Not All Is Lost: Repairing Lossy User Preference States of Personalization Encoders

1 Introduction

• A task head operating only on the compressed preference state has no direct access to evidence retained elsewhere in the encoder computation.The paper calls the possibility of correcting that state from cached representations the recoverability gap.Source: §1 · p. 1
• REPAIR compares the exposed state with cached timestep representations, learns corrective evidence across time, and selectively adds it to the state before the task head.Its temporal views emphasize extended history, recent evidence, and localized bursts; the encoder remains unchanged and does not reprocess raw history.Source: §1 · p. 2
• K = 128 gives the best MRR and nDCG@10 in all five tested settings, while increasing K to 192 lowers both metrics.Combining long-, short-, and episodic-support views is strongest in 11 of 12 controlled settings; removing temporal resolution can make REPAIR worse than the original frozen host.Source: §1 · p. 2
• Under the frozen-host protocol, REPAIR outperforms two evaluated post-head alternatives in all eight recommendation settings.In generation, personalization measured by PerSEval can improve substantially even when ROUGE and METEOR change little.Source: §1 · p. 2
• REPAIR improves MRR and nDCG@10 for all twelve recommendation hosts with both encoder and task head frozen, and outperforms head-only finetuning in every evaluated setting.The paper also reports gains in personalized generation and distinguishes those gains from generic text quality.Source: §1 · p. 2

Preference Encoders And State Compression

• Preference encoders use diverse architectures to contextualize and compress user histories into a task-facing state; REPAIR asks whether cached representations can improve that state after training.Source: §2 · p. 2

Temporal And Multi-Interest Preference Modeling

• Multi-timescale and multi-interest preference modeling are established, but REPAIR applies temporal bases to corrective evidence after compression rather than to construction of the host’s history state.Its questions concern the compactness of the correction and which temporal evidence remains useful across frozen hosts.Source: §2 · p. 2

Personalized Generation And Downstream Utilization

• Personalized generation draws on history, profiles, retrieval, or prompts, while REPAIR examines how downstream interfaces use a repaired state and separates personalization gains from generic text-quality changes.Source: §2 · p. 3

Representation Intervention, Model Editing, And Denoising

• REPAIR uses cached per-timestep representations to correct an already formed preference state before downstream use, with the host encoder fixed and without re-encoding raw history.This distinguishes its correction target and evidence source within a broader intervention setting that includes parameter updates, model editing, denoising, and diffusion-based recommendation representations.Source: §2 · p. 3

3 Preliminaries: Preference State Compression & Recoverability

• A user history is represented as a time-ordered sequence of interactions, each pairing an action with associated content.Actions include positive, negative, and task-command types; examples include clicking, skipping or rejecting, and specifying a personalization command.Source: §3 · p. 3
• The encoder compresses per-timestep representations into a task-facing preference state before passing it to the downstream task head.Recommendation predicts or ranks the next item, while generation additionally uses a task-side query or content to produce a personalized response.Source: §3 · p. 3
• The recoverability gap is useful preference evidence that may be weakly expressed in the compressed state while remaining available in cached per-timestep representations.Such evidence may support state correction with the encoder frozen, while performance gains also depend on how effectively the downstream head uses the corrected state.Source: §3 · p. 3

4 Methodology

• REPAIR is inserted between a pretrained personalization system’s encoder and downstream task head, using the existing frozen host computation rather than rerunning the history.The cache is bounded by the host history window and costs O(Lde) for maximum length L.Source: §4 · p. 4
• Learned projections map cached representations and the compressed state into a common space, where REPAIR forms a state-relative corrective signal and probes it with K learned patterns.The factorization is learned rather than computed by SVD; K = 128 is the default, and varying K tests compactness.Source: §4 · p. 4
• REPAIR centers each pattern’s activation by removing its prefix-wide average, so patterns activated equally throughout the prefix vanish before later state-conditioned steps.Source: §4 · p. 4
• Long-, short-, and episodic-support views resolve timestep-wise corrective activations for extended history, recent evidence, and localized bursts.They are views of the same corrective activation, not labels that classify interactions; their combined support is learned.Source: §4 · p. 4
• REPAIR retains selected timestep-pattern activations, forms write candidates, pools them, and adds the resulting correction to the original state before passing it to the task head.A hard forward mask retains activations while a straight-through surrogate trains the retention fraction; the original state remains the base during residual fusion.Source: §4 · pp. 4–5
• The evaluation design ranks one recommendation target against 499 sampled negatives.Source: Table 1 · p. 5
• For prompt-conditioned LLM hosts, interaction evidence is encoded separately because the frozen LLM does not directly expose an equivalent interaction-level sequence.Training combines downstream task loss with relative-position and source-evidence coherence supervision, using fixed coefficients across hosts.Source: §4 · p. 5

