When History Misleads: Asymmetric Margin Supervision for Instruction-Guided LLM Generative Recommendation

• History editing can remove information useful for personalization, so AIMS uses deletion effects as supervision rather than changing the training input.The request-dependent score gap is between the target and a competitor near the recommendation cutoff.Source: §1 · p. 2
• History-event influence does not determine signed support for the target, and target-score gains alone do not determine the decoded ranking.Removing an event can raise a competitor's score more than the target's, reducing their margin.Source: §1 · p. 2
• AIMS uses a frozen reference model to select deletions that improve both target score and target–competitor margin, then learns request-specific margins from complete histories.The auxiliary loss penalizes margin shortfalls and routes its gradient only through the competitor score; inference uses the original decoder without deletion search.Source: §1 · p. 2
• AIMS improves Recall and NDCG over strong baselines across an industrial dataset and two public benchmarks.Matched ablations support request-specific margins, margin-based deletion selection, complete-history training, and ACS; selected deletions remove violating history events more often than chance.Source: §1 · p. 2

2 Diagnosing Instruction-Conditioned History Use

• The audit freezes an SFT checkpoint and intervenes on one history event at a time, masking it while leaving the request and other events unchanged.It defines event utility for a candidate from the score change between the full and intervened views.Source: §2.1 · p. 3, §2 · p. 3

2.1 Request dependence of utility

• For both backbones, factual-request utility rankings aligned better with paraphrases than generic requests, whose sign disagreements were higher by 17.10 points for Llama and 4.65 for Gemma.The comparison held histories, targets, and masked events fixed.Source: §2.1 · p. 3

2.2 Influence–utility misalignment

• Prediction influence is Jensen–Shannon divergence between full- and masked-history first-token distributions, but it does not determine whether an event supports the target.The most influential event was not reliably more likely than a random event to support the target.Source: §2.2 · p. 4
• Influence correlated more with absolute than signed target utility: Spearman correlations were 0.300 versus 0.045 for Llama and 0.167 versus 0.002 for Gemma.Maximum-influence selection barely raised supportive-event rates versus uniform selection under factual requests; both intervals included zero, and estimates were negative under generic requests.Source: §2.2 · p. 4

2.3 From target scores to margins and decisions

• A fixed competitor is chosen at the top-R boundary of the initial beam, and the target–competitor margin is the difference between their scores.The competitor is the first excluded candidate when the target is already in the top-R list, and otherwise the candidate at rank R.Source: §2.3 · p. 4
• The audit compares the target and fixed competitor after deleting the highest-uplift event from initially missed requests when uplift exceeds 10−4 nats.The margin improves only when target uplift exceeds competitor uplift; re-decoding measures whether the target enters the top-10 list.Source: §2.3 · p. 4
• On initially missed requests, deletions with the largest target uplift still decreased the target–competitor margin for 11.8% of Llama and 16.7% of Gemma requests.Re-decoding recovered 1.08% and 1.50% of them, respectively, as top-10 hits.Source: §2.3 · p. 4
• A fixed-competitor margin is only a pairwise surrogate: top-R membership also depends on other candidates and beam-search output.The target can enter the decoded list even when this margin decreases.Source: §2.3 · p. 4

3 Asymmetric Intervention-Guided Margin Supervision

• AIMS builds counterfactual margins for initially correct requests and trains the student to reach them from complete histories.The frozen SFT checkpoint serves as reference and initializes the student; continued cross-entropy can recover missed targets but also lose initial top-10 hits.Source: §3 · p. 4, §3.1 · p. 4

3.1 Counterfactual margins from history deletions

• For each protected request, the competitor is fixed at rank R + 1 of the initial decoded list, and the margin is the target score minus that competitor’s score.Reference and student use the same full-sequence SID log-probability scoring procedure.Source: §3.1 · p. 5
• A deletion qualifies only if it raises both target score and target–competitor margin; among qualifying views, AIMS selects the largest margin gain.If none qualifies, the complete-history margin is retained; selection runs once offline and leaves the protection set unchanged.Source: §3.1 · p. 5
• AIMS uses a single deletion and one competitor by default; larger deletion budgets or averaging over competitors changed Recall@10 only marginally, while maximizing over competitors degraded as the set grew.The larger deletion budget required up to 2.9× as many scoring views.Source: §3.1 · p. 6

