发表机构
Xiaohongshu Inc.; Tianjin University（小红书公司; 天津大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
SIRF通过继续预训练将平台策略内化到模型权重中，在秒级延迟下实现高精度自动风险控制，显著提升召回率并降低误罚。
AI 中文摘要
对于工业内容风险控制，实际部署的约束条件并非平均准确率，而是在高精度和秒级延迟下能够自动处理多少风险。我们提出了SIRF（规范内化风险基础模型），该模型通过继续预训练（CPT），将平台的复杂策略内化到模型权重中，这些策略通过EntiGraph、MAGA重写和账户级思维链（CoT）合成，无需额外的人工标注，从而在超低延迟、仅输出裁决的部署下以高精度应用规则。一项受控的同源对比（Qwen3-8B-SFT与SIRF-8B-SFT，采用相同的策略注入和仅输出裁决的输出形式，仅在基于策略的CPT上有所不同）将性能提升归因于内化：SIRF-8B-SFT达到了71.3%的Black Recall@P95，比基线高出15.1个百分点，仅使用了约7000万个CPT token，且不损害通用能力；在包含的、可获取logprob的模型中，该接口下它达到或超过了远大于自身的系统。SIRF作为树模型裁决层部署（多恢复了20%的误罚样本），并以低成本迁移到冻结场景（相对误罚减少约70%）。
英文摘要
For industrial content risk control, the real deployment constraint is not average accuracy but how much risk can be auto-handled under high precision and second-level latency. We present SIRF (Spec-Internalized Risk Foundation Model), which internalizes a platform's complex policies, synthesized without additional human annotation via EntiGraph, MAGA rewriting and account-level chain-of-thought (CoT), into the weights via continued pretraining (CPT), so rules are applied at high precision under an ultra-low-latency, verdict-only deployment. A controlled same-source comparison (Qwen3-8B-SFT vs. SIRF-8B-SFT, identical policy injection and verdict-only output form, differing only in policy-grounded CPT) attributes the gain to internalization: SIRF-8B-SFT reaches 71.3% Black Recall@P95, +15.1pp over the baseline, using only ~70M CPT tokens without harming general ability, and among included, logprob-available models under this interface it matches or exceeds far larger systems. SIRF is deployed as a tree-model adjudication layer (20% more mis-penalized samples recovered) and transfers to a freezing scenario at low cost (~70% relative mis-penalization reduction).
Comments14 pages, 12 figures. Accepted at the Industry Track of EMNLP 2026