I. Experimental Setup

• The study evaluates recommendation and personalized generation on MovieLens, MIND, PENS, and Amazon Reviews 2023, with TG-ReDial as a separate discourse-derived stress test.Histories use user-disjoint train/test splits; each prediction uses an observed history prefix and its next supervised target.Source: §5 · p. 5

IV. Downstream-Utilization Diagnostics; V. Statistical Evaluation

• Recommendation ranks one target against 499 sampled negatives; generation uses quality and personalization-sensitive metrics, while selected hosts receive three-seed variability and user-clustered bootstrap intervals.The bootstrap uses 10,000 user-clustered resamples; extended settings without repeated-run uncertainty are reported as point estimates.Source: §5 · pp. 5–6

I. Experimental Setup

• Hosts span recurrent, self-attentive, state-space, news, multimodal, and personalized-generation systems, testing correction across compression mechanisms and downstream interfaces.The inventory also includes frozen LLM hosts receiving repaired states through continuous prefixes.Source: §5 · p. 6

II. Frozen-Host Protocol; III. Corrective-Structure Diagnostics

• The frozen-host comparison trains only REPAIR, contrasting it with the unchanged host, head-only finetuning, and REPAIR-head alignment while keeping the encoder frozen.The diagnostics vary corrective-pattern count, construct long-range, recent, or episodic evidence controls without retraining, and ablate temporal resolution and selective retention.Source: §5 · p. 6

6.1 RQ1: Post-Compression Recoverability With A Frozen Host

• Average MRR/nDCG@10 gains from frozen-host REPAIR are +2.66/+2.63 on MovieLens, +2.29/+2.16 on PENS, +1.96/+3.08 on MIND, and +2.19/+2.55 on Amazon Reviews 2023.The gains cover all twelve representative hosts; cached per-timestep representations improve the state used by the unchanged task head.Source: §6.1 · p. 6, Table 2 · p. 7
• Head-only finetuning gains +0.29, +0.76, +0.72, and +0.68 average MRR across the four datasets, versus REPAIR’s +2.66, +2.29, +1.96, and +2.19; the ordering holds in all twelve settings.Joint REPAIR-head alignment further improves MRR and nDCG@10 in 11/12 settings while keeping the encoder frozen.Source: §6.1 · pp. 6–7

6.2 RQ2: Structure Of The Learned Corrective Representation

• K = 128 gives the best MRR and nDCG@10 in all five tested settings, while K = 192 lowers both metrics in all five.The mean learned read-coordinate energy ratio at K = 128 is 96.46% across those settings; it is distinct from a leading-eigenspace energy fraction.Source: §6.2 · p. 7, Table 3 · p. 7
• The input-timing pattern predicts corrective timing in only 6/12 probes; Mamba4Rec selects short-support correction across all three controlled input patterns.Full REPAIR’s combined long-, short-, and episodic-support views outperform the best isolated view in 11/12 settings.Source: §6.2 · p. 7, Table 4 · p. 8

6.3 RQ3: Where And How Is Recovered Preference Evidence Used?

• Table 5 compares pre-head REPAIR with two adapted post-head correctors using the same task-trained host and two-epoch downstream alignment protocol.The caption identifies the evaluated comparison but does not report which method performs best.Source: Table 5 · p. 8

6.2 RQ2: Structure Of The Learned Corrective Representation

• Selective write-back improves both MRR and nDCG@10 for three of four hosts by prefix length 40 and all four by length 50; source-coherence supervision gives the strongest complete objective across four hosts.Selection acts at the timestep-pattern level, and timestep provenance adds gains beyond task supervision alone.Source: §6.2 · p. 8

6.3 RQ3: Where And How Is Recovered Preference Evidence Used?

• Walk2Pers gains +13.25%/+15.32% and IMPerSumm +20.76%/+25.23% on PSE-W-JSD/PSE-W-RG-L; IMPerSumm also gains 39.29% on PSE-P-RG-L.These are relative gains on PENS; PerSEval measures personalization with respect to evolving user preference.Source: §6.3 · p. 8, Table 6 · p. 9
• For frozen LLM hosts, DeepSeek improves generation quality more than personalization, whereas Qwen improves both; the repaired state enters through a continuous prefix.The gains differ across decoders, and the comparison does not identify the responsible decoder mechanism.Source: §6.3 · pp. 8–9

