SIRF: a foundation model that internalizes industrial content-risk rules¶
Concept diagnostic. Implements rule-relation synthesis, trigger/exemption learning, and P95 threshold selection. A linear model stands in for 8B LLM continued pretraining.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv v1 |
| Organization | Xiaohongshu |
| First public release | 2026-09-10 (arXiv v1) |
| Authors' code | No public author repository was found when checked on 2026-09-14. |
| Adapter | sirf |
| Local reproduction code | src/auto_research/reproductions/sirf/ |
Original paper¶
Motivation and method¶
SIRF synthesizes long-tail policy examples with EntiGraph, MAGA rewriting, and account-level chain-of-thought. Continued pretraining writes complex trigger/exemption rules into model weights. A single verdict and adjustable threshold support decisions on a seconds-scale serving path.
flowchart LR
S[Platform rules] --> E[EntiGraph, MAGA and account CoT]
E --> C[About 70M continued-pretraining tokens]
C --> F[SFT verdict model]
F --> T[High-precision threshold and online verdict]Key figure from the paper¶
Figure 1: Original method design and key components. From the paper; copyright remains with the authors.
Core computation¶
The local model represents rules with triggers, exemptions, and interactions. After training, it scans thresholds on held-out policy cases and reports maximum Black Recall at Precision≥95%.
Paper results¶
Against a matched Qwen3-8B-SFT control, Black Recall@P95 rises 15.1 percentage points. The adjudication layer releases about 20% more incorrectly penalized cases. In the account-freeze scenario, wrongful penalties decline by about 70% relative, at a cumulative million-user scale. The full text also reports a significant rise in weekly active penetration in randomized treatment/control A/B, disclosing only a high-single-digit to low-double-digit range.
Local reproduction¶
Local comparison: Against matched linear features, adding rule interactions increases Black Recall@P95 by 65.81% relative over three seeds, about 33.84 absolute percentage points. These are not results for the 8B model or online traffic.
See the three-seed results. The paper's online results and local synthetic-data measurements are kept separate.
Reproduction boundaries¶
No private account data, 8B continued pretraining, EntiGraph/MAGA language generation, or online serving were reproduced. Local results cannot verify the reported business lift.