Source-linked AI summary
Less Is Personal: Learning Minimal Sufficient User Profiles for Personalized Language Models
Minghang Liu, Qiang Qiu, Yuanzhuo Wang, Huawei Shen, Xueqi Cheng
TL;DR
Fixed-length retrieval profiles can include redundant, harmful, or non-personal history, motivating profiles that preserve personalization utility at lower context cost. ENOUGH constructs ordered profiles with adaptive stopping using counterfactual utility, user specificity, and token cost, then invokes the frozen generator once. Across six tasks and two backbones, it ranks first in all 24 task–metric settings while using 43.6% fewer profile tokens than the strongest baseline.
Problem
Fixed-size retrieved profiles do not determine how much history is useful or which relevant records are genuinely specific to the target user.
Method
ENOUGH uses bounded counterfactual search and long-horizon supervision to select and order behavioral records or Stop, accounting for downstream gain, user specificity, and token cost.
Results
ENOUGH ranks first in all 24 task–metric settings and uses 43.6% fewer profile tokens than the strongest baseline across six tasks and two generator backbones.
Takeaways & Limitations
Effective personalization depends on identifying when selected user evidence is sufficient rather than exposing models to more history.
Abstract
from arXiv · showhide
Retrieval-augmented personalization enables large language models to produce more accurate and preference-aligned outputs using relevant records retrieved from user histories. Personalized language models typically prepend a fixed number of retrieved user records, even when additional history is redundant, harmful, or unrelated to a user's distinctive behavior. We study minimal sufficient personalization: constructing the least costly ordered profile for each input while preserving the utility achievable from a retrieved candidate pool. We introduce ENOUGH, a method that iteratively appends behavioral records or emits STOP to construct profiles with adaptive lengths. Offline, bounded counterfactual search evaluates profile prefixes by jointly considering downstream gains, user specificity, and token costs. The resulting long-horizon targets are distilled into a multi-head value controller with explicit ranking and stopping supervision. At inference, the controller selects and orders records through lightweight decisions, and the frozen generator is invoked once after stopping. Extensive experiments on six personalized tasks demonstrate that ENOUGH consistently outperforms strong heuristic and retrieval-augmented baselines in both effectiveness and efficiency, achieving minimal sufficient profiles that preserve personalization utility while reducing unnecessary context costs.
1 Introduction
Retrieval-augmented personalization must determine both how much history to use and which relevant records genuinely reveal the target user. ENOUGH formulates this as minimal sufficient profile construction, selecting ordered records or stopping adaptively while accounting for utility, specificity, and cost.
- Motivation: Fixed-size retrieval profiles cannot determine when additional history is redundant, conflicting, or unnecessary for a request.Requests may require no profile, one behavior, or several selected records.
- Motivation: Semantic relevance and general generation utility do not by themselves establish that a record is specific to the target user.The paper compares target-user records with content-matched other-user replacements to measure user-specific marginal value.
- Objective: Minimal sufficient personalization seeks the least costly ordered profile that approaches the best attainable personalized utility within a retrieved candidate pool.The objective jointly makes profile length emerge from conditional record value rather than treating it as an independent hyperparameter.
- Method: ENOUGH constructs profiles from an empty state by appending behavioral records or emitting Stop, allowing it to reject augmentation or terminate at any feasible length.Offline bounded counterfactual search evaluates downstream gain, user specificity, and incremental token cost, with long-horizon backups supervising a multi-head value controller.
- Evaluation: The evaluation spans six personalized tasks across classification, regression, and text generation, with diagnostics for stopping, specificity, robustness, and efficiency.These experiments assess ENOUGH across varied personalization settings rather than a single task type.
2 Methodology
ENOUGH constructs ordered user profiles adaptively by adding behavioral records or stopping, optimizing downstream utility, user specificity, and prompt cost. Offline bounded search supplies long-horizon supervision for a multi-head controller that selects records and Stop under context constraints.
- Profile construction: ENOUGH starts from an empty profile and chooses feasible non-repeating records or Stop within candidate-pool and context limits.A frozen retriever forms the candidate pool, while selecting a record appends it to the ordered profile and Stop terminates construction.
- Profile objective: The objective combines downstream gain, matched-replacement user specificity, and incremental prompt cost into cost-regularized personalized utility.Task gain compares reference likelihood with and without a profile; specificity credits target-user records over content-matched records from other users.
- Controller learning: The framework represents the request, selected profile, remaining candidates, and budget state while using a frozen generator after profile construction.The profile representation is order-sensitive, the remaining-candidate summary is permutation-invariant, and the generator is invoked through the prompt template.
- Profile objective: ENOUGH prevents harmful profiles from receiving specificity credit when they underperform no personalization, even if matched controls perform worse.This safeguard conditions the specificity bonus on improvement over the empty profile.