• REPAIR avoids re-encoding history by reusing the frozen forward pass, but requires retaining cached evidence and running the repair module.Across evaluated predictive hosts, parameter-derived memory overhead is 31–36% and latency overhead is 38–44%.Source: §7 · p. 9
• Evidence discarded before the host processes history is outside REPAIR’s recovery path.Source: §7 · p. 9
• On TG-ReDial, preference actions must be inferred from recommendations, acceptance or rejection, and topic shifts; sequential hosts benefit, but mixed TextCNN results do not support uniformly beneficial correction across inferred interaction interfaces.Source: §7 · p. 9

• REPAIR learns compact corrective patterns from cached representations and selectively updates compressed preference states before downstream use.Source: §8 · p. 9
• Across twelve recommendation hosts, REPAIR improves MRR and nDCG@10 with both encoder and task head frozen, outperforming head-only finetuning in every evaluated setting.Source: §8 · p. 9
• The analyses find host-dependent corrective timing and show that personalization gains in generation can differ from generic text-quality gains.Future work can study retaining useful corrective evidence under tighter cache budgets and improving downstream interfaces’ use of repaired states.Source: §8 · p. 9

Reproducibility Statement

• The reproducibility statement locates the repair computation, host interfaces, training details, evaluation protocols, and extended analyses across Appendices A–G and Tables 7–9.The implementation maps state and cached interactions into a common space, reads corrective coordinates, resolves and selects evidence, pools it, then fuses the correction with the state.Source: Reproducibility Statement · p. 9, A REPAIR Implementation Details · p. 15

A.1 User Interaction Graph And Eventization

• User Interaction Graph eventization represents heterogeneous histories as ordered action–content pairs, with actions marking positive feedback, negative feedback, or task commands.Predictive personalization uses the next content item as its target; generative personalization stores the reference response as the target and keeps the task-side source or query separate from the preference state.Source: §A.1 · p. 15

A.2 Frozen Host Interface

• A frozen host supplies both a task-facing state and cached per-position representations from the same history computation, without introducing a second history encoder.The cache is taken before terminal compression or readout; its form varies by host architecture, including recurrent, contextualized, state-space, and fused multimodal sequences.Source: §A.2 · pp. 15–16

A.3 State-Relative Corrective Evidence

• Because cached vectors and the final state may differ in dimension or meaning, projections map them into a shared space before residuals are read into corrective activations.Centering emphasizes variation across history positions; the common state term cancels there, while state conditioning remains in later gating, aggregation, and fusion.Source: §A.3 · pp. 16–17, Figure 2 · p. 17

A.4 Temporal Resolution

• Temporal resolution weights corrective activations by lag through long-support decay, a short-support recency envelope, and an episodic function that can emphasize localized lags.The L, S, and E labels name temporal functions, not user intent; trainable maps mix their weighted coordinates into a shared field.Source: §A.4 · pp. 17–18

A.5 Adaptive Top-ν Retention; A.6 Correction Aggregation And Fusion

• Adaptive top-ν retention selects eligible timestep activations independently for each corrective coordinate, using a hard forward mask and a smooth training surrogate.State gating scales coordinates, normalized weights pool selected evidence across timesteps, and residual-gated fusion adds the correction to the original state; the objective families supervise task utility, temporal provenance, and source coherence.Source: §A.5 · p. 18, §A.6 · pp. 19–20

A.7 Prompt-Conditioned LLM Hosts

• For prompt-conditioned LLM hosts, REPAIR compares separately constructed interaction vectors with the frozen prompt’s terminal state and converts the repaired state into a continuous soft prefix.Prefix insertion changes the decoder input, so decoding uses an attention cache corresponding to that modified input rather than reusing the original prompt cache.Source: §A.7 · p. 20

B Theoretical Justification; B.1 Compact Approximation Of The State-Relative Residual

• The theoretical analysis follows the encoder-host repair pathway and gives conditional results on residual approximation, state-correction realizability, temporal structure, and objective-family control.Its compactness analysis considers spectral and learned subspaces under a finite-second-moment assumption; the experiments test observable consequences, not the assumptions themselves.Source: B Theoretical Justification · pp. 20–21, §B.1 · pp. 22–23

