OnePiece: Bringing Context Engineering and Reasoning to Industrial Cascade Ranking¶

Fidelity: core mechanism. Paper results, public-data measurements, and omissions are separate.

Paper information¶

Field	Details
Paper	arXiv 2509.18091
Organization	Shopee
First public release	2025-09-22 (arXiv v1)
Authors' code	The paper did not provide official or author code when checked on 2026-08-09.
Adapter / method	`onepiece`
Local reproduction code	`src/auto_research/reproductions/onepiece/`

Original paper¶

Motivation and method¶

Combine context tokens, block-level latent reasoning, and progressive multitask objectives in a cascade ranker. Retrieval, clicks, and value share representations but converge in stages.

flowchart LR
 A["Public input or history"] --> B["OnePiece core path"]
 B --> C["Paper-specific state or objective"]
 C --> D["Matched-budget evaluation"]

Key figure from the paper¶

Figure 2: Core architecture, modules, and connections. From the original paper; copyright remains with the authors.

Core equations¶

\[ H^{k+1}=H^k+F_k(H^k,c),\quad\mathcal L=\sum_t\alpha_t(k)\mathcal L_t. \]

Paper results¶

Shopee reports GMV/UU above +2% and ad revenue +2.90%.

Local reproduction¶

Local comparison: The baseline is the shared transition + content scorer. Adding only the OnePiece core mechanism changes NDCG@10 by +22.87% relative.

MovieLens-100K, 260 users and 420 items, seed 42: NDCG@10 0.0354 → 0.0435 (+22.87%).

auto-research reproduce --paper onepiece --dataset-dir data --seed 42
auto-research evolve --model rankmixer --dataset movielens-100k --direction "组合 onepiece 与已安装论文算子" --generations 2 --population 4

See the fixed metrics.

Reproduction boundaries¶

Full-catalog MovieLens-100K exercises the paper's core state, objective, or inference path. No private features, production traffic, or online serving are used. Paper A/B results are cited separately and local numbers do not represent the original large model, data, traffic, or custom kernels.