Taiji: Pareto policy optimization for semantic and collaborative IDs¶

简体中文

Fidelity: core mechanism. Runs the defining mechanism verifiable on public data. Private data, production models, and serving remain outside scope.

Paper information¶

Field	Details
Paper	arXiv 2606.03866
Organization	Kuaishou Technology
First public release	2026-06-02 (arXiv v1)
Authors' code	No public author code was found on 2026-09-02.
Adapter / method	`taiji`
Local reproduction code	`src/auto_research/reproductions/taiji/`

Original paper¶

Motivation and method¶

Use reverse-engineered preference reasoning and rejection sampling to activate recommendation chain-of-thought. POPO then balances LLM semantic reward and collaborative-ID reward dynamically, avoiding the nonconvex Pareto-frontier blind spots of fixed weights.

flowchart LR
  A["Public behavior and candidates"] --> B["Taiji core mechanism"]
  B --> C["Matched-budget method output"]
  A --> D["Common direct baseline"]
  C --> E["Full-catalog ranking and structural diagnostics"]
  D --> E

Key figure from the paper¶

Figure 1: Core architecture, modules, and connections. From the original paper; copyright remains with the authors.

Core equations¶

\[ r(x,y)=w_s(x,y)r_{sem}(x,y)+w_{id}(x,y)r_{collab}(x,y),\\qquad w\\in\\Delta^2. \]

Paper results¶

A one-week Kuaishou ads experiment with 10% traffic per arm reports overall ADVV +2.83% and revenue +3.30%; full deployment began in 2026-05. This is paper evidence, not a local public-data outcome.

Local reproduction¶

Local comparison: Under the same full-catalog protocol, baseline NDCG@10 is 0.05401 and Taiji NDCG@10 is 0.06448, +19.39% relative. The local proxy does not extrapolate to online lift.

See the three-seed artifact for complete results, means, standard deviations, and 95% confidence intervals.

auto-research reproduce --paper taiji --dataset-dir data --seeds 42,43,44

Reproduction boundaries¶

The local run uses a public MovieLens-1M subset and auditable proxies to validate the core computation. It does not reproduce private logs, production models, online traffic assignment, or serving. No claim is made to reproduce absolute paper metrics or online gains.