- Offline supervision: Offline counterfactual search builds a bounded prefix tree with stratified profile depths, expanded feasible actions, and retained prefixes for exact backups within the sampled tree.Search combines retriever-score, diversity, and uniformly sampled actions while representing unique ordered prefixes once and batching generator evaluations.
- Controller learning: A multi-head controller predicts normalized task gain, matched-replacement specificity, and prompt cost for each action, then combines them into a net score for ranking and stopping.The learned policy distills sampled long-horizon continuations and applies the composite value to record additions and Stop.
3 Experiments
Across six personalized tasks, ENOUGH consistently improves effectiveness while reducing profile and inference costs. Ablations and diagnostics show that long-horizon supervision, adaptive stopping, user specificity, and token cost jointly support minimal sufficient profiles.
- 3.2 Main Results: ENOUGH ranks first in all 24 reported metric cells across six tasks and two frozen generator backbones.With Qwen3.5-9B, gains over the strongest baseline span classification, ordinal rating, and all three generation tasks; the same pattern transfers to Llama-3.1-8B-Instruct.
- 3.3 Ablation Analysis: Replacing long-horizon supervision with myopic targets degrades task quality, lengthens profiles, and increases harm.The result indicates that locally attractive records do not reliably form a useful final profile.
- 3.3 Ablation Analysis: Disabling adaptive stopping produces the same quality, length, and harm pattern, showing that fixed-length selection cannot avoid unnecessary evidence.The ablation supports reasoning about future selections when additional personalization is no longer beneficial.
- 3.3 Ablation Analysis: Removing user specificity reduces task quality, whereas removing token cost slightly improves aggregate quality but substantially increases token usage and suppresses early stopping.The token-cost term therefore improves the quality–cost–harm trade-off by preventing marginal gains from requiring disproportionately long profiles.
- 3.4 Is More Personalization Always Beneficial?: Fixed-profile quality peaks at k = 5 (98.7) but falls by 0.8 points at k = 10, while harm rises from 23.2% to 31.7%.These measurements motivate instance-adaptive rather than fixed profile lengths.
- 3.4 Is More Personalization Always Beneficial?: After predicted Stop, the beneficial share falls from 18.0% at +1 to 9.9% at +3, while the harmful share rises from 16.9% to 35.9%.Deeper continuation becomes less likely to improve the objective and more likely to hurt performance.
- 3.6 Hyperparameter Sensitivity: With λ = 0.10, mean profile length falls from 741 to 512 tokens, a 30.9% reduction, while the quality index changes from 100.2 to 99.9.The sweep shows that token-cost weighting can substantially reduce context length with a small quality change.
- 3.7 Retriever Independence: ENOUGH outperforms the strongest baseline in all 30 retriever–task combinations, with Accuracy and ROUGE-1 gains ranging from +0.2 to +1.8 points.The largest margin varies across retrievers and tasks, suggesting the benefit comes from selection and stopping rather than a particular retriever.
4 Conclusion
The paper situates ENOUGH within retrieval-augmented personalization and concludes that adaptive profile construction can jointly optimize record selection, stopping, utility, specificity, and token cost. Its inference procedure uses lightweight controller decisions followed by one frozen-generator call.
- Objective: The framework distinguishes user-specific benefit from generic task improvement by combining gain over an empty profile with matched-replacement comparisons.This comparison tests whether selected evidence helps because of the target user rather than merely serving as a generally useful example.
- Method: ENOUGH extends retrieval-augmented personalization by selecting and ordering behavioral records while allowing an empty profile or termination at any feasible length.The method frames profile construction as a sequential decision process rather than fixed-size retrieval.
- Offline supervision: The offline procedure constructs bounded prefix trees, scores terminal profiles and matched replacements, and propagates labels for controller supervision.The procedure uses sampled roots, search budgets, matched replacements, and backups of three objective components.
- Inference: At inference, ENOUGH scores remaining actions in batches for at most Kmax steps and calls the autoregressive generator exactly once after profile termination.Reference outputs, matched replacements, likelihoods, counterfactual rollouts, and generator calls are excluded from selection and stopping decisions.
D.1 Proposition 1: Exact-value stopping
The proposition establishes exact-value stopping on a finite deterministic candidate tree: with exact optimal action values, repeatedly maximizing over Stop and feasible record actions reaches an optimal terminal profile. The guarantee applies to exact values, not necessarily to the learned controller.
- Finite-tree setting: The candidate subtree is finite because records cannot repeat and profile length is bounded, so every reachable terminal profile can be evaluated.The proof uses induction on the maximum remaining record-action depth.
- Bellman recursion: The Bellman characterization compares stopping immediately with the best continuation value among feasible record actions.Stop represents the current prefix as a terminal profile, while record actions descend to child prefixes.