3.2 Learning the counterfactual margin on the complete history

• On complete histories, cross-entropy supervises targets and ACS penalizes margin shortfalls below the cached target minus ε, routing its gradient only through the competitor score.The target score is detached; inference is unchanged, and ACS is active at initialization for requests with a positive margin gain greater than ε.Source: §3.2 · p. 6

4.3 Ablation Studies and Analysis

• The request-specific intervention gains and their assignment both matter: removing gains, shuffling them, or deriving them from generic requests or initial-margin difficulty weakens AIMS.ConstantGate and Magnitude Shuffle fall between R-comp and AIMS, while AIMS exceeds Shuffle on both backbones; random-item interventions yield gains about ten times smaller.Source: §4.3 · p. 7
• Margin-based deletion selection outperforms uplift-first and influence-first choices, which leave fewer protected requests with eligible deletions and lower Recall@10 and NDCG@10.AIMS leaves 88.2/85.7% of protected requests with an eligible deletion, versus 75.9/66.1% and 28.8/25.9% for those alternatives.Source: §4.3 · pp. 7–8

4.2 Main Results

• AIMS achieves the best Recall@10 and NDCG@10 for all six backbones on all three datasets, improving Recall@10 over Continued CE by 4.0–10.9% in relative terms.It also exceeded Continued CE and the strongest baseline in every repeated run with three training seeds.Source: §4.2 · p. 6, Table 1 · p. 7, Table 2 · p. 8

4.3 Ablation Studies and Analysis

• Using selected deletions as training inputs does not match using them as targets: deletion-view variants trail AIMS, including after initial hinge gaps are matched.Deletion CE does not beat Continued CE; the gap-matched control does not equalize later training dynamics.Source: Figure 3 · p. 8, §4.3 · p. 8
• Competitor-only auxiliary-gradient routing outperforms target-only and symmetric routing; target-only is on par with Continued CE, while symmetric routing falls below AIMS in Recall@10.The routing variants share cached targets and the forward hinge, differing only in where the auxiliary gradient flows.Source: §4.3 · p. 8, Table 3 · p. 9
• AIMS improves both retention and recovery over Continued CE, with higher recovery accounting for 68/85% of the gain; training without history or with history dropout does not reproduce its gains.History dropout trails AIMS by 0.176/0.378 Recall@10 points and gains less on conflict requests; AIMS exceeds it overall and in both groups.Source: §4.3 · pp. 8–10, Figure 4 · p. 9
• On conflict requests, Recall@10 gains over Continued CE are +0.558/+0.874 points, versus +0.321/+0.664 on the full test set; observed violation rates fall from 29.0/25.4% to 21.6/18.3%.Both effects grow with conflict severity; these comparisons are observational.Source: Figure 5 · p. 9, §4.3 · p. 9
• On protected requests with a satisfying target and mixed history, selected deletions remove a violating event for 80.9/83.8% of requests, above the 64.0/67.2% chance expectation.The excess shrinks under generic requests and reverses when the target violates the constraint, indicating selection depends on both request and target.Source: Figure 5 · p. 9, §4.3 · p. 9

5 Related Work

• Personalized search asks when to rely on a user’s history, and Lu et al. (2024) train LLM recommenders with category-based instructions and reinforcement learning.Source: §5 · p. 10
• Other approaches synthesize or correct histories, train retrieval filters on label-dependent utility, or use ranking, preference, and teacher-margin losses (Zhang et al., 2021; Lin et al., 2023, 2024; Izacard et al., 2023; Wang et al., 2023; Wang et al., 2024; Li et al., 2024; Chen et al., 2024; Deng et al., 2025; Hofstätter et al., 2020).Source: §5 · p. 10
• AIMS uses single-event deletions to set request-specific margin targets, trains on complete histories, and penalizes margin shortfalls through the competitor only.Source: §5 · p. 10

