PaperlayerView exact revision on arXiv

Source-linked AI summary

Does Every User Need a Private LoRA? Decoupling Personalization from Per-User Adaptation

Songyuan Sui, Srikanth Malla, Chiho Choi, Joon Hee Choi

arXiv:2610.02353v1cs.LGcs.CL

TL;DR

Personalized LLMs must represent differences in user preferences without requiring a complete adaptation state for every user. The paper proposes LINEUP, which composes reusable factors using user history and queries while limiting user-specific fitting to a compact code; it reports leading results across six tasks and all 12 metrics.

  • Problem

    The paper asks how much personalization capacity can be shared across users, how it should be composed, and how much must remain user-specific.

    Source: §1 · p. 2

  • Method

    LINEUP learns reusable factors across source users, composes them through history-conditioned recall and query-dependent calibration, and refines only an eight-dimensional target-user code while shared components remain fixed.

    Source: §1 · p. 2

  • Results

    LINEUP leads on all 12 reported metrics across six tasks spanning personalized classification, prediction, and generation, with scores averaged over three independent runs.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The results show that rich personalization need not require a full private adapter for every user, with independent adaptation reserved for compact individual correction.

    Source: §6 · p. 10

    User histories, codes, and shared components still require privacy safeguards, and history-based personalization may reflect biases or manipulation in user data.

    Source: ETHICS STATEMENT · p. 10

Abstract

from arXiv · show

Personalized large language models often require a complete adaptation state for each user. However, this paradigm scales poorly as the user population grows. We revisit this design through the lens of personalization capacity allocation: how much adaptation capacity can be shared across users, how the shared capacity should be composed, and how much must remain user-specific. We answer them through three complementary empirical analyses. We find that independent user adapters contain substantial cross-user reusable structure, that the utility of reusable directions reflects both user relevance and variation across queries, and that user histories provide transferable signals for compact individual correction. Motivated by these findings, we propose LINEUP. It learns a bank of reusable low-rank personalization factors, composes them through user-conditioned recall and query-dependent calibration, and restricts target-user adaptation to a tiny user code over a shared correction space. This design decouples expressive personalization capacity from per-user trainable state. Each target user optimizes only eight scalars, while all shared components remain fixed. By comparison, the evaluated private-LoRA configuration uses 4.19 million per-user parameters. Our theoretical analysis gives a finite-step, finite-history risk bound and sufficient conditions for user-code refinement to improve on history initialization. Across six tasks spanning personalized classification, prediction, and generation, LINEUP leads on all 12 metrics, each averaged over three independent runs (e.g., reducing LaMP-3 RMSE by 11.4% relative to the strongest baseline). It maintains advantages under limited history. These results show that rich personalization can be supported primarily by reusable, conditionally composed shared capacity, while independent user adaptation remains confined to a tiny correction state.

1 INTRODUCTION

The paper asks how personalization capacity can be shared, composed for each request, and limited to compact user-specific adaptation. Empirical findings motivate LINEUP, which combines reusable shared factors with a tiny user code.

1 INTRODUCTION

  • The central question is how much of a personalized adapter is truly personal, given that pooled adaptation captures common behavior but not differing user preferences.The paper frames this as a problem of allocating capacity between cross-user reuse and target-user adaptation.Source: §1 · pp. 1–2
  • Independent adapters contain reusable structure, direction utility reflects both user association and query variation, and user histories provide transferable individual-correction signals.These three analyses motivate shared capacity, conditional composition, and compact individual correction.Source: §1 · p. 2
  • LINEUP learns reusable low-rank factors across source users, recalls and calibrates them using history and the current query, and confines target-user fitting to an eight-dimensional code.Shared components remain fixed during code refinement; the evaluated OPPU configuration uses 4.19 million per-user parameters.Source: §1 · p. 2
  • LINEUP leads on all 12 reported metrics across six tasks spanning personalized classification, prediction, and generation, with scores averaged over three runs.It also maintains advantages with limited user history.Source: §1 · p. 2
  • The paper derives a finite-step, finite-history risk bound and sufficient conditions for regularized user-code refinement to improve on history initialization.The bound accounts for shared-space approximation and history initialization.Source: §1 · p. 2

2 PROBLEM SETUP AND PRELIMINARIES

The setup separates population-shared personalization capacity from target-user state and conditions prediction on user history and the current query. The preliminaries describe parameter-space LoRA updates and representation-space correction directions.

