TagLLM: fine-grained post tag generation¶

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 2603.21481
Organization	Tongji University
First public release	2026-03-23 (arXiv v1)
Authors' code	No public author code was found on 2026-09-05.
Adapter / method	`tagllm`
Local reproduction code	`src/auto_research/reproductions/tagllm/`

Original paper¶

Motivation and method¶

Build a user-interest handbook by post category. Multimodal chain-of-thought extracts fine-grained tags by interest dimension. Two-stage knowledge distillation transfers tagging to a smaller model for explainability and serving efficiency.

flowchart LR
  A["Public behavior and candidates"] --> B["TagLLM 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 2: Overall workflow, key stages, and data flow. From the original paper; copyright remains with the authors.

Core equations¶

\[ \mathcal L=\mathcal L_{tag}+\lambda\,D_{KL}\!\left(p_{teacher}(y|x,H)\Vert p_{student}(y|x)\right). \]

Paper results¶

Online A/B reports watch time per user +0.31%, interactions +0.96%, and cold-start PV CTR +32.37%. 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 TagLLM NDCG@10 is 0.05147, -4.69% 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 tagllm --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.