- Stopping rule: Choosing Stop on a tie preserves optimality because Stop is itself a maximizing action.Applying this argument at the root proves the proposition.
- Scope: The guarantee is conditional on exact Q⋆ values and does not imply that the learned controller Qθ recovers the complete combinatorial-tree optimum.The bounded-tree construction provides an exact target only within the sampled tree and approximates the full optimum.
E.2 Implementation Details
The implementation fixes retrieval, generation, serialization, leakage controls, evaluation, and cost accounting to isolate profile-construction effects. ENOUGH and baselines are compared under matched pools and audited computational costs.
- Models and retrieval: The main setup uses frozen Contriever retrieval with at most M = 20 records, frozen Qwen3.5-9B generation, and deterministic decoding.Llama-3.1-8B-Instruct is used for generator transfer, with tokenizer and likelihood computations restricted to reference-response tokens.
- Supervision: Offline policy-induced aggregation adds controller-visited states, relabels them with the frozen generator, and retrains on the union of state pools.No controller update occurs during test-time profile construction.
- Serialization and leakage: All selectors use generator-specific tokenization and explicit record boundaries, while temporal truncation and duplicate removal prevent future or repeated interactions from entering retrieval or matching.The current input is serialized separately and records are not silently truncated.
- Cost accounting: Profile cost is measured as actual incremental tokenizer tokens, alongside root Stop rate, harm rate, generator calls, latency, and throughput.Reranker and generator-equivalent forward passes are included in the cost audit.
- Statistical protocol: Each automatic result uses three complete pipeline seeds with user-clustered paired bootstrap intervals and Holm-corrected comparisons against main baselines.Development data determines operating settings before test evaluation.
E.3 Generator-Scale Sensitivity
Generator-scale sensitivity is evaluated across Qwen3.5-4B, 9B, and 27B while holding the retrieval, search, prompting, decoding, and leakage controls fixed. Larger generators generally improve quality, but the magnitude varies by method, task, and metric.
- Experimental design: The experiment independently recomputes labels, caches, likelihoods, and controllers for Qwen3.5-4B, 9B, and 27B without transferring generator-dependent artifacts.The retriever pool, search budget, task prompt, deterministic decoding, and leakage controls remain fixed.
- Evaluation: Figure 7 reports both official metrics for every task and all nine main methods, with each cell aggregating three complete pipeline seeds.The 9B condition reproduces the corresponding main-table results, while the other scales are independently measured.
- Results: From 4B to 27B, baseline changes range from 1.2–2.5 points for classification accuracy, 1.2–2.6 for F1, 1.3–2.7 for MAE reduction, 2.5–5.0 for RMSE reduction, and 0.4–2.0 for ROUGE.These ranges show that scale effects differ across metrics rather than following a uniform offset.
E.4 Profile Length and Stopping Diagnostics
The diagnostics show that fixed profile lengths trade quality against rising cost and harm, motivating adaptive stopping. ENOUGH’s stop signal is informative but imperfect and does not establish globally minimal profiles.
- Fixed profiles: Quality rises from 96.15 with no profile to 98.71 at k = 5, then falls to 97.93 at k = 10.Across the same grid, mean profile cost reaches 1,755 tokens and harm rises to 31.7%.
- Fixed profiles: Fixed profile length produces non-monotonic quality while cost and instance-level risk increase with length.The diagnostic motivates instance-adaptive length selection rather than identifying a universally optimal k.
- Forced continuation: 18.0% of controller-ordered +1 continuations are beneficial, declining to 9.9% at +3 as harmful continuations rise from 16.9% to 35.9%.Cumulative added cost increases from 177 to 521 tokens; retriever-ordered continuation is more harmful at every reported depth.
- Stopping quality: Stop confidence aligns with bounded-tree teacher decisions, with Brier/ECE of 0.116/0.028 and AUROC/AUPRC of 0.839/0.794.The calibration target is the bounded-tree teacher rather than exact global stopping optimality.
- Stopping quality: Stopping errors are 8.8% premature, 10.7% late, and 3.0% forced, with mean utility loss of 0.076.These rates are descriptive audit events, and the diagnostics do not establish globally minimal profiles.
E.5 Specificity, Matching, and Human Evaluation
Specificity evaluation combines matched-control audits, frozen-generator interventions, and blinded human judgments to distinguish target-user information from generic contextual utility. Matching and replacement analyses characterize conditional comparisons rather than causal effects of user identity.
- Matching: Matched controls require three other-user records satisfying a caliper, so replacement comparisons describe the matched subset rather than all selected profiles.Coverage measures feasible matching, while exact-match rates and standardized mean differences assess balance.
- Specificity interventions: Content-matched, random-user, whole-profile, action-shuffling, and ownership interventions progressively test whether personalization depends on target-user records and their actions.The interventions preserve or disrupt content, ownership, and context-action pairing in distinct ways.