2.1 PERSONALIZATION PROBLEM SETUP

  • Each task is studied independently; shared components are learned from source users, while unseen target users are personalized using their available history and query inputs, never the target output.History records are input–output pairs, and support construction need not use every history record.Source: §2.1 · pp. 2–3

2.2 ADAPTATION PRELIMINARIES

  • A user-specific LoRA update is a scaled low-rank update to a frozen linear transformation, fitted independently per user and reused across that user’s queries.The parameter-space analysis concerns effective updates rather than a particular factorization.Source: §2.2 · p. 3
  • Representation-space correction adds a positional-scope-weighted vector to a hidden state, with the loss gradient at zero intervention defining a first-order loss-reducing direction.This unrestricted direction serves as a diagnostic before imposing a low-dimensional user parameterization.Source: §2.2 · p. 3

3 EMPIRICAL ANALYSIS OF PERSONALIZED ADAPTATION

Three analyses examine reuse across adapters, variation in reusable-direction utility, and transfer from user histories. Their findings motivate learning shared capacity, composing it conditionally, and reserving compact individual correction.

3 EMPIRICAL ANALYSIS OF PERSONALIZED ADAPTATION

  • The analyses examine adapter reuse and direction utility on a fixed backbone, then test representation-space correction without user adapters; reported intervals are 95% user-level bootstrap confidence intervals unless stated otherwise.The design choices informed by these analyses are evaluated through later ablations.Source: §3 · pp. 3–4

3.1 RQ1: HOW MUCH STRUCTURE IS REUSABLE ACROSS USER ADAPTERS?

  • 32 directions, or 0.8% of output dimension, capture 54.1%–67.9% of held-out update energy, versus 0.78% for random subspaces.Genuine-user bases also cover 14.3–21.0 percentage points more than size-matched random-partition bases on the same updates.Source: §3.1 · p. 4

3.2 RQ2: HOW DOES DIRECTION UTILITY VARY ACROSS USERS AND QUERIES?

  • Within-user versus cross-user Spearman agreement is 0.145 versus 0.008 on LaMP-2M and 0.400 versus 0.059 on LaMP-2N, while query differences explain 82.7% and 61.3% of response variation.The within-user advantage holds across analyzed layers and projections; the query-variation figures control for overall sensitivity scale.Source: §3.2 · p. 4

3.3 RQ3: CAN USER HISTORY GUIDE COMPACT INDIVIDUAL CORRECTION?

  • User histories provide positive transferable correction, with intervals excluding zero and means above both random and label-matched donor histories.An eight-dimensional space retains 34.7% and 29.4% of full-space own-history transfer on LaMP-2M and LaMP-2N, respectively, while capturing 12.9% and 17.0% of residual query-correction energy.Source: §3.3 · p. 4

4 METHODOLOGY: LINEUP

LINEUP allocates most personalization capacity to shared reusable factors, composed using user history and the current query, while limiting target-user fitting to a compact code. Its analysis bounds finite-step code refinement under limited history.

4.2 HIERARCHICAL HISTORY–QUERY COMPOSITION

  • History recalls candidate factors, then each query calibrates their sparse composition; query calibration can reweight recalled factors or change the active subset.When candidate and active set sizes match, calibration reweights recalled factors; when the active set is smaller, it can also change which factors are active.Source: §4 · p. 5, §4.2 · pp. 5–6

4.1 FINE-GRAINED SHARED PERSONALIZATION FACTORS

  • LINEUP learns reusable low-rank personalization factors across source users, allowing overlapping factor subsets to produce distinct personalized updates without retraining.Factors are jointly learned rather than extracted from separately trained user adapters.Source: §4.1 · p. 5, §4 · p. 5

4.3 HISTORY-INITIALIZED ADAPTATION WITH TINY USER CODES

  • A single eight-dimensional code, initialized from population-informed information and fitted on user evidence, controls shared correction maps across layers and queries.Shared maps are fixed during target-user fitting; the code coordinates corrections across layers and is reused across queries.Source: §4 · p. 5, §4.3 · p. 6

4.4 THEORETICAL ANALYSIS OF COMPACT USER ADAPTATION

  • The finite-step risk bound relates shared-space approximation and history-initialization error to fitting progress and support–query disagreement, giving a sufficient condition for fitting to improve on initialization.The analysis accounts for history reuse and does not require convergence.Source: §4.4 · p. 6

5 EXPERIMENTS

Across six LaMP tasks, LINEUP leads on all reported metrics. Its results include strong performance with limited history and lower adaptation-storage growth, while ablations examine its retrieval, query-calibration, and correction choices.

