Echoes in Filter Bubble: Diagnosing and Curing Popularity Bias in Generative Recommenders

Researchers have identified and addressed popularity bias in Generative Recommenders (GRs), a emerging class of AI systems that use unified end-to-end frameworks for recommendations. The study reveals that this bias stems from token-level optimization flaws and undifferentiated item tokenization, proposing Ghost, a novel system using asymmetric unlikelihood optimization and skeleton-founded tokenization to mitigate the problem while maintaining recommendation quality.

Generative Recommenders represent a significant shift in how recommendation systems operate, moving away from traditional approaches toward more unified, end-to-end frameworks powered by generative models. However, this research exposes a critical vulnerability: GRs perpetuate popularity bias, where the system disproportionately recommends popular items while marginalizing niche content. This phenomenon creates filter bubbles that limit user discovery and skew platform economics.

The technical root cause proves more nuanced than previous studies suggested. Rather than a singular flaw, popularity bias emerges from the interaction between how GRs optimize their token-level objectives and how they tokenize items using semantic indices. Traditional debiasing methods, borrowed from classical recommenders, fail to address this specific architectural vulnerability in generative systems.

The Ghost framework tackles this through two innovations: asymmetric unlikelihood optimization, which penalizes popular items differently during training, and skeleton-founded tokenization, which ensures item representations maintain differentiation regardless of popularity distribution. Empirical testing across three datasets demonstrates Ghost substantially reduces popularity bias while incurring only slight degradation to overall recommendation accuracy.

For the broader AI ecosystem, this work matters because recommendation systems underpin content discovery across major platforms. As AI adoption accelerates, ensuring fair algorithmic distribution becomes increasingly critical for user satisfaction, creator opportunity, and platform legitimacy. The research validates that architectural choices in generative models directly impact fairness properties, suggesting future GR designs must incorporate debiasing considerations from inception rather than treating them as afterthoughts.

• →Popularity bias in Generative Recommenders stems from token-level optimization flaws combined with undifferentiated item tokenization, not just data distribution.
• →Ghost's asymmetric unlikelihood optimization and skeleton-founded tokenization substantially reduce popularity bias across multiple datasets.
• →Traditional debiasing methods prove ineffective for GRs, requiring architecture-specific solutions to address fairness issues.
• →The research demonstrates a tradeoff between fairness and overall recommendation utility, with Ghost achieving significant bias reduction at minimal accuracy cost.
• →This work highlights that generative AI systems require fairness considerations embedded into design, not applied retrospectively.