TRACER: Token ReAssignment for Concept ERasure in Generative Recommendation

Researchers introduce TRACER, a novel framework for removing sensitive concepts from generative recommendation systems while preserving overall utility. The method uses token reassignment to handle the unique challenge that semantic IDs in recommendation systems are shared across items to forget and retain, unlike discrete tokens in language models.

Generative recommendation systems operate similarly to large language models by predicting next items through autoregressive sequences of semantic IDs derived from user behavior. As these systems gain prominence, regulatory pressures and safety concerns demand the ability to unlearn harmful or sensitive concepts—yet existing LLM unlearning techniques fail when applied directly to recommendation contexts. The core problem stems from structural differences: while LLMs use semantically explicit word tokens, recommendation systems employ abstract semantic IDs that naturally overlap between items requiring removal and those requiring preservation, creating inherent conflicts.

TRACER addresses this by avoiding direct suppression of shared tokens and instead reassigning concept-related items to alternative tokens optimized for forgetting. A coherence regularizer maintains semantic consistency among retained items during the unlearning process. This architectural innovation reflects growing maturation in AI safety research, moving beyond one-size-fits-all solutions toward domain-specific unlearning methods.

For the AI industry, TRACER demonstrates that recommendation systems require specialized approaches as they scale to LLM-scale complexity. The framework's success in balancing concept removal with utility preservation suggests that future commercial recommendation engines will incorporate similar unlearning mechanisms to address privacy regulations and safety standards. This advancement enables platforms to comply with data protection requirements without sacrificing recommendation quality—a critical capability as jurisdictions worldwide implement stricter AI governance frameworks.

• →TRACER uses token reassignment rather than suppression to remove sensitive concepts from recommendation systems while preserving functionality.
• →Semantic IDs in recommendation systems create unique unlearning challenges that differ fundamentally from LLM token structures.
• →The framework incorporates a coherence regularizer to maintain semantic consistency across retained items during unlearning.
• →Experimental results show TRACER substantially outperforms existing unlearning baselines in preserving recommendation quality.
• →This research addresses growing regulatory and safety demands for concept unlearning in production recommendation systems.