Tencent researchers proposed TGR-EvoDAG, a multi-agent code evolution framework for generative recommendation. It models evolution as a Directed Acyclic Graph to merge complementary innovations. The framework adds stage-aware cost planning to optimize GPU-hour budgets. It also introduces active downstream co-adaptation to handle temporary regressions. On Amazon datasets, it improved offline metrics by up to 37.4%. It also raised Agent4Rec simulated user satisfaction by 19.7%. In Tencent’s commercial platform, it achieved a 12.16% offline Hit@1 gain. Online click-through rate rose by 0.40%.
What changed
Existing evolution methods used linear pipelines and ignored stage-asymmetric retraining costs.
What it unlocks
Autonomous evolution of a complete generative recommendation pipeline with cross-stage gains.
- 37.4% offline metrics improvement
- 19.7% Agent4Rec satisfaction gain
- 12.16% offline Hit@1 improvement
- 0.40% online CTR increase
Sources