Organizations: Department of Artificial Intelligence and Data Engineering, Istanbul Technical University, Istanbul, Türkiye · Artificial Intelligence and Data Science Application and Research Center, Istanbul Technical University
Large language models (LLMs) have introduced a generative approach to recommendation, producing item lists from textual representations of user histories and item information. However, this generative setting raises challenges concerning output validity, prompt consistency, and sensitivity to demographic descriptors. In this work, we investigate how preprocessing choices in fine-tuning data affect recommendation utility and fairness-related behavior in a generative LLM-based recommender system. We first establish a controlled recommendation pipeline that combines supervised fine-tuning, catalog-constrained decoding, and prompt augmentation. We then study three preprocessing choices: item-title representation, demographic balancing, and auxiliary item-knowledge enrichment. Experiments on MovieLens-1M show that the catalog-constrained pipeline eliminates out-of-catalog items and ensures complete ten-item outputs, while prompt augmentation improves prompt consistency and utility. Original MovieLens titles increase HR@10 from 10.67% to 12.47% and Coverage@10 from 41.36% to 47.88%. Fairness-related effects are attribute-dependent: age balancing and knowledge enrichment reduce selected group-level utility gaps while largely preserving utility, whereas gender balancing reduces utility without improving the evaluated gender-related indicators. Overall, preprocessing choices in fine-tuning data should be evaluated jointly in terms of utility, output behavior, and demographic sensitivity.
Figures & tables
Figure 1. Overview of the controlled experimental pipeline. A left-to-right overview of the data preparation, preprocessing choices, fine-tuning, catalog-constrained recommendation generation, and controlled evaluation pipeline.
Figure 2. Prefix-tree constrained decoding within an item span. After the model emits <SOI> , only tokens that continue a valid catalog path remain available. Invalid continuations are masked, and <EOI> becomes available only after a complete catalog-item sequence has been generated. A prefix tree rooted at the start-of-item token. Two allowed branches lead to valid catalog-item sequences and end-of-item nodes, while one invalid branch is masked.
Family
Character
Templates
Share
T1
Direct request
8
30%
T2
Explicit rule list
7
25%
T3
Preference framing
7
20%
T4
Output format emphasis
6
15%
T5
Compressed phrasing
5
7%
T6
Misspelled instructions
4
3%
Table 1. Prompt template families used during fine-tuning. Shares give the proportion of recommendation examples assigned to each family.
Model
HR@10 ↑
Exact-10 ↑
OOD@10 ↓
Repeat@10 ↓
Prompt HR Spread ↓
Prompt Jaccard ↑
SFT (unconstrained decoding)
7.47
0.06
2.08
3.12
–
–
SFT (constrained decoding)
7.43
0.02
0.00
4.86
–
–
Catalog-constrained
9.89
100.00
0.00
10.56
2.18
0.6799
Prompt-augmented
10.67
100.00
0.00
8.01
0.26
0.7489
Table 2. Pipeline reliability, recommendation quality, and prompt consistency. Arrows give the preferred direction.
Model
HR@10 ↑
Coverage @10 ↑
Repeat @10 ↓
Unique Items ↑
Normalized titles
10.67
41.36
8.01
1,606
Original titles
12.47
47.88
8.43
1,859
Table 3. Effect of item-title representation on recommendation utility and catalog usage. Arrows give the preferred direction.
Model
HR@10 ↑
Neutral Gap ↓
Neutral-cond. Jaccard ↑
SNSR ↓
Pair Jaccard ↑
Gender
Baseline
12.47
0.45
0.8312
0.0519
0.8110
+ Gender-balanced
11.32
1.21
0.8240
0.0636
0.7918
Age
Baseline
12.47
0.69
0.8195
0.0017
0.8407
+ Age-balanced
12.39
0.38
0.8305
0.0077
0.8471
Table 4. Effect of gender- and age-balanced fine-tuning on recommendation utility and demographic sensitivity under the original-title representation. Arrows give the preferred direction.