6 Conclusion

• The paper identifies influence–utility misalignment and shows that target-score gains alone need not improve rankings.Source: §6 · p. 10
• AIMS learns counterfactual margin targets from request-specific deletions on complete histories through an asymmetric auxiliary loss, leaving inference unchanged.Source: §6 · p. 10
• Across six backbones and three datasets, AIMS achieves the best Recall@10 and NDCG@10 in every setting and improves Recall@10 over matched Continued CE by 4.0–10.9% relative.The conclusion also reports larger gains under instruction–history conflict and says history-drop training does not reproduce them.Source: §6 · p. 10

A Protection-Set Statistics

• For the two ablation backbones, most protected requests have an eligible deletion and almost all eligible deletions activate the hinge at initialization; the median gain is about 0.25 nats.That gain multiplies the target-to-competitor likelihood ratio by about 1.3; larger gains prolong hinge activation but do not scale its gradient.Source: A Protection-Set Statistics · p. 15, Table 4 · p. 15

B.1 Training Pipeline

• All experiments use three SID tokens from level-specific vocabularies of 2,048 codes, with dataset-specific mappings fixed across methods and backbones.The 6,144 SID tokens are added to each backbone’s vocabulary.Source: §B.1 · p. 15
• Each dataset and backbone first trains to generate item SIDs, then fine-tunes with CE on request, history, and target instances; AIMS and baselines share the resulting SFT checkpoint and training setup.Source: §B.1 · p. 15

B.2 Datasets

• The industrial data come from anonymized production logs; Qilin uses its search subset, and KuaiSearch-Lite uses a released e-commerce search subset with interaction records for 102,086 users.The industrial split uses three training days and a later test day separated by one day; the public benchmarks use user-disjoint splits.Source: §B.2 · pp. 15–16

B.3 Evaluation Protocols

• On the industrial dataset, methods retrieve 100 unique valid SIDs from the full catalog, with Recall@k and NDCG@k computed from the top k.Recall@k measures the fraction of test requests with the target in the top k; NDCG@k uses binary relevance with one target per request.Source: §B.3 · p. 16
• On public benchmarks, requests’ candidate pools are sorted by full-sequence SID log-probability; sampled-pool metrics compare methods on identical pools, not across datasets.All clicked items are relevant, and NDCG@k uses binary gains.Source: §B.3 · p. 16
• Pools with fewer than 40 items are extended using training-split impressions, with logged candidates excluded and the same pools used for every method.Sampling is seeded by request identifier and does not use click labels.Source: §B.3 · p. 16
• For candidate-pool AIMS training, an instance is protected when its target ranks in the top ten under θ0; the eleventh candidate becomes its competitor.Other clicked items from the same request are removed when ranking the initial pool; inference is unchanged.Source: §B.3 · p. 16
• Significance is assessed with paired user-cluster bootstrap resampling, calling differences significant when the two-sided 95% interval excludes zero.The protocol uses 10,000 resamples and keeps each sampled user’s test requests together.Source: §B.3 · p. 16

B.4 Hyperparameters

• All baselines share AIMS’s initialization, SIDs, positive instances, schedule, and evaluation protocol; fixed-in-advance settings include one deletion, one competitor, R = 10, and beam size 100.Hyperparameters are tuned on the same validation split, and test splits are never used for selection; Table 5 lists the values.Source: §B.3 · p. 16, §B.4 · p. 16, Table 5 · p. 17