Appendix A.3 constructs the unreduced state-relative residual

• At K = 128, mean learned read-coordinate energy is 96.46% across five encoder-host settings, while this statistic is not the leading-eigendirection energy fraction or a numerical test of the spectral bound.K = 128 also gives the best MRR and nDCG@10 among four tested capacities; the learned read vectors need not be orthogonal, so the energy ratio can exceed one and 1 − Rread(K) is not a spectral-tail fraction.Source: Appendix A.3 · p. 24

B.2 Conditional Recoverability Through Selective Write-Back And Fusion

• Theorem 2 bounds conditional recovery through write-space, coordinate, aggregation, and fusion errors, assuming an appropriate oracle correction exists.The bound covers the complete pre-head pathway; it does not establish that information is absent from the original state.Source: §B.2 · pp. 25–26
• Improvements across all twelve representative recommendation hosts with a frozen encoder and head show useful state corrections can be realized from cached representations; head-only continuation gives substantially smaller gains.The experiments do not identify the oracle state or separately estimate the four approximation terms, and the jointly adapted REPAIR + Head Alignment condition is outside this attribution.Source: §B.2 · p. 26

B.3 Identifiability Of Temporal Residual Structure

• The temporal dictionary’s coefficients are identifiable when its atoms have full column rank, while its finite-prefix expressivity result is not an identifiability or generalization guarantee.Across controlled host and trace settings, the combined L + S + E pathway is strongest in eleven of twelve; full-history selective write-back improves all four representative hosts while retaining broad prefix support.Source: §B.3 · pp. 27–28

B.4 Backpropagation Through Temporal Recovery Kernels

• Temporal-parameter gradients depend jointly on host-specific post-compression residuals and the supervision reaching them, so two frozen hosts on the same raw trajectory need not learn the same temporal mapping.Theorem 4 bounds local temporal-kernel gradients under bounded residuals, gradients, and operator norms.Source: §B.4 · p. 30
• The combined pathway is strongest in eleven of twelve settings, while preferred repair branches vary by host under controlled trace comparisons.Mamba4Rec favors short-support correction across L-Tr, S-Tr, and E-Tr; TiSASRec, NAML, and EBNR show different host and trace responses. Simultaneous gradient credit to WL, WS, and WE is compatible with mixed residuals, but the results do not verify Theorem 4’s bounded-gradient assumptions.Source: §B.4 · p. 30

B.5 Local Control Induced By The Repair Loss Families

• Under stable-decomposition and local family-control assumptions, positive weights on task utility, timestep provenance, and source coherence provide local control through their corresponding error components.The theorem concerns functional loss families, not individual primitive losses; the implementation groups its terms into these three families.Source: §B.5 · pp. 31–32
• Adding timestep-provenance supervision improves the intermediate ablation configuration, and adding source-coherence supervision yields the strongest configuration across four hosts.The progressive ablation does not isolate each family's causal contribution or verify the theorem's local-decomposition and quadratic-control assumptions.Source: §B.5 · pp. 31–32

C.1 Shared Evaluation Protocol

• Train and test users are disjoint, and paired conditions share held-out users, history prefixes, targets, and candidate sets.Recommendation uses one ground-truth item and 499 negatives per test instance; representative repeated-run results average three seeds.Source: §C.1 · p. 33, C Evaluation Details · p. 33

C.2.1 Common Interaction Representation

• The common interaction representation maps records into action-content pairs, distinguishing actions such as clicks and rejections without adding a trainable history encoder.Each resulting prefix is converted to the input format required by its host.Source: §C.2.1 · p. 33

C.2.2 Encoder-Host Representation Interface

• The frozen encoder supplies the task-facing state and cached timestep representations, which REPAIR uses to change the state passed to the task head without re-encoding the sequence.The cache boundary is defined in Appendix A.2.Source: §C.2.2 · p. 33

C.2.3 Dataset-Specific Eventization / C.2.4 Supervised Target Construction

• Movie and product histories use chronologically ordered positive or negative interactions, while news histories retain observed order and use the next positive item as the recommendation target.Generation supplies the source article or query separately; held-out references are excluded from training-history construction.Source: §C.2.3 · p. 33

