PaperlayerView exact revision on arXiv

Source-linked AI summary

Measuring and Mitigating Identity-Cue Preference Drift in LLM-based Recommender Systems

Zhuoxiong Gan, Qiang Dong

arXiv:2609.34229v1cs.IRcs.SI

TL;DR

Factual identity cues can change LLM recommendations even when a user’s interaction history is unchanged, yet prior evaluations do not isolate this effect. PromptShift measures and mitigates the resulting drift without model training. It reduces macro-mean SliceShift by 62.42% while improving several utility metrics, with a slight NDCG@20 decrease.

  • Problem

    Prior evaluations do not isolate whether a factual demographic cue added to an otherwise unchanged prompt shifts recommendations toward the corresponding group’s item distribution.

    Source: §1 · p. 1

  • Method

    PromptShift uses controlled prompts, a shared identity-slice-by-item table, Drift and SliceShift measures, and a user-adaptive post-hoc reranker.

    Source: §1 · p. 2, §4.1 · p. 4

  • Results

    62.42% lower macro-mean SliceShift; identity-cued prompts also have higher mean Drift than paraphrase controls, with positive SliceShift across all six dataset–model settings.

    Source: §1 · p. 2, §7 · p. 10

  • Takeaways & Limitations

    Across the evaluated settings, identity-cue drift can be measured and mitigated without model training, with a modest, metric-dependent utility cost.

    Source: §1 · p. 2, §7 · p. 10

    The reranker only operates within the generated candidate pool and cannot recover relevant items that are absent from it.

    Source: §6 · p. 10

Abstract

from arXiv · show

In large language model-based recommender systems, identity cues embedded in prompts can steer recommendations toward group-level patterns even when the underlying behavioral evidence remains unchanged. We introduce PromptShift, an interpretable, training-free framework for quantifying and mitigating such identity-cue preference drift. We define Drift as the divergence, in both item membership and ranking order, between a recommendation list generated under an identity-cued prompt and the reference list produced from the same user's interaction history alone. SliceShift then measures the extent to which a cued list gravitates, relative to the history-only reference, toward items that are more popular within the cued slice than among the global user population. Beyond conventional accuracy, we propose DifHitRate, a difficulty-weighted hit metric that credits only relevant items, assigning higher credit to hits that are less popular within the cued slice and ranked higher in the list. All components are supported by an identity-slice-by-item table constructed from positive interactions, which further enables an adaptive post-hoc reranking strategy: the reranker interpolates between the original LLM ranking and inverse slice-popularity, with personalized interpolation weight. Experiments on two datasets with three LLMs show that identity-cued prompts incur higher mean Drift than identity-free paraphrase controls, an effect beyond generic wording sensitivity, and that SliceShift is positive across all six dataset-model settings. PromptShift consistently reduces both Drift and SliceShift, lowering macro-mean SliceShift by 62.42%, while improving DifHitRate, HitRate and MRR. These results demonstrate that identity-cue preference drift can be measured and mitigated without any model training, albeit with a modest, metric-dependent utility cost.

1. Introduction

The introduction asks whether factual identity cues change recommendations when behavioral history is fixed, and whether changes favor items popular in the cued group. PromptShift measures and mitigates this drift; experiments find it beyond wording sensitivity and reduce it with a modest, metric-dependent utility cost.

1. Introduction

  • Prompt sensitivity makes attribution difficult, while a changed list alone does not show whether recommendations move toward group-typical items or merely reshuffle content.Meaning-preserving rewording can change outputs; demographic cues state a fact rather than an explicit recommendation instruction.Source: §1 · p. 1
  • Prior evaluations do not isolate whether adding one factual demographic attribute to an unchanged prompt shifts recommendations toward the corresponding group’s item distribution.The paper calls a systematic cue-associated list change unsupported by a change in user history identity-cue preference drift.Source: §1 · p. 1
  • PromptShift uses paraphrase controls, a positive-interaction identity-slice-by-item table, Drift and SliceShift measures, and a user-adaptive reranker interpolating the LLM ranking with inverse slice-popularity.Its proposed difficulty-weighted hit metric gives more credit to relevant items that are less popular within the cued slice and ranked higher.Source: §1 · pp. 2–3
  • PromptShift reduces Drift and SliceShift, lowering macro-mean SliceShift by 62.42%, while improving Difficulty@20, HitRate@20, and MRR@20 at the cost of a slight decrease in NDCG@20.Source: §1 · p. 2

2. Related work

Related work covers LLM recommendation, recommender-system bias, and prompt sensitivity alongside post-hoc reranking. PromptShift addresses the remaining question of whether a factual identity cue itself shifts an otherwise unchanged recommendation prompt.

