RecLoop study: Generative recommenders less prone to information cocoons

A new paper by Jiyuan Yang and colleagues introduces RecLoop, a simulation framework that uses LLM-powered user agents to study information cocoons in generative recommendation systems. Unlike traditional recommenders that rely on item IDs, generative recommenders use Semantic ID (SID) sequences, which fundamentally change how recommendations are generated. The authors compare two generative and two traditional sequential recommenders over multiple feedback cycles on two Amazon datasets, using both standard exposure-level metrics and a novel model-level metric called Code-Space Structural Cocoon. This metric measures concentration in the generated SID space.

Results show that generative recommenders are generally less prone to exposure-level cocoon formation, preserving broader diversity and slowing cross-user homogenization. However, feedback loops still induce concentration within the SID space. Crucially, cocoon severity depends heavily on tokenization strategy: collaborative-signal tokenization produces stronger cocoon effects than semantic tokenization. Larger models maintain greater code-space diversity and better retain access to niche content. These findings suggest that although generative recommenders may mitigate some traditional echo-chamber effects, the design choices around tokenization and model capacity remain critical for ensuring diverse recommendations.