C.2.5 Training-Only Trajectory Augmentation

• Training-only trajectory augmentation can replace later trajectory segments with contiguous training segments, selecting surrogate responses using context similarity and recency weighting.Validation and test trajectories are not substitution sources.Source: §C.2.5 · p. 34

C.2.6 Evaluation Data Path

• Within each host comparison, the encoder checkpoint, held-out trajectory, target, and evaluation candidates remain fixed across original-host, head-only, and REPAIR conditions.The trainable component varies by condition.Source: §C.2.6 · p. 34

C.3 Dataset-Specific Evaluation Tasks

• MovieLens, MIND, PENS, and Amazon Reviews 2023 use sampled-candidate recommendation with one positive and 499 negatives, across 15,000, 20,000, 20,000, and 2,000 held-out instances, respectively.Amazon uses frozen SigLIP2, OpenCLIP, or DINOv2-E5 representations in its host settings.Source: §C.3 · pp. 34–35
• PENS headline generation uses 20,000 held-out examples containing user history, news content, and a user-specific headline; TG-ReDial separately evaluates dialogue-derived recommendation and response generation.PENS generation metrics are computed from the same generated outputs, while TG-ReDial uses SASRec, GRU4Rec, and TextCNN for recommendation and ReDial-style or Transformer sequence-to-sequence interfaces for responses.Source: §C.3 · p. 35
• Evaluation-only L-Tr, S-Tr, and E-Tr controls place target-conditioned evidence across about 60% of a trajectory, its final 40%, or a random contiguous 3–4-timestep segment.These controls preserve the supervised target, leave trained parameters unchanged, and are not ordinary next-item inference conditions; results are reported separately in Table 4.Source: §C.3 · p. 35

C.4.2 Head-Only Continuation Control

• The Head-only Continuation control updates only the original task interface for five additional epochs, leaving the host encoder and preference state unchanged and omitting REPAIR.It retains the checkpoint with the strongest validation performance across those continuation epochs.Source: §C.4.2 · p. 35

C.4 Training And Optimization Protocol

• Stage 2 trains REPAIR while freezing both the host encoder and task interface; Stage 3 jointly trains REPAIR and the task interface while keeping the encoder frozen.Stage 2 selects a validation checkpoint, which initializes Stage 3; recommendation and generation alignment each run for two epochs.Source: §C.4.3 · pp. 35–36, §C.4.5 · p. 38

C.4.4 Stage-2 Repair Objective

• Stage 2 supervises task utility, timestep provenance, and source coherence; its better-than-base penalty applies during training but does not guarantee non-degradation on every held-out instance.Training-side auxiliary branches perturb item identities, action labels, or adjacent interactions, while clean prefixes supply task-utility losses.Source: §C.4.4 · pp. 36, 38

C.4.7 Randomness, Pairing, And Uncertainty

• Repeated-run uncertainty is reported only for OpenCLIP and SigLIP2 on Amazon and EBNR on PENS; paired conditions share held-out users and evaluation inputs, with intervals bootstrapped by user.The 95% intervals use 10,000 user-level bootstrap samples, retaining all observations for each sampled user.Source: §C.4.7 · p. 39

D Evaluation Metrics And Reporting

• Recommendation metrics rank a single target, generation-quality metrics compare outputs with references, and personalization-sensitive metrics compare outputs for different users given the same source.Recommendation scores are presented after multiplying native values by 100, while PerSEval values use a 1000× display factor; relative gains are reported separately from score-point gains.Source: D Evaluation Metrics And Reporting · p. 39, §D.1 · pp. 39–41

D.2 Recommendation Metrics

• With one relevant target per test instance, Hit@K is equivalent to Recall@K, and nDCG is reported at @5 or @10 when specified by the result table.These are recommendation-evaluation conventions for the corresponding result tables.Source: §D.2 · p. 41

D.3 Generation Quality Metrics

• Generation quality is evaluated with ROUGE, METEOR, BLEU, and BERTScore, alongside weighted and paraphrase-aware variants.The reported ROUGE variants are those named in the corresponding result table.Source: §D.3 · pp. 41–42
• Weighted ROUGE distinguishes weighted n-gram overlap from weighted longest-common-subsequence overlap, while the P prefix permits paraphrase matches and P-RG-L denotes the reported LCS variant.For weighted LCS, a superlinear match-length function rewards uninterrupted matches more strongly than fragmented ones.Source: §D.3 · pp. 41–42