5.1 EXPERIMENTAL SETUP

  • Evaluation uses six LaMP tasks spanning classification, rating prediction, and generation, with source and target users kept disjoint and scores averaged over three seeds.Baselines include private per-user adapters, retrieval, reusable pieces, representation shifts, and hypernetwork-generated adapters.Source: §5.1 · pp. 6–7

5.2 MAIN RESULTS ON LAMP

  • All 12 metrics across six tasks favor LINEUP; it improves LaMP-2M F1 by 0.020 over PerFit, reduces LaMP-3 RMSE by 0.050, and improves LaMP-7 ROUGE-L by 0.029 over OPPU.Tables 1 and 2 cover classification, prediction, and generation; the strongest baseline varies by task.Source: Table 1 · p. 7, Table 2 · p. 7, §5.2 · p. 7

5.3 COLD-START PERSONALIZATION

  • With one history record on LaMP-2N, LINEUP reaches 0.806 Accuracy and 0.560 Macro-F1, versus OPPU’s 0.793/0.536; full history raises Macro-F1 to 0.621.LINEUP has the best mean Accuracy and best or tied-best reported Macro-F1 across history budgets.Source: §5.3 · p. 8, Figure 3 · p. 8

5.4 EFFICIENCY AND SCALABILITY

  • At 1,024 users, projected total adaptation storage is approximately 151.2 MB for LINEUP, 1.61 GB for PerFit, and 8.60 GB for OPPU.LINEUP adds 16 B per user; its total storage becomes lower than OPPU’s at 17 users and PerFit’s at 95 users.Source: §5.4 · pp. 8–9

5.5 ABLATION STUDIES

  • Random candidates cause the largest degradation versus S1-only; removing query calibration also reduces performance, while initialized correction and subsequent support fitting add gains.The ablations cover LaMP-2M and LaMP-5; increasing code dimension from 2 to 8 improves mean performance, with no consistent further gains for larger codes.Source: §5.5 · p. 9

6 CONCLUSION AND FUTURE WORK

LINEUP combines reusable factors, conditional composition, and compact user-specific correction, and achieves the best mean performance across six LaMP tasks while optimizing eight scalars per target user. The authors identify cross-task personalization as future work.

6 CONCLUSION AND FUTURE WORK

  • Across six LaMP tasks, LINEUP achieves the best mean performance while optimizing only eight scalars per target user, with advantages under limited history and low incremental storage costs.The conclusion frames rich personalization as achievable without a full private adapter for every user.Source: §6 · pp. 9–10
  • Future work will extend the task-isolated formulation to cross-task personalization, using reusable factors to capture stable user preferences.Source: §6 · p. 10
  • LINEUP reduces persistent user-specific model state, but histories, codes, and shared components still require privacy safeguards; history-based personalization may also reflect bias or manipulation.Source: ETHICS STATEMENT · p. 10

A.1 PERSONALIZED LARGE LANGUAGE MODELS

The paper frames personalization as an allocation problem: what reusable capacity can be shared, how it should be composed, and what must remain user-specific. Its related-work discussion and appendix diagnostics position LINEUP’s factor-based reuse and test cross-user structure against controls.

A.2 MIXTURE OF EXPERTS

  • LINEUP’s factor-centric allocation differs from individualized adapters and collaborative adapter reuse by learning fine-grained factors jointly for cross-user composition from the outset.In the related-work account, MoE-style routing is hierarchical: history narrows candidates, then the query calibrates them.Source: §A.1 · pp. 17–18, §A.2 · p. 18

A.1 PERSONALIZED LARGE LANGUAGE MODELS

  • LINEUP learns reusable low-rank factors across source users, then composes them by user history and current query instead of constructing a complete adaptation state per user.Target-user fitting is confined to an eight-dimensional code over shared correction maps.Source: §A.1 · p. 18

B ADDITIONAL EMPIRICAL ANALYSIS

  • The appendix diagnostics use held-out users and independently fitted OPPU adapters for RQ1–RQ2, while RQ3 probes a task-adapted backbone without user-specific adapters.The diagnostic folds and learned bases are separate from LINEUP’s shared-training procedure.Source: B ADDITIONAL EMPIRICAL ANALYSIS · p. 19, §B.1.1 · p. 19

