Aplaud: Adaptive Personalized Low-Rank Decomposition for User-Specific LLM
Xinyu Li, Ruoming Jin, Jianfeng Zhu, Ruixin Guo, Zhi Liu
cs.AI
Quick verdict
Promising shared-subspace LoRA for per-user survey prediction: storage savings are solid, but small accuracy gains over OPPU lack significance testing.
Summary
The paper introduces personalized survey response prediction: given an individual's earlier answers, a fine-tuned LLM must predict their responses to unseen multiple-choice questions. The proposed Aplaud is a two-stage framework. Stage 1 trains a global LoRA adapter on pooled data and decomposes it via SVD into shared subspaces U and V; Stage 2 learns, per user, only a small r×r correction matrix Cu modulating the shared singular values, plus a rank-one residual αuβu. Aplaud+ further factorizes Cu as PuQu. Per-user cost drops to ~110K parameters (2.11% of OPPU's 5.24M) for Aplaud and ~96K (1.84%) for Aplaud+ (Table 3). On four Pew ATP waves, GSS Panel 20, and LAMP Movie Tagging with Llama2-7B and Mistral-7B, Aplaud+ averages ~4.6% higher ACC and 4.5% higher Macro-F1 than OPPU on Llama2 and ~2.5%/10.2% on Mistral; GPT-5 profile and five-neighbor RAG baselines are mostly beaten, though GPT5-RAG ties Aplaud+ on the Llama2 average. V-subspace probing, CKA analysis, noise robustness, and a no-user-overlap check support the design. Main limitations: seed-based significance on only two datasets, possible test-set-tuned hyperparameters, internal inconsistencies in dataset naming and counts, and no measured inference latency.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 5
● low 1
5 strengths · 0 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Per-user adaptation can be compressed to a small correction matrix inside an SVD-derived shared LoRA subspace plus a rank-one residual, cutting per-user parameters to ~2% of OPPU (Table 3: 110,599 vs 5,242,880).
- Aplaud+ outperforms OPPU by ~4.6% ACC / 4.5% Macro-F1 on Llama2-7B and ~2.5% / 10.2% on Mistral-7B averaged over six datasets.
- The learned V-subspace captures task-relevant shared semantics (Table 4: delta > 0 on all five datasets; CKA shows V stable across topics).
- Stage-1 pretraining does not leak evaluation users: excluding overlap leaves performance virtually unchanged (Table 11).
- Aplaud+ establishes new state of the art on the majority of datasets, including against GPT-5 profile and retrieval baselines.
Strengths
Large, verifiable per-user storage reduction
Table 3 computes OPPU at 5,242,880 per-user parameters vs 110,599 (2.11%) for Aplaud and 96,263 (1.84%) for Aplaud+; Table 18 shows Aplaud+ at 0.038% of backbone parameters outperforming OPPU at rank 64 / 2.120% on TS with Qwen2.5-7B (Macro-F1 0.7253 vs 0.6926).
Coherent shared-subspace theory with empirical support
Appendix A frames the method as a Tucker-2-style factorization; Table 4 V-space probing shows Δ > 0 on all five datasets (GSS +0.0138, p=1.14e-11) and Fig 4 CKA shows V stable (≈0.88-0.98) across topics while U varies.
Leakage and robustness checks beyond the minimum
Broad evaluation surface
Honest limitation reporting in places
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper defines personalized survey response prediction: given an individual's answers to a subset of survey questions, predict their multiple-choice responses to held-out questions (§1). Three claimed contributions: (i) the task as a new benchmark built from four Pew ATP waves (G&L, TS, F&R, EI), GSS Panel 20, and LAMP Movie Tagging; (ii) Aplaud, a two-stage parameterization W = W0 + sU(Σ+Cu)V⊤ + αuβu⊤, where U and V are SVD factors of a stage-1 global LoRA update shared by all users, with a per-user r×r correction Cu plus a rank-one residual; (iii) Aplaud+, which factorizes Cu ≈ PuQu. Headline results (§4): +4.6% ACC / +4.5% Macro-F1 vs OPPU on Llama2-7B and +2.5% / +10.2% on Mistral-7B averaged over six datasets (Tables 15–16), with ~2% of OPPU's per-user parameter count.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 0 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the Aplaud method?
Aplaud is a parameter-efficient way to personalize an LLM per user. It first trains one global LoRA adapter on all users' data, decomposes it with SVD into shared directions U and V, then learns for each user only a tiny correction matrix that reweights those shared directions plus a rank-one residual — cutting per-user parameters to about 2% of training a separate LoRA adapter per user (the OPPU approach).
What data was used to evaluate Aplaud?
Four waves of the Pew American Trends Panel (gender & leadership, trust in science, family & relationships, economic inequality), the GSS Panel 20 cohort, and the LAMP Movie Tagging dataset. For each survey the 200 most active respondents were kept, and models predict held-out multiple-choice answers, scored with accuracy and Macro-F1 (plus Wasserstein Distance for response distributions).
Does Aplaud actually beat GPT-5?
It beats a GPT-5 baseline that sees only a textual user profile by large margins, and beats a GPT-5 retrieval baseline using the five most relevant past Q&A pairs on Mistral-7B (+2.4% accuracy on average), but on Llama2-7B the retrieval baseline essentially ties it (+0.3% accuracy, -0.1% Macro-F1 on average).
Is code or data available for Aplaud?
No code release is mentioned in the reviewed text. The surveys are partly public: GSS and LAMP are freely available, while Pew ATP data requires an approved researcher application. The appendix documents prompts, hyperparameters, and per-wave data statistics in detail.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt with [...sampling gap...] regions (parts of §3.3–§4 setup, appendix tables such as 7–10, and portions of the D.13 narratives); judgments about those regions rely on the surrounding visible text only.
- Figures 1–4 could be assessed only through their captions; the underlying plots were not part of the extracted text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.