D.4 Personalization-Sensitive Generation Metrics

• PerSEval combines agreement between expected and generated user-specific output separation with an accuracy penalty, rather than treating output diversity alone as personalization.It is computed after generation and is not part of generation or model selection unless explicitly stated.Source: §D.4 · p. 43
• PerSEval variants use plug-in text distances, including Jensen-Shannon and weighted-JSD lexical distances, as well as weighted and paraphrase-aware ROUGE-L; all are computed from the same generated outputs.Similarity-to-distance conversions use native scores before display scaling.Source: §D.4 · p. 44

D.5 Deployment Metrics

• The memory column is a parameter-storage proxy and does not measure peak device memory, activations, or KV-cache residency.Deployment measurements compare REPAIR with its corresponding frozen host within the same host-specific execution environment.Source: §D.5 · p. 44
• Throughput may be measured directly or derived from reciprocal latency, so the reported basis distinguishes those two cases.The passage specifies these alternatives but does not provide their formulas.Source: §D.5 · p. 44
• The predictive score-resource ratios compare multiplicative MRR improvement with multiplicative resource cost; values above one mean relative score gains exceed relative cost increases.These ratios are descriptive, not training objectives or guarantees of absolute serving quality.Source: §D.5 · p. 44

E Task-Specific Frozen Hosts

• Hosts span recurrent, attentive, bidirectional, time-aware, state-space, denoising, frequency-aware, hierarchical, multimodal, and prompt-conditioned interfaces.The selection is intended to test repair across different ways of forming and consuming a task-facing state, not to rank all hosts as one architecture leaderboard.Source: E Task-Specific Frozen Hosts · p. 44, §E.1 · p. 45, §E.2 · p. 45

E.1 Movie Recommendation Hosts

• Movie recommendation hosts vary in sequence modeling, including selective state-space models and hosts that already model noise or temporal structure.The two cross-host transfer rows are distinguished from directly trained repair modules in Table 11.Source: §E.1 · p. 45

E.2 News Recommendation Hosts

• News hosts test repair after content-view attention, recurrent compression, self-attention, long- and short-term modeling, multiple interests, or history filtering.SCAPE-GRU is evaluated through a recommendation head, with REPAIR attached after the task-trained user representation is formed.Source: §E.2 · p. 45

E.3 Product Recommendation Hosts

• Product recommendation extends state correction to histories containing frozen image and text features.SigLIP2, OpenCLIP, and DINOv2-E5 provide product representations before the host forms its task-facing state.Source: §E.3 · p. 45

E.4 Personalized Generation Hosts

• Personalized generation hosts vary how preference representations condition headline or text generation, including through a continuous prefix for frozen LLMs.PENS T1 initializes the decoder hidden state, whereas T2 personalizes attention over source-news words.Source: §E.4 · p. 45

E.5 Conversational Hosts

• TG-ReDial tests recommendation and response-generation consumers using interaction sequences inferred from dialogue.The recommendation hosts include SASRec, GRU4Rec, and TextCNN; generation uses a ReDial-style system and a Transformer sequence-to-sequence model.Source: §E.5 · pp. 45–46

F.1 Shared Comparison Protocol

• The corrective comparison uses the same task-trained checkpoint, split, held-out instances, targets, candidate construction, and two-epoch alignment protocol.Parameter count and FLOPs are not matched unless stated, and the adapted correction settings are not exact reproductions of their source systems.Source: §F.1 · p. 46

F.2 Rec-Denoiser-Corr

• Rec-Denoiser-Corr changes output-side timestep influence, while SERAC-Corr retrieves training examples for a post-head counterfactual correction.Rec-Denoiser-Corr uses host attention when available or prediction-side timestep sensitivity otherwise; neither adaptation changes the pre-head preference state, and each is evaluated only in the RQ3 recommendation comparison.Source: §F.2 · p. 46, §F.3 · pp. 46–47

F.4 Scope Of The Intervention-Locus Comparison

• REPAIR modifies the task-facing preference state before the original ranking head, unlike the two evaluated post-head corrective adaptations.The comparison changes both correction architecture and intervention location, so it does not isolate intervention locus as a single causal factor or establish dominance over other possible architectures.Source: §F.4 · p. 47

G Additional Results And Extended Analysis

• The additional analyses distinguish state changes supplied to a fixed head from changes after head alignment and cover generation, temporal resolution, supervision, selection, noise, serving cost, and dialogue-derived histories.The experiments do not estimate the oracle state or certify the assumptions in Appendix B.Source: G Additional Results And Extended… · p. 47