B.1.1 SHARED SUBSPACES AND UPDATE COVERAGE

  • Shared-subspace coverage measures how much held-out user-update energy falls in a basis learned from other users, with random subspaces as a reference.The random reference averages five Haar-random subspace draws; its coverage is approximately 0.0078 at m = 32 and 0.0312 at m = 128.Source: §B.1.1 · p. 19

B.1.2 RANDOM-PARTITION CONTROL

  • At m = 32, genuine-user bases exceed random-partition coverage by 0.143–0.210 on all four tasks, with confidence intervals excluding zero.Removing the task-adapter output span leaves these coverage gaps nearly unchanged.Source: §B.1.2 · p. 20, Table 4 · p. 20, Figure 4 · p. 20

B.1.3 TASK-ADAPTER CONTROL

These diagnostics test whether reusable directions reflect task-adapter span, user-specific utility, and query variation, and whether compact history-derived corrections transfer to held-out queries. The reported checks compare diagnostic predictions with exact intervention outcomes.

B.1.3 TASK-ADAPTER CONTROL

  • At m = 32, residual coverage after removing the task-LoRA output span is 0.533, 0.613, 0.672, and 0.615 on LaMP-2M/2N/5/7, respectively.Mean overlap with the removed span is 2.85%, 3.68%, 3.65%, and 2.00% on those tasks; coverage remains nearly unchanged.Source: §B.1.3 · p. 20

B.2.1 DIAGNOSTIC SETUP AND DIRECTION UTILITY

  • Direction utility is evaluated using fixed shared directions and diagnostic gates, with the first eight held-out queries per user and scale-normalized sensitivity responses.The gates are diagnostic variables, not a trained router.Source: §B.2.1 · p. 21

B.2.2 WITHIN-USER AND CROSS-USER AGREEMENT

  • Within-user versus cross-user agreement is 0.145 [0.107, 0.185] versus 0.008 [0.004, 0.014] on LaMP-2M, and 0.400 [0.330, 0.475] versus 0.059 [0.021, 0.112] on LaMP-2N.The within-user value is higher at all 64 sites on both tasks; the diagnostic uses fixed adapters and repeated held-out queries per user.Source: §B.2.2 · p. 21, §B.2.1 · p. 21

B.2.3 SCALE-CONTROLLED QUERY VARIATION

  • After sensitivity-scale control, within-user query variation is 0.827 [0.778, 0.869] on LaMP-2M and 0.613 [0.527, 0.699] on LaMP-2N.This indicates substantial query-dependent variation remains after normalization.Source: §B.2.3 · p. 21

B.2.4 EXACT DIRECTION-REMOVAL VALIDATION

  • First-order direction-removal predictions correlate with exact query-loss changes at 0.998955 on LaMP-2M and 0.999295 on LaMP-2N.Each task contributes 512 interventions.Source: §B.2.4 · p. 22, Figure 5 · p. 22

B.3.2 SUPPORT-TO-QUERY TRANSFER

  • Own-history correction transfer is positive on both tasks after removing the label-mean span, while random and matched donor results are weaker or negative.Own-history point estimates are positive at all eight layers; the label span is built from analysis-training users, not held-out query signals.Source: §B.3.1 · p. 22, §B.3.2 · p. 23

B.3.3 COMPACT CORRECTION SPACES

  • At r = 8, transfer retention exceeds residual-energy coverage on both tasks: 0.347 versus 0.129 on LaMP-2M and 0.294 versus 0.170 on LaMP-2N.At r = 16, coverage and retention rise to 0.168 and 0.425 on LaMP-2M, and 0.244 and 0.416 on LaMP-2N.Source: §B.3.3 · p. 23

B.3.4 FIRST-ORDER INTERVENTION VALIDATION

  • Median layerwise correlations between first-order predictions and exact intervention changes are 0.9998 on LaMP-2M and 0.9997 on LaMP-2N.The validation uses projected own-history directions in the r = 8 diagnostic space.Source: §B.3.4 · p. 23, Figure 6 · p. 24

C METHODOLOGY DETAILS

LINEUP learns shared factors and correction maps in stages, then personalizes unseen users by refining only a code initialized from history. The accompanying analysis bounds finite-step query risk and gives conditions under which fitting improves on that initialization.

