MetaStrategy: Generating executable ranking strategies for industrial recommendation¶
Fidelity: Core mechanism. Trains a context-conditioned strategy generator and executes typed strategies with a deterministic compiler. Negative local results are retained.
Paper information¶
| Field | Value |
|---|---|
| Paper link | arXiv 2608.09440 |
| Organization | Alibaba / Taobao |
| First published | 2026-08-10 (arXiv v1) |
| Original author code | No author code found (checked 2026-08-13). |
| Adapter | metastrategy |
| Local reproduction code | src/auto_research/reproductions/metastrategy/ |
Core training is in latest_20260813_common.py.
Original paper summary¶
Background and main changes¶
A fixed scoring network has difficulty adapting objectives to the user's current intent. MetaStrategy has an LLM emit typed JSON with objective weights, content and category preferences, constraints, and position rules. A deterministic compiler executes the strategy on candidates so free text does not directly control live ranking. Production training also uses replay, a self-competition curriculum, and evaluator-routed reward-augmented OPD to distill a 4B teacher into a 0.8B student.
flowchart LR
C["User and request context"] --> G["LLM Strategy Generator"]
G --> J["typed JSON strategy"]
J --> V["schema / safety validation"]
V --> E["deterministic compiler"]
I["Candidates and multi-objective signals"] --> E
E --> R["Final ranking"]Key figure from the paper¶
Original Figure 1 (pipeline): Key stages and data flow. Image from the original paper; copyright belongs to the authors. Click the image for its source.
Core equation¶
The generator only outputs auditable parameters. The compiler deterministically executes the ranking.
Offline and online results in the paper¶
A seven-day randomized user A/B test on Taobao's home-page “Guess You Like” feed reports click PV +2.11%, IPV +3.12%, transaction value +2.83%, and impression PV +1.49%.
Local reproduction¶
Control: Same-budget single-objective ranker versus typed strategy generator and deterministic compiler. Relative NDCG@10 change is -8.51%.
MovieLens-1M uses 260 users, 420 items, 50 steps, and seed 42. NDCG@10 is 0.02083 for the single-objective baseline versus 0.01906 (-8.51%) with typed strategies. Head share decreases from 0.06538 to 0.03731, and 2,400 strategy bundles execute. Short-budget multi-objective constraints trade off relevance; this is not a reproduction of the online lift.
Stable metrics: metrics/movielens1m-seed42.json.
Reproduction limits¶
Taobao's private logs, production replay, 4B-to-0.8B OPD, and nearline serving are not reproduced. The public experiment tests only typed-bundle generation, validation, and deterministic execution.