Less Is Personal: Learning Minimal Sufficient User Profiles for Personalized Language Models
Minghang Liu, Qiang Qiu, Yuanzhuo Wang, Huawei Shen, Xueqi Cheng
cs.AI · cs.CL
Quick verdict
Claims minimal-sufficient user profiles via ENOUGH (adaptive length, Stop); best on all 24 LaMP cells with modest gains; rigorous but no code release stated.
Summary
The paper proposes ENOUGH, a method for retrieval-augmented personalization of frozen LLMs that constructs per-input user profiles of adaptive length. Instead of prepending a fixed number of retrieved history records, a lightweight controller sequentially appends behavioral records or emits Stop, aiming for minimal sufficient profiles: the cheapest ordered profile whose utility approaches the best attainable from the candidate pool. Offline, bounded counterfactual search scores profile prefixes using downstream log-likelihood gain, user specificity measured against content-matched other-user replacements, and token cost; the resulting long-horizon targets are distilled into a multi-head value controller with ranking and stopping losses. At inference, the frozen generator is called once after stopping. Across six LaMP tasks, ENOUGH ranks first in all 24 metric cells with Qwen3.5-9B and Llama-3.1-8B-Instruct generators, beating the strongest baseline PURPLE by 3.51 normalized quality points while using 43.6% fewer profile tokens. Diagnostics show fixed-length profiles peak at k=5 and degrade by k=10, and continuation beyond predicted Stop becomes increasingly harmful. Limitations: gains are modest (ROUGE deltas as low as 0.3 points), no code release is stated, offline labeling costs 336 GPU-hours, and the specificity analysis covers only the 87.5% of records with valid matched controls.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
4 strengths · 5 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
- ENOUGH achieves the best result in all 24 task-metric cells across six LaMP tasks with Qwen3.5-9B and Llama-3.1-8B-Instruct as frozen generators (Table 1).
- An ε-minimal sufficient profile objective can be optimized by distilling bounded-tree counterfactual targets into a multi-head value controller with an explicit Stop action (Eqs. 6, 20-31).
- Dual counterfactual utility separates user-specific marginal value from generic in-context gains, as content-matched replacement drops normalized quality from 100.0 to 99.1 and likelihood gain from 0.041 to 0.024 per token (Table 2).
- ENOUGH beats PURPLE by 3.51 normalized quality points while using 43.6% fewer profile tokens (510 vs 905) and a single generator call at inference (Table 3).
- Continuation beyond predicted Stop is increasingly harmful: the harmful share rises from 16.9% at +1 to 35.9% at +3 appended records (Table 7).
Strengths
Well-posed unification of selection and stopping
Eqs. 3-7 define ε-minimal sufficient profiles over ordered, budget-constrained sequences; Proposition 1 is honestly scoped — optimality holds only given exact Q*, with an explicit disclaimer that the learned controller does not recover the combinatorial optimum (D.1).
Dual counterfactual utility with a harm gate
Eq. 14-17 contrast leave-one-out marginal gains with K−=3 content-matched other-user replacements matched only on pre-action variables, and credit specificity only when the full profile beats the empty profile; balance audit Table 8 (87.5% coverage, mean SMD 0.040) and the Table 2 intervention ladder (content-matched swap: quality 100.0→99.1, gain 0.041→0.024; random swap SMD 0.180) support the construct.
Unusually broad empirical coverage
Strong leakage and tuning discipline, with transparency
Weaknesses
Modest margins against a same-signal baseline
Significance protocol internally inconsistent
Specificity evidence is subset-only
Reproducibility gated by compute, no code statement
Normalization and editorial issues
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper addresses retrieval-augmented personalization of frozen LLMs and asks three coupled questions: whether to personalize at all, which history records to include and in what order, and when additional history stops paying. It defines an ε-minimal sufficient profile (Eq. 6): the cheapest ordered profile whose cost-free utility approaches the best attainable from the retriever's candidate pool, optimized via a cost-regularized surrogate (Eq. 7). ENOUGH is claimed to (i) unify selection and stopping in one sequential decision process, (ii) separate general downstream gain from user-specific value via a dual counterfactual utility, and (iii) deliver one-call inference. Headline empirical claims: first place in all 24 metric cells across six LaMP tasks with Qwen3.5-9B and Llama-3.1-8B-Instruct (Table 1); +3.51 normalized quality over PURPLE with 43.6% fewer profile tokens (510 vs 905) and p50 latency 370 vs 398 ms (Table 3).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 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 ENOUGH in personalized LLMs?
ENOUGH (Efficient Net-utility Optimization for User-profile Growth and Halting) is a method from arXiv 2609.08180 that builds a user profile of adaptive length for retrieval-augmented personalization: a learned controller appends one history record at a time or emits a Stop signal, so a frozen LLM receives only the minimal sufficient context and is called exactly once after stopping.
Does using more user history always improve LLM personalization?
Not according to this paper: fixed-length profiles peak around k=5 records (quality index 98.71) and degrade at k=10 (97.93) while harm relative to no profile rises to 31.7%; forcing one to three extra records beyond the learned Stop raises the harmful share from 16.9% to 35.9%.
How much cheaper is ENOUGH than competing profile-selection methods?
Compared with PURPLE, the strongest baseline, ENOUGH reaches 100.00 vs 96.49 normalized quality while using 510 vs 905 profile tokens (43.6% fewer) and one generator call with p50 latency of 370 ms. However, its offline training required 336 GPU-hours of teacher scoring plus 95 GPU-hours per additional cost weight.
Is code available for ENOUGH?
The paper text does not include a code or data release statement. The benchmark (LaMP, six tasks, time-based split) is public and all six prompt templates are fully specified in the appendix, but the offline labeling pipeline (bounded-tree counterfactual search with matched other-user replacements) would need to be reimplemented.
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.
- No truncation detected: the extraction ends cleanly after the Figure 12 caption and Appendix F.
- Numeric values in Figures 3, 5, 6, 7, 9-11 are embedded as axis/label text, so figure-level visual trends could not be independently verified beyond the quoted numbers.
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.