C.1 ALGORITHM OF LINEUP

  • Stage 1 learns shared factor composition; Stage 2 freezes it while learning shared correction maps and a history-based code initializer, after which all shared parameters remain fixed for target-user fitting.The task-adapted backbone and feature extraction also remain fixed.Source: §C.1 · p. 24

C.2 SHARED FACTORS AND HIERARCHICAL COMPOSITION

  • History recalls candidate low-rank factors and the current query calibrates their sparse mixture, which is applied across adapted sites and request tokens.Factors are learned through source-user prediction under conditional composition rather than derived from diagnostic adapters.Source: §C.2 · pp. 24–26

C.3 TINY USER CODES AND SUPPORT ADAPTATION

  • History-derived initialization sets a shared correction code, and target-user support fitting refines only that code while shared correction maps remain fixed.The final code is reused across queries; the method uses rc = 8.Source: §C.3 · p. 26

D.1 HISTORY-INITIALIZED FITTING AND QUERY RISK

  • The finite-step comparison bounds query-risk change using support-loss progress, regularization, and discrepancy between query risk and support loss along the code update.The analysis assumes a bounded code displacement and continuously differentiable losses on the segment from initialization to the final code.Source: §D.1 · pp. 26–27

D.2 FINITE HISTORY AND SHARED CONDITIONING

  • For finite histories, the gradient-discrepancy bound depends on record replacement effects in conditioning and initialization, with terms scaling as O(κ^2/N + B^2/n) under the stated conditions.The assumptions include independent records, a data-independent support subset, bounded output-length ratios, bounded per-example code gradients, and controlled gradient sensitivity.Source: §D.2 · p. 27
  • The routing-stability condition preserves TopK sets under record replacement only when each nontrivial selection gap exceeds 2∆s; otherwise routing jumps enter the influence constants.History averaging changes scores by at most ∆s = O(1/N) under the stated bounded-feature and encoding assumptions.Source: §D.2 · p. 27

D.3 FINITE-STEP ADAM AND THE COMBINED BOUND

  • The numerical Adam bounds assume fixed learning rates and specified moment settings, while query improvement still requires control of the gradient-discrepancy term.The paper explicitly notes that the numerical bounds are not measurements.Source: §D.3 · p. 28
  • Fitting improves on history initialization when support-fit progress plus the regularization term exceeds the finite-history discrepancy term.The bound is for finite-step Adam; illustrative settings yield a step-radius bound of 0.029–0.230, not measured query improvements.Source: §D.3 · p. 28

D.4 CORRECTION GEOMETRY AND TRANSFER

The section analyzes when compact code refinement improves query risk and how support-to-query transfer is governed by the learned correction geometry. Its bounds are finite-step results under stated assumptions, not convergence claims.

D.4 CORRECTION GEOMETRY AND TRANSFER

  • Refinement lowers expected query error in a positive-curvature direction exactly when the remaining correction signal outweighs estimation noise under the stated moments.The directional risk difference depends on the correction magnitude, noise variance, sample size, curvature, and regularization.Source: §D.4 · p. 29
  • With S1 active, the first support-fitting step transfers to query-risk reduction through the positive semidefinite, rank-at-most-r operator V M1V^⊤.This connects transfer to learned correction geometry rather than an orthogonal projector.Source: §D.4 · p. 29
  • The finite-step NLL bounds require stated smoothness and influence conditions; the quadratic corollary additionally assumes common curvature and stated moments.The quadratic result does not predict an optimal code size or gains in every metric.Source: §D.4 · p. 29

E.1 DATASETS

The evaluation uses six diverse LaMP tasks, common scoring and backbone protocols, and independently seeded runs. A supplementary study extends the comparison to long-form generation under a focused baseline and user-split setup.

E.1 DATASETS

  • The six retained LaMP datasets span diverse tasks; LaMP-1 is excluded for unreliable citation-format evaluation under the OPPU prompt and Llama-2-7B, and LaMP-6 for privacy concerns.The task statistics are summarized in Table 5.Source: §E.1 · p. 29, Table 5 · p. 30

E.2 EVALUATION METRICS

  • Generation uses ROUGE-1 and ROUGE-L F-measures, while LaMP-3 numeric outputs are parsed as floats and unparsable outputs receive the farther endpoint in {1, 5}.All methods use the same evaluator and task metrics.Source: §E.2 · p. 30, Table 6 · p. 30