Model
HR@10 ↑
Neutral Gap ↓
Neutral-cond. Jaccard ↑
SNSR ↓
Gender
Baseline
12.47
0.45
0.8312
0.0519
+ Knowledge
12.24
0.10
0.8270
0.0291
Age
Baseline
12.47
0.69
0.8195
0.0017
+ Knowledge
12.24
0.32
0.8201
0.0123
Table 5. Effect of knowledge enrichment on recommendation utility and demographic sensitivity. Arrows give the preferred direction.
Large language model-based recommendation uses language models to personalize ranking, retrieval, and conversational suggestions from text and behavior.
Large language model (LLM) recommenders can produce semantically different recommendations when a user's protected attribute changes, even if the interaction history remains fixed. Controlling this disparity is challenging when model weights and full-vocabulary logits are inaccessible. We propose FAIDER (Fairness-Aware Item-level DEcoding for Recommendation), which incorporates fairness into list construction by selecting complete item candidates rather than manipulating token-level logits or repeatedly repairing finished outputs. The central idea is to separate the operational criterion for a violation from how strongly generation is constrained. Conformal-inspired empirical calibration fixes a cross-group semantic-disparity threshold, while online decoding applies either hard filtering or a soft penalty adjusted through dual-inspired feedback to target a prescribed violation rate. Separately, we establish a pathwise guarantee: under a uniform enforceability assumption, the soft decoder's empirical violation rate exceeds its target by at most O(1/T). Experiments across recommendation benchmarks and LLM backbones show that Soft FAIDER improves both measured semantic fairness and ranking quality over the evaluated prompt-repair baselines. It also generally achieves a more favorable fairness-utility balance than the hard variant, indicating that stricter per-recommendation constraints do not necessarily yield lower measured disparity.
Large Language Models (LLMs) have introduced new capabilities to recommender systems, enabling dynamic, context-aware, and conversational recommendations. However, LLM-based recommender systems inherit and may amplify social biases embedded in their pre-training data, especially when demographic cues are present. Existing fairness solutions either require extra parameters fine-tuning, or suffer from optimization instability. We propose a lightweight and scalable bias mitigation method that combines a kernelized Iterative Null-space Projection (INLP) with a gated Mixture-of-Experts (MoE) adapter. Our approach estimates a closed-form projection that removes single or multiple sensitive attributes from LLM representations with no additional trainable parameters. To preserve task utility, we introduce a two-level MoE adapter that selectively restores useful signals without reintroducing bias. Experiments on two public datasets show that our method reduces attribute leakage across multiple protected variables while maintaining competitive recommendation accuracy.
Nan Cui, Wendy Hui Wang, Yue Ning
Department of Computer Science, Stevens Institute of Technology, Hoboken NJ, USA
Fairness audits for LLM-based recommenders have largely focused on observable outputs, implicitly assuming that stable recommendations reflect stable internal processing. We challenge this assumption with FairGap, the first benchmark to jointly evaluate recommendation fairness at two levels: observable output shift (OBS) and hidden representation shift (IBS), measured through controlled counterfactual identity probes across gender, age, and race. Their relationship is summarized via Representation-Output Alignment (ROA), with quadrant diagnostics for identifying user-level hidden-output mismatch. Applied to six open-weight LLM families across three domains, FairGap reveals pervasive hidden-output decoupling: ROA rarely exceeds 0.22, and a non-negligible user population shows stable outputs despite substantial internal shifts, a mode that output-only audits cannot detect by design. Further, activation steering that reduces IBS by up to 8x simultaneously worsens OBS, demonstrating a fundamental tension between internal and output-level fairness that existing frameworks are unequipped to diagnose.
Chan Aristella Lu, Arya Fayyazi, Junhao Zhang +6
University of Georgia · University of Southern California · Carnegie Mellon University +1