B.5 Ablation variants

• Difficulty-based controls replace AIMS’s request–gain assignment with difficulty-based targets while retaining its protection set, competitors, and training objective.D-cal preserves the gain distribution and initial hinge-loss distribution but may change which requests begin active.Source: §B.5 · p. 17
• Difficulty-matched controls use empirical midrank percentiles to assign larger gains to requests with smaller initial margins, with the scale selected by validation Recall@10.The percentile and gain construction uses neither deleted-history scores nor AIMS gains.Source: §B.5 · p. 17
• Generic-matched gains are computed using a request that omits specific intent, then reassigned while preserving the AIMS gain distribution and initial hinge losses.Training still uses the factual request and complete history; these synthetic targets need not equal margins from generic deletion views.Source: §B.5 · p. 17
• Uplift-first and Influence-first change only which deletion defines the gain; each assigns zero when its selected deletion is ineligible.AIMS instead uses the largest margin gain among eligible deletions, so these rules’ per-request gains cannot exceed AIMS’s.Source: §B.5 · p. 17

C.1 All cutoffs on the public benchmarks

• Tables 6 and 7 report public-benchmark performance at all cutoffs under candidate-pool ranking, with Recall and NDCG metrics.The supplied captions identify Qilin and KuaiSearch-Lite, respectively, but do not state table cell values.Source: §C.1 · p. 17, Table 6 · p. 18, Table 7 · p. 19

C.2 Training seeds

• Across three seeds, AIMS is higher in Recall@10 than each backbone’s strongest baseline in every comparison.The table reports means, standard deviations, and paired per-seed differences; cached AIMS competitors and margins are identical across seeds.Source: §C.2 · p. 18

C.3 Deletion budget and competitor set

• Larger deletion budgets yield small, backbone-dependent changes while using up to 2.9× as many scoring views; Pair-mean changes are small and inconsistent, and Set-max degrades as M grows.The sensitivity analysis varies deletion budget and competitor set around K = 1 and M = 1.Source: §C.3 · p. 18

C.4 History use

• History-use controls compare request-only CE, 0.5-probability history dropout, and replacement with matched histories from other users; inference uses complete history.The history-swap replacements match request category, history length, and coarse category composition, and results are averaged across three replacements.Source: §C.4 · p. 19

C.5 Placebo interventions and selection rules

• Placebo insertions produce fewer eligible views and much smaller positive gains than deletions, but the comparison does not isolate why the perturbations differ.Eligible views occur for 66.4%/69.7% of protected requests with placebo versus 88.2%/85.7% with deletion; median positive gains are 0.018/0.023 versus 0.234/0.256 nats (Llama/Gemma).Source: §C.5 · pp. 19–20, Table 11 · p. 21
• Uplift-first and Influence-first leave fewer deletions eligible than AIMS, and both have smaller median positive gains.Influence-first is eligible for only 28.8%/25.9% of protected requests, versus 88.2%/85.7% for AIMS (Llama/Gemma).Source: §C.5 · p. 20, Table 11 · p. 21

C.6 Retention–recovery decomposition

• Higher recovery accounts for 68%/85% of the Recall@10 gain over Continued CE on the industrial test set for Llama-3.1-8B/Gemma-3-12B.The decomposition uses a test split defined by θ0’s top-10 decisions.Source: §C.6 · p. 21

C.7 Signed label-free attribution

• When θ0’s top-1 prediction differs from the target, model-prediction attribution is negatively correlated with signed target utility, so neither attribution reliably identifies target-supporting events.When the top-1 prediction equals the target, the attribution equals signed target utility.Source: §C.7 · p. 21, Table 12 · p. 22

Constraint and history labels

• History labels apply to positive events preceding the request: pass satisfies every explicit constraint, fail demonstrably violates at least one, and unknown lacks sufficient attributes to verify all.Item categories come from frozen platform metadata, and evaluation labels are shared across methods and backbones.Source: D Instruction–History Conflict… · p. 22

History-conflict severity

• Severity analysis stratifies eligible requests into four κ_i bands and excludes the 3.8% that fail the evaluable-history coverage condition.The bands are κ_i = 0, 0 < κ_i < 0.5, 0.5 ≤ κ_i < 0.8, and 0.8 ≤ κ_i ≤ 1.Source: D Instruction–History Conflict… · p. 22

