Self-Evolving Recommender Systems: A Survey and an Evaluation Audit
Year: 2026
Type: preprint
Abstract: In 2026 a new kind of recommender research appeared: large language model (LLM) agents that propose changes to a recommender's configuration or source code, train the changed model, and keep what the evaluation says is better. Within nine months about thirty such systems were reported, from open-source research harnesses to production loops at Google, Meta, Amazon, Alibaba, Kuaishou, Tencent and NetEase. We organise them, together with search-space AutoML and the agentic recommenders that update their own memory or weights, by what the loop changes, who proposes the change, and which chain of signals decides whether it survives, and we synthesise their recurring designs and the changes they keep. Because a self-evolving system is only as good as the signal it selects on, we then audit how 27 configuration-, code- and component-level loops are evaluated, coding each paper's full text against eight criteria with released evidence. Of the 16 loops that select candidates offline, three of the ten that report ranking metrics rank the full catalogue, eight do not say which data keep or reject candidates, four select on the data they report, and two state that the test set was used once; none reports its headline result with the spread over repeated evolution sessions, and three compare against a non-LLM search at equal budget. Of the 19 loops that run online experiments, none corrects for the many rounds it tests and one confirms its final candidate in a fresh experiment. Where transfer to held-out data is checked, some selected gains survive and some do not. Four loops report that their agent inflated the metric or leaked the target; where the paper says what caught it, it was people or the agent's own diagnosis, not the acceptance rule. We distil these findings into a two-tier reporting standard mapped to the audit criteria.
Source: Preprints.org
Authors: Shanglin Yang, Zhan Shi
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7223202107