- Interpretation: The replacement comparisons estimate frozen-generator effects conditional on achieved covariate balance, not a causal effect of user identity.A negative-control audit tests whether the matching pipeline itself can manufacture an apparent specificity effect.
- Human evaluation: Blind comparisons with three annotators on 200 examples per task separately judge faithfulness, style consistency, and overall preference.Presentation order is randomized and method identity is concealed, preventing task correctness from being conflated with stylistic personalization.
E.6 Additional Ablations
Ablations and exact-oracle diagnostics evaluate ENOUGH’s controller design and its approximation to minimal sufficient profiles. The bounded-tree teacher remains close to the exact oracle, while ENOUGH consistently outperforms myopic and fixed-k selectors on the reported diagnostics.
- Ablations: The ablation figure isolates long-horizon targets, adaptive stopping, user specificity, token cost, value-head decomposition, ranking and STOP losses, and order-sensitive encoding.These ablations target the principal components of the controller objective and representation.
- Exact-oracle diagnostic: The exact-oracle diagnostic enumerates 517 legal ordered profiles per instance using M′ = 6 candidates and K′ = 4 maximum length.Because it scores profiles with the gold reference, the exact oracle is a diagnostic upper bound rather than a deployable selector.
- Exact-oracle diagnostic: The bounded-tree teacher achieves 0.028 utility regret, 91.6% ϵ-sufficiency, 41 excess tokens, and 92.3% oracle-stop agreement.These results quantify the gap introduced by bounded search relative to the exact oracle.
- ENOUGH controller: ENOUGH reaches 0.069 regret, 81.8% ϵ-sufficiency, 89 excess tokens, and 83.7% oracle-stop agreement.The additional gap from the teacher to ENOUGH reflects controller approximation.
- Task-level results: ENOUGH regret ranges from 0.064 on LaMP-1 to 0.082 on LaMP-3, below myopic and fixed-4 alternatives on every task.Pooled quantities are computed directly over diagnostic instances rather than by averaging heterogeneous task-level metrics.
E.8 Additional Hyperparameter Sensitivity
One-at-a-time development sweeps place the selected structural, search, and loss settings in favorable regions rather than at sharply isolated optima. The reported trends are descriptive within the evaluated grids and do not establish statistical significance.
- Structural and search parameters: M = 20 achieves quality 99.4 and the lowest stop regret, 0.069, among evaluated candidate-pool sizes.Larger pools produce small degradations in both metrics.
- Structural and search parameters: Increasing Kmax from 2 to 10 substantially improves quality and stop regret, while Kmax = 15 preserves quality 99.4 and lowers regret by only 0.001.The selected structural settings lie near favorable regions of their sweeps.
- Training-loss weights: The selected loss weights η = 0.2 and ζ = 0.4 each reach quality 99.9, outperforming their omitted and unit-weight alternatives.The neighboring settings remain competitive, indicating a favorable region rather than a sharply isolated optimum.
- Interpretation: The sweeps support balancing ranking and stopping objectives with value regression, but Figure 10 reports only aggregate quality.They do not establish effects on stopping calibration, profile length, or task-level variability.
E.9 Robustness and Subgroups
The evaluation probes ENOUGH under controlled perturbations and across user and input subgroups, while separately measuring training and online inference costs. ENOUGH achieves high quality with comparatively low task-specific training time and remains in the low-latency group without collapsing distinct cost dimensions into one score.
- Robustness: Near duplicates change selected length by 0.07 records, while action conflicts increase harm from 12.0% to 19.0% and reduce quality by 0.7 index points.The robustness ladder inserts near duplicates, random records, and behaviorally inconsistent action conflicts into an otherwise controlled candidate pool.
- Subgroups: Subgroup analysis stratifies users and inputs by history length, matched-candidate coverage, and input difficulty using development-set quantiles frozen before test evaluation.The subgroup buckets are defined before test evaluation rather than tuned on the test set.
- Measurement protocol: Online latency reports tokens, generator-equivalent calls, and p50/p95 end-to-end timing, with median decomposition into retrieval, selection, generation, and residual preprocessing.Offline teacher scoring, cache construction, and additional λ operating points are excluded from online latency and reported separately.
- Efficiency: ENOUGH remains in the low-latency group for classification, scoring, and text-generation tasks, although BM25/Contriever have slightly lower ranges and some baselines reach higher upper bounds.The latency comparison uses batch-size-one measurements under common hardware and generator settings.
- Prompt protocol: The frozen generator receives shared task delimiters and official LaMP instructions, with ENOUGH-selected records serialized in selected order; LaMP-7 histories omit separate pre-action context.The prompt templates cover six personalized LaMP tasks and distinguish shared instructions from task-specific fields.