLS rank improves Abstract Generation but harms Scholarly Title after a small optimum, indicating task-dependent shared-subspace capacity.
5 Conclusion
PLUME enables fine-grained user adaptation through low-rank user modulation and shared subspaces, matching or exceeding personalized LoRA variants while reducing per-user parameters by over 95%.
- PLUME combines a User-Conditioned Subspace Mixer, cross-layer shared personalized subspace, and rank-1 residuals for compact personalization.These components support rich personalization with only a fraction of the parameters required by conventional per-user adapters such as OPPU.
- Over 95% lower per-user parameters accompany performance that consistently matches or exceeds existing personalized LoRA variants.
- PLUME provides a more effective and highly compact user representation for personalized LLMs.
- LoRA reduces trainable parameters from O(dk) to O(r(d + k)) by replacing full weight updates with a low-rank update.
A.2 LLM personalization
LLM personalization methods primarily use prompts or adapters, but explicit-signal dependence, context limits, parameter growth, and overfitting constrain their scalability and fine-grained user adaptation.
- Personalization strategies are broadly divided into prompt-based and adapter-based approaches.
- Prompt-Based Approaches: Prompt-based methods encode user history or behavior into prompts without modifying LLM parameters.They use handcrafted or learned prompts, including summarized or sampled user histories, to guide model behavior.
- Prompt-Based Approaches: Prompt-based personalization relies on strong explicit user signals and often struggles with implicit stylistic preferences or personality traits.These limitations can produce unstable and degraded performance in content generation tasks.
- Adapter-Based Approaches: Adapter-based methods either train a shared model across users or assign each user a dedicated PEFT-trained model.OPPU assigns each user an individual LoRA module, while other methods explore shared models and mixture-of-experts structures.
- Adapter-Based Approaches: Per-user PEFT modules create storage overhead that grows with the user population and can overfit limited user data.These modules may also contain redundant parameters, motivating reductions in user-specific parameterization.
- Linear CKA measures scale- and rotation-invariant similarity between learned LoRA representations across layers.The analysis uses column-centered, vectorized LoRA matrices from different layers of the same user model.
B Data Processing Detail and Data Statistics
The evaluation uses temporally partitioned user interactions and compares generation configurations on LLaMA-2-7B across five personalized tasks.
- Each user’s historical interactions are split into training, validation, and test subsets using an 8:1:1 temporal ratio.Prompt-based retrieval uses only the training portion to avoid test leakage.
- Table 5 compares performance on LLaMA-2-7B across five personalized text-generation tasks.
- Table 6 reports average generated tokens across ground-truth length bins for each configuration.Percentages for PLUME and PLUME-s indicate changes relative to OPPU.
C More Experiment
The paper evaluates personalized generation with complementary metrics covering lexical overlap, structural similarity, semantic matching, and word-order penalties.
- ROUGE-1: ROUGE-1 measures unigram overlap between generated and reference texts as a proxy for lexical content coverage.The reported F1-score balances unigram precision and recall.
- ROUGE-1: ROUGE-1 overlap is defined from unigram counts in the generated text G and reference text R.
- ROUGE-L: ROUGE-L measures sentence-level structural similarity using the Longest Common Subsequence between generated and reference texts.Its F-measure combines recall and precision, with β typically favoring recall.
- METEOR: METEOR extends lexical overlap with exact, stem, and synonym matches plus a fragmentation penalty for word ordering.Its penalty uses matched chunks and hyperparameters γ and θ.
- Together, ROUGE-1, ROUGE-L, and METEOR provide complementary perspectives on personalized text-generation quality.
C.2 Training Details and Reproducibility
The experiments use fixed training and generation settings while tuning learning rates, component coefficients, and LoRA ranks through predefined search spaces.
- All experiments run for 5 epochs with AdamW and generation temperature set to 0.
- Learning rates are selected from six candidates ranging from 3e−4 to 1e−6.
- Component coefficients s and s′ are selected from eight values spanning 0.01 to 50.0.
- Sensitivity studies vary LoRA rank across values from 1 through 64.
C.3 Llama2-7B Results
With LLaMA-2-7B-Chat, both PLUME variants outperform prior personalized baselines, with PLUME-s providing the strongest balance between performance and parameter efficiency.
- Both PLUME variants outperform prior personalized baselines on the LLaMA-2-7B-Chat backbone.
- PLUME achieves slightly higher scores on several tasks.
- PLUME-s attains the best overall balance between performance and parameter efficiency.
- Results across LLaMA-2-7B and Mistral-7B demonstrate robustness across different backbone models.
D More Analysis
Additional analyses examine reference-example counts, output-length bins, and model variants, revealing task-dependent effects and an optimal length range for abstract generation.
- For Abstract Generation, adding reference examples produces small gains before performance plateaus around 3–4 examples.
- For Scholarly Title Generation, quality peaks with 1–2 examples and then declines as more examples are added.
- Figure 5 analyzes performance changes as the number of retrieved interaction-history examples varies.
- Figure 6 compares Abstract Generation across five ground-truth length bins for OPPU, PLUME, and PLUME-s.
- Both PLUME variants consistently outperform OPPU across most ground-truth length bins.
- ROUGE-1 and METEOR peak in Bin 2, covering ground-truth lengths of 126–140 tokens, then decline as length increases.
E Case Study
The case studies evaluate PLUME outputs for abstract and scholarly-title generation using style-conditioned prompts and comparisons against reference and baseline outputs. The examples emphasize writing-style fidelity, relevance, and conceptual alignment in generated text.
- Case Study Setup: The case studies present representative examples from Abstract Generation and Scholarly Title Generation.
- Abstract Generation: Abstract-generation prompts request new academic abstracts based on reference examples for style and tone, while prohibiting sentence copying.
- Abstract Generation: The abstract-generation constraint specifies 150–250 words of formal, concise, self-contained text without headings or commentary.
- Abstract Generation: The NFC reference describes a coopetitive architecture supporting cooperative and competitive business models in dynamic, scalable mobile services.
- Abstract Generation: The NFC case compares Gold, non-personalized LoRA, OPPU, and PLUME-s outputs, with highlights marking generated relevance and relevant ground truth.
- Scholarly Title Generation: The scholarly-title case prompts a title matching the author's writing style based on an abstract and reference title examples.
- Scholarly Title Generation: The face-recognition example combines PCA-based feature extraction with a feed-forward back-propagation neural network and evaluates 400 images across 40 classes.
- Qualitative Findings: The PLUME-s NFC output is described as preserving academic style while improving fidelity to the target writing style and conceptual alignment with the gold abstract.