2.1. LLM-based recommendation

  • LLM-recommender research spans zero-shot ranking, interactive and explainable systems [2], prompting strategies, and near-cold-start recommendation, with surveys consolidating the field [5].Prior evaluation also shows that utility metrics alone do not fully characterize these systems.Source: §2.1 · p. 3

2.2. Bias and fairness in recommender systems

  • Prior bias research documents popularity, exposure, and selection biases, demographic effects, feedback-loop amplification, and popularity bias in zero-shot LLM recommenders.These studies do not isolate whether one factual identity cue shifts recommendations toward the cued group’s item distribution.Source: §2.2 · p. 3

2.3. Prompt sensitivity, controllability, and post-hoc reranking

  • Because outputs vary with demonstration order and meaning-preserving formatting, cue effects require wording controls; unlike deliberate steering [8], identity cues implicitly condition the model.PromptShift applies post-hoc reranking, adapting intervention strength to within-slice history mainstreamness.Source: §2.3 · p. 3

3. Preliminaries and problem statement

The paper defines users, histories, demographic identity slices, and controlled prompt conditions for studying identity-cue effects. It frames drift measurement and mitigation as paired problems under fixed model and prompt settings.

3.1. Notation

  • Demographic values partition users into identity slices, and g=⋆ denotes the all-user reference; each user has an ordered interaction history.The prompt conditions are history-only, identity-conditioned, and paraphrase-control prompts.Source: §3.1 · p. 3

3.2. Problem statement

  • The paired prompt design holds model version, message, history, decoding, list length, and output format fixed so the factual identity cue is the only input difference.The LLM is treated as a black box, and the cue states a fact rather than an instruction.Source: §3.2 · pp. 3–4
  • Identity-cue preference drift is a systematic change from the history-only list to the identity-conditioned list that exceeds changes from meaning-preserving rewording and is attributable to the cue alone.Paraphrase controls estimate ordinary wording sensitivity under the paired design.Source: §3.2 · pp. 3–4
  • The measurement problem is to quantify both list-change magnitude and whether the identity-conditioned list moves toward items disproportionately popular in the cued slice.This distinguishes the direction of change from a mere reshuffling of items.Source: §3.2 · p. 4
  • The mitigation problem is to rerank the identity-conditioned candidate pool toward zero drift while preserving utility against held-out positives.The intervention is constrained to prompt-side or generated-list operations because model parameters, logits, and training data are inaccessible.Source: §3.2 · p. 4

4. The proposed PromptShift framework

PromptShift reuses one identity-slice-by-item table across four stages to measure cue-associated changes and determine the strength of post-hoc correction.

4.1. Framework overview

  • The shared slice table gives cue-associated list changes a data-grounded direction and determines how strongly the reranker should intervene.At the item level it characterizes slice association; at the user level it sets correction strength.Source: §4.1 · p. 4

4.2. Identity-slice-by-item popularity table

  • The table maps recorded identity values to slices and uses positive interactions to connect a cue with items common in that slice.Distinct-user counts and shrinkage toward the all-user column address highly active users and small-slice estimates.Source: §4.2 · pp. 4–5

4.3. Measuring drift: magnitude and direction

  • Drift quantifies recommendation-list change using rank-biased overlap, while SliceShift measures whether cue-associated changes move toward slice-overrepresented items.A positive SliceShift indicates movement toward items more popular in the active slice than among all users; Drift and SliceShift capture complementary properties.Source: §4.3 · p. 5
  • Three frozen, meaning-preserving paraphrase controls provide a wording-sensitivity baseline for distinguishing cue-associated change from ordinary prompt variation.Cue Drift is compared with Paraphrase Drift under identical decoding settings.Source: §4.3 · p. 6

4.4. Within-slice history mainstreamness

  • Within-slice history mainstreamness adapts reranking strength because users sharing an identity value can have different histories and preferences.The score compares a user's history with the scenario mean to guide trust in the original order versus less prevalent candidates.Source: §4.4 · p. 6

4.5. Adaptive post-hoc reranking

  • Generating more than K candidates lets reranking promote relevant items that began below the original top K.The final list is selected from the larger candidate pool rather than only rearranging the original top K.Source: §4.5 · p. 6
  • The reranker balances the LLM's original ranking against inverse slice-popularity, with the user's mainstreamness determining their relative weight.The original rank serves as a proxy for personal relevance, while inverse slice-popularity gives less prevalent candidates a signal.Source: §4.5 · p. 6
  • Geometric retention scores favor early LLM positions while preserving a nonzero floor for candidates ranked later.The reranker also protects an initial prefix and sorts the remaining candidates by the final score.Source: §4.5 · p. 6

5. Experiments

Across two datasets and three LLMs, the experiments test whether factual identity cues shift recommendations beyond ordinary wording sensitivity and whether post-hoc reranking reduces that shift. PromptShift reduces cue-associated movement while producing metric-dependent utility changes.