E.3 BASELINES

  • Baselines use official implementations under the common Llama-2-7B backbone and OPPU target-user protocol, with method-specific source-data construction and shared training retained.Per-Pcs and MTA lack complete released configurations for LaMP-2N; MTA also excludes LaMP-7.Source: §E.3 · pp. 30–31

E.4 IMPLEMENTATION

  • LINEUP’s shared components are trained on source-training users, hyperparameters are selected on source-validation users, and selected configurations are fixed before target-user evaluation.Means and standard deviations are reported over three independently trained seeds.Source: §E.4 · p. 31

E.5 HYPERPARAMETERS

  • Across tasks, LINEUP uses K = 32 rank-4 LoRA factors and 256-dimensional history and query representations; each target user fits only an eight-dimensional code for 10 Adam steps.Shared parameters remain fixed during target-user fitting.Source: §E.5 · p. 31

F.1 LONG-FORM GENERATION ON LONGLAMP

  • The long-form extension evaluates LongLaMP-2 abstract generation, LongLaMP-3 product review generation, and LongLaMP-4 topic writing.LongLaMP-1 is omitted because its Avocado email corpus requires licensed access.Source: §F.1 · p. 31, §F.1.1 · p. 31

F.1.1 EXPERIMENTAL SETUP

  • The LongLaMP setup uses user-disjoint source and target splits, task-specific history-derived supervision, and a focused comparison against Non, Basic, RAG, and OPPU.OPPU is the direct private-adapter reference; P2P is omitted because it uses a different protocol.Source: §F.1.1 · pp. 31–32

F.1.2 RESULTS ON LONGLAMP

LINEUP leads the LongLaMP comparison across all six metrics, with OPPU the strongest baseline throughout. The reported advantage extends the paper’s capacity-allocation findings to long-form generation.

F.1.2 RESULTS ON LONGLAMP

  • LINEUP achieves the best mean performance on all six LongLaMP metrics across three tasks, with OPPU the strongest baseline throughout.Table 9 reports the long-form generation comparison.Source: §F.1.2 · p. 32
  • On Product Review Generation, LINEUP improves over OPPU by 0.056 in ROUGE-1 and 0.048 in ROUGE-L.Source: §F.1.2 · p. 32
  • The long-form advantage is consistent with reusable factors, history–query composition, and compact user codes supplying shared capacity and residual individual correction.The paper concludes that long-form personalization can benefit from rich shared capacity without a complete private adapter for each user.Source: §F.1.2 · p. 32

F.2 ADDITIONAL COLD-START RESULTS

The supplementary analyses test performance under restricted histories, storage growth, and user-code dimension choices. They also specify the controlled ablation variants used to probe LINEUP’s components.

F.2 ADDITIONAL COLD-START RESULTS

  • With one history record on LaMP-4, LINEUP reaches 0.219/0.197 ROUGE-1/ROUGE-L versus OPPU’s 0.215/0.191, and leads on both metrics at every history budget.Figure 7 also reports that one interaction nearly recovers full-history ROUGE-1 and that LINEUP maintains the highest ROUGE-L across budgets.Source: Figure 7 · p. 33, §F.2 · p. 33, Table 13 · p. 36

F.3 ADDITIONAL EFFICIENCY AND SCALABILITY RESULTS

  • At 1,024 users, projected storage is approximately 151.18 MB for LINEUP, 18.09 GB for Per-Pcs, and 22.47 GB for MTA.For LaMP-2N, LINEUP becomes more storage-efficient than OPPU at 17 users and PerFit at 95 users, but PerFit remains lower at the evaluated 49-user size.Source: Figure 7 · p. 33, §F.3 · p. 34

F.4 ABLATION PROTOCOL

  • The ablation protocol contrasts history-guided retrieval with random candidates, query calibration with λ = 0, and full LINEUP correction with S1-only or initialized-code variants.Initialized-code comparisons separate shared maps and history initialization from the additional contribution of support fitting.Source: §F.4 · p. 34

F.5 USER-CODE DIMENSION SENSITIVITY

  • Increasing code dimension from 2 to 8 improves all four reported means, but larger codes yield no consistent gains; the default rc = 8 matches the best reported generation means.At rc = 32, LaMP-2M Accuracy reaches its highest mean, while Macro-F1 and both ROUGE means decrease relative to rc = 16.Source: §F.5 · p. 35, Table 11 · p. 35

Related papers

Loading 2610.02353v1…