A search framework adapts recommendation pipelines across stages
TGR-EvoDAG: A DAG-based Multi-Agent Code Evolution Framework for Generative Recommendation on Tencent · arXiv:2610.32326
In plain words
Generative recommendation systems predict a person's next item through linked training stages. A change early in the pipeline can make scores drop until later stages are adjusted, and different changes can take different amounts of computing time.
TGR-EvoDAG is a framework that searches through versions of this pipeline, keeps promising changes in branches, and can combine branches or adapt later stages. On Amazon Office Products, it beat a general search framework that also received diagnostic feedback, on tests of next-item recommendations.
The research
When a recommendation pipeline has linked training stages, an upstream change can briefly lower its score before later stages adapt, while retraining costs vary across stages. Can code search preserve useful upstream changes and budget for the work needed to make them pay off?
TGR-EvoDAG searches recommendation pipelines by tracking candidate branches, planning stage costs and adapting downstream stages after upstream changes.
On Amazon Office Products, the authors evolved candidate recommendation pipelines from a shared starting codebase under the same 24 GPU-hour budget. They compared TGR-EvoDAG with Arbor + Diag. , which also received diagnostic feedback, and measured how often each user's held-out next item appeared in the top 10 and how highly it ranked.
TGR-EvoDAG scored 6.8% higher in Hit@10 and 16.9% higher in NDCG@10 on the test set. Because both approaches had diagnostic feedback, the result suggests that feedback alone does not explain the difference; their search procedures also differed.
The evidence comes from evolution on one public dataset and transfer to one industrial platform. The framework comparisons change several search mechanisms together, and the reported search ablation gives a qualitative pattern rather than numerical effects for each mechanism.
What the evidence shows
- Why pipeline search is difficult
TGR-EvoDAG is a multi-agent code evolution framework for generative recommendation, which predicts a user's next item through linked training stages. The search problem is unusually coupled: changing an upstream stage can temporarily lower scores until a downstream stage adapts, and retraining costs differ by stage.
A search that considers only immediate scores or treats every edit as equally costly can miss changes whose benefits arrive later.
- Keep and combine branches
The framework records candidate codebases and their dependencies in a Code DAG, where branches preserve different lines of work. Explore tries a new direction, Mutate refines a candidate, and Merge combines changes found in separate branches.
This gives the search a way to retain and combine candidate improvements across stages, rather than advancing only along one sequence of edits.
- Plan for cost and adaptation
For each proposed edit, the Operator Selector identifies the affected stage and estimates retraining cost from earlier measured stage costs. It weighs estimated cost against anticipated performance gain; after execution, the budget is charged for the measured GPU-hours.
If an upstream edit causes a regression because a later stage is incompatible, the Pruner can preserve it with a tag that directs the next edit to adapt a downstream stage.
- Compare feedback and search
On Amazon Office Products, under a shared 24 GPU-hour budget and starting codebase, TGR-EvoDAG exceeded Arbor + Diag. on test Hit@10 and NDCG@10. Arbor + Diag. already had mathematical probes and simulated-user diagnoses, and adding those diagnostics to Arbor alone also improved validation Hit@10.
These comparisons indicate that richer feedback helps, while it does not by itself account for the gap between the two tested frameworks.
- Check controls and transfer
Restricted-search results show a qualitative rise in test and simulated-user measures from greedy search to the full framework, though the paper passage does not quantify each restriction's effect. The evolved architecture also transferred to one Tencent platform: it improved offline next-item ranking against APAO, and adding it as a retrieval channel in a live test raised overall system click-through rate. These results extend the evidence beyond the public dataset, but only to that industrial setting.
Deep
- Original paper authors
- Junlin He, Junzhang Zhu, Yudong Li, Yihong Tang, Xuefeng Jiang
- Submitted to arXiv
- License
- CC BY 4.0