Large Language Model (LLM) agents have enabled automated code evolution for algorithm discovery. However, current recursive self‑improvement (RSI) research for sequential recommendation faces a dilemma: general‑purpose frameworks lack domain‑specific feedback, while recommender‑specific frameworks only support linear evolution. Naïvely combining them fails for Generative Recommendation (GR), which differs from end‑to‑end models by relying on a module‑cascaded pipeline (tokenization, pre‑training, alignment). Existing evolution methods thus suffer from two major issues:
- Cost blindness – they ignore asymmetric retraining overheads across stages, wasting GPU budgets;
- Short‑sighted evaluation – they discard upstream innovations that temporarily regress metrics before downstream adaptation.
To address these gaps we propose TGR‑EvoDAG, a multi‑agent code evolution framework tailored for GR. It models code evolution as a Directed Acyclic Graph (DAG), preserving multiple evolution branches and merging complementary innovations while unifying advanced search with rich GR diagnostic feedback. The framework introduces two key mechanisms:
- Stage‑aware cost planning – dynamically allocates GPU‑hour resources per stage to avoid unnecessary expenditure;
- Active downstream co‑adaptation – retains upstream improvements despite temporary regressions, allowing downstream retraining to unlock cross‑stage gains.
On public Amazon datasets, TGR‑EvoDAG achieves the first fully autonomous evolution of a complete GR pipeline, improving offline metrics by up to 37.4% and boosting Agent4Rec simulated user satisfaction by 19.7%, surpassing the strongest baseline. In industrial deployment on Tencent’s large‑scale commercial media platform, the evolved GR yields a 12.16% relative improvement in offline Hit@1 and a statistically significant 0.40% increase in online click‑through rate (CTR).
Review