Output labels and metrics

• The protocol materializes ten ranks per request, labels each pass, fail, or unknown, and keeps unknown and invalid slots in the fixed denominator without counting them as satisfying constraints.A rank is invalid if missing, duplicated, or unmappable to the frozen candidate and attribute snapshot.Source: D Instruction–History Conflict… · p. 22

Uncertainty

• Confidence limits use 10,000 paired user-cluster bootstrap resamples, keeping each sampled user’s requests together and evaluating every method on identical resamples.Reported limits are empirical 2.5th and 97.5th percentiles, not centered or symmetrized around the estimate.Source: D Instruction–History Conflict… · p. 22

Severity trend

• Across κ_i strata, Continued CE FAIL and AIMS violation reductions rose monotonically; Recall@10 gains were non-decreasing, but these observational trends do not identify causal effects.From κ_i = 0 to κ_i ≥0.8, absolute reductions grew from 0.6 to 8.6 points on Llama and 0.5 to 8.4 points on Gemma.Source: D Instruction–History Conflict… · p. 23, Table 13 · p. 23

Labels of selected deletions

• Table 14 shows selected deletions removed fail events 16.8/16.6 points more often than expected; the excess fell to 2.9/4.1 for generic requests and reversed to −12.3/−11.1 when targets failed.Deletion labels describe the selected events but do not enter AIMS.Source: D Instruction–History Conflict… · p. 23

Request–history conflict

• Prior context-control work studies architecture, decoding, or training data; AIMS uses real user queries and observed click targets to evaluate individual history events offline, leaving inference unchanged.Lu et al. (2024) use synthesized control instructions and reinforcement learning, so the paper treats that pipeline as complementary rather than directly comparable.Source: E Extended Related Work · p. 24

Label-dependent context utility as supervision

• Unlike RAG filters trained on label-dependent context utility, AIMS turns history-deletion effects into recommender margin targets and keeps complete history at training and inference.Training on the same deletion views is less effective; an inference-time filter would need to predict target-dependent effects without the target.Source: E Extended Related Work · pp. 24–25

History synthesis, selection, and correction

• History transformation methods synthesize, correct, or select sequences; AIMS instead evaluates evidence for the current target–competitor comparison, and an accepted deletion need not mark an erroneous interaction.A recorded behavior can be valid yet contribute adversely under the present instruction.Source: E Extended Related Work · p. 25

Context and output debiasing in LLM recommenders

• CFT uses a label-directed full-history versus history-absent logit objective, whereas AIMS builds an individual-request margin; AIMS also keeps its initially decoded competitor fixed during training.At decoding, AIMS retains the original SID-constrained deployment decoder.Source: E Extended Related Work · p. 25

Ranking objectives and asymmetric negative supervision

• Unlike ranking and preference objectives with general thresholds or candidate losses, AIMS derives its comparison threshold from reference margins strengthened by eligible history interventions and gates the penalty by margin shortfall.LETTER (RG) applies a ranking-guided generation loss, while LTRGR adds a rank-loss phase over retrieved candidates.Source: E Extended Related Work · p. 25

Transferring supervision and preserving recommendations

• AIMS transfers a scalar comparison like score-difference distillation, but eligible history deletions strengthen it and the student learns a minimum desired margin with tolerance rather than an equality target.Margin-MSE trains rankers to match cached teacher score differences between document pairs.Source: E Extended Related Work · p. 26

Preserving recommendations

• AIMS constrains selected target–competitor comparisons rather than reproducing a reference distribution; R-comp improves Recall@10 over Continued CE by +0.093/+0.336.The paper reports request-specific gains account for the remaining 71%/49% of AIMS improvement.Source: E Extended Related Work · p. 26

Attribution and counterfactual explanations

• Unlike attribution methods tracing retraining effects, AIMS evaluates input deletions with a frozen model and treats utility as a predictive contrast, not an identified causal effect of user behavior.The paper distinguishes its target-free distributional sensitivity measure from signed influence-function quantities.Source: E Extended Related Work · p. 26