G.1 Full Predictive Recommendation Results Across Frozen Encoders

• Stage-3 repair with head alignment improves reported recommendation hosts across MovieLens, PENS, MIND, and Amazon Reviews 2023.Examples include Mamba4Rec’s 4.97-point MRR gain on MovieLens, NAML’s increase from 1.29 to 12.38 MRR on PENS, and Amazon MRR gains of 15.33 to 17.91 points across three multimodal hosts.Source: §G.1 · pp. 48–49, Table 14 · p. 50, Table 13 · p. 50

G.2 Generative Accuracy And Personalization-Sensitive Metrics

• IMPerSumm’s first three PerSEval variants improve 20.76%, 25.23%, and 39.29%, even as some standard quality scores change little or decline.For example, its ROUGE-L decreases from 14.67 to 14.27 despite a small METEOR gain; PerSEval evaluates user-specific responsiveness, distinct from reference-overlap quality.Source: §G.2 · p. 51, Table 16 · p. 52

G.3 Interpretation Of Temporal Resolution Ablations

• The full L + S + E temporal configuration has the highest MRR for all six hosts, while removing temporal resolution scores below full repair in every row.In controlled probes, the combined pathway is strongest in eleven of twelve settings; the six-host ablation and twelve probes are distinct comparisons.Source: Table 18 · p. 53, §G.3 · p. 53

G.4 Training Objective Ablation

• The complete progressive supervision objective yields the highest reported MRR for all four hosts; provenance and coherence add gains beyond task supervision alone.Adding provenance raises NRMS from 46.39 to 51.29 MRR, and adding coherence raises it further to 52.74; the fixed-order ablation does not test every possible combination.Source: §G.4 · pp. 53–54, Table 19 · p. 54

G.5 Length-Controlled Selective Retention

• At history length 50, selective retention improves both MRR and nDCG@10 over non-selective repair for all four hosts.At shorter histories, selection reverses for some host-metric settings, so the evidence supports a more consistent benefit at the longest tested prefix, not at every length.Source: Table 20 · p. 54, §G.5 · p. 55

G.6 Noisy-Trajectory Stress Test

• With approximately 35 to 40% of valid history positions replaced by random training-pool content, repair improves all four reported settings.Mamba4Rec rises from 18.21 to 21.43 MRR under the same corruption; these results do not establish gains under arbitrary corruption or recovery of clean-history scores.Source: Table 21 · p. 55, §G.6 · p. 55

G.7 Serving Cost And Compute Footprint

• Encoder score-to-latency ratios range from 0.75 to 0.84 and score-to-memory ratios from 0.79 to 0.89, indicating resource increases exceed relative MRR increases.Mamba4Rec gains 16.96% MRR with 40% latency overhead; parameter-derived memory excludes activations and caches, and throughput change is derived from reciprocal latency.Source: Table 22 · p. 56, §G.7 · p. 56
• At history length 50, repair adds 0.18 GFLOPs to Mamba4Rec’s 0.75 base and 0.02 GFLOPs to NAML’s 0.03 base.The equal repair workload implies different relative overheads across hosts; serving results assess evaluated budgets without defining a universal acceptable cost or numerically establishing the spectral-tail condition.Source: §G.7 · pp. 56–57

G.8 Broader Applicability On TG-ReDial

• On TG-ReDial, SASRec improves from 4.11 to 4.56 MRR and TextCNN’s Hit@10 decreases from 5.08 to 5.04.GRU4Rec MRR rises from 4.14 to 4.56, while TextCNN MRR rises from 2.57 to 2.78; its nDCG@5 and nDCG@10 also decrease, so results are not uniformly beneficial across hosts.Source: §G.8 · pp. 57–58, Table 25 · p. 58
• On TG-ReDial, the ReDial-style generator improves from 12.33 to 27.76 ROUGE-1, while the Transformer changes from 31.81 to 31.97.The comparison does not establish decoder sensitivity or show that both systems recovered the same internal correction; generation scores are distinct from recommendation ranking scores.Source: §G.8 · p. 58, Table 25 · p. 58

H License And Usage Statement

• The usage statement distinguishes model and code licenses from dataset access terms, and says raw user histories and third-party host checkpoints are not redistributed.It also states that benchmark data are used without attempting to identify users or combine records to deanonymize individuals.Source: H License And Usage Statement · pp. 58–59