5.1. Datasets

  • MovieLens 1M supplies explicit ratings and Last.fm 1K implicit listening events; users’ interactions are split chronologically into fit, validation, and test data.The datasets contain 1,000,209 ratings from 6,040 users and nearly 1,000 profiles with approximately 19 million listening events; eligible users are sampled with a frozen scenario-stratified seed.Source: §5.1 · p. 7

5.2. Prompt templates and controls

  • The prompt comparison adds one factual identity sentence to the history-only prompt and uses three identity-free paraphrases to measure ordinary wording sensitivity.Cue values are drawn from recorded user attributes, and the cue sentence makes no claim about group preferences or instruction to emphasize identity.Source: §5.1 · p. 7, §5.2 · p. 7

5.3. Baselines

  • B0 keeps the identity-conditioned list’s first K items; B1 and B2 use fixed-weight all-user and slice-conditioned popularity, while B3 adds adaptive weighting and a protected prefix.The baseline progression compares all-user versus slice-conditioned popularity and the complete adaptive design.Source: §5.3 · pp. 7–8

5.4. Evaluation metrics

  • The evaluation reports Drift and SliceShift against the history-only reference, a difficulty-weighted hit score, and HitRate@20, MRR@20, and NDCG@20 for utility.Reranking comparisons reuse the same users and candidate pool; Table 6 reports unweighted macro means across six dataset–model settings.Source: §5.4 · p. 8

5.5. RQ1: Effects of identity cues on recommendation lists

  • Cue Drift exceeded Paraphrase Drift in all six averages (Δd 0.0270–0.0759), while SliceShift was positive in every average (0.0550–0.0680).Paraphrase Drift itself ranged from 0.4520 to 0.6110, underscoring the need to compare against ordinary wording sensitivity.Source: §5.5 · p. 8, Table 4 · p. 9

5.6. RQ2: Effects of reranking

  • B1’s all-user popularity penalty raised SliceShift to 0.0650 and reduced HitRate@20 to 0.7421; slice-conditioned B2 lowered SliceShift to 0.0480, and adaptive B3 to 0.0230.B3 also achieved the highest HitRate@20 and Difficulty@20; the results identify slice-conditioned popularity as the effective signal, with adaptive weighting adding improvement.Source: §5.6 · p. 8
  • Across six settings, B3 reduced macro-mean SliceShift from 0.0612 to 0.0230 and raised Difficulty@20 from 21.2200 to 25.9600 versus B0; Drift fell from 0.5827 to 0.5423.The reranker reduces cue-associated movement without simply reverting to the history-only list.Source: §5.6 · p. 8, Table 5 · p. 9
  • B3 raised macro-mean HitRate@20 from 0.8423 to 0.8845 and MRR@20 from 0.4775 to 0.5032, while NDCG@20 fell from 0.4957 to 0.4866.The authors associate this pattern with surfacing lower slice-popularity hits, which can disturb multi-hit orderings.Source: §5.6 · p. 9

6. Discussion and limitations

The discussion concludes that identity cues redirect recommendations toward group-level consumption patterns beyond ordinary wording sensitivity, and that post-hoc correction can reduce this drift. The evaluation is limited to two datasets, coarse recorded identity labels, and a frozen prompting configuration; reranking cannot recover relevant items absent from the candidate pool.

6. Discussion and limitations

  • The shared slice table grounds both drift measurement and post-hoc correction, which uses slice-conditioned popularity and per-user strength based on within-slice history mainstreamness.The reported correction reduces cue-associated movement while preserving—and in most cases improving—conventional utility metrics.Source: §6 · p. 9
  • The framework requires no model training or access to model internals, but adds paraphrase list generations, a pass over historical interactions, and an O(NlogN) per-user reranking sort.The authors describe slice-conditioned auditing as a plausible complement to outcome-disparity metrics in deployed LLM recommenders.Source: §6 · p. 9
  • The study covers two datasets, coarse dataset-recorded identity labels, and one frozen prompting configuration; reranking cannot recover relevant items missing from the generated candidate pool.SliceShift reflects historical interaction frequencies, not individual intrinsic preferences or inherent group properties.Source: §6 · p. 10

7. Conclusion and future work

Across two datasets, three LLMs, and three cue scenarios, identity cues shifted recommendation lists toward cued-slice patterns beyond wording sensitivity; reranking reduced SliceShift with improvements on several ranking metrics. The paper concludes that slice-conditioned auditing complements outcome-disparity metrics and identifies future transfer tests.

7. Conclusion and future work

  • 62.42% lower macro-mean SliceShift accompanied improvements in Difficulty@20, HitRate@20, and MRR@20, with a modest NDCG@20 cost.The reported pattern spans two datasets, three LLMs, and three cue scenarios.Source: §7 · p. 10
  • PromptShift measures identity-cue drift with Drift and SliceShift and mitigates it using a user-adaptive reranker driven by a shared identity-slice-by-item representation.The framework is training-free.Source: §7 · p. 10
  • The findings support slice-conditioned auditing as a practical complement to outcome-disparity metrics; future work will test transfer across broader benchmarks and interventions.Proposed directions include privacy-preserving benchmarks, broader candidate pools, and generator-level interventions.Source: §7 · p. 10

CRediT authorship contribution statement

Zhuoxiong Gan is credited with conceptualization, methodology, software, investigation, and the original draft; Qiang Dong is credited with supervision, review and editing, and funding acquisition.

CRediT authorship contribution statement

Related papers

Loading 2609.34229v1…