SIRF improves industrial content risk control with faster high precision decisions
SIRF: A Spec-Internalized Risk Foundation Model for Industrial Content Risk Control
Artificial IntelligenceComputation and LanguageMachine Learning
Summary
Controlling risky online content in industries requires very accurate and fast decisions, not just good average results. The authors created SIRF, a system that learns complex policies directly inside its model by training on specially processed data without extra human labeling. This internal knowledge helps SIRF make precise decisions quickly with less delay, outperforming a similar baseline. It is already used to improve decision-making layers and reduce unnecessary penalties in content management.
What this means in practice
- •For content moderation teams: Automatically enforce complex content policies quickly at scale with minimal false penalties.
- •For industrial ai operations: Deploy risk control models that produce fast, precise verdicts without extra human annotation costs.
Authors
Suwan Wu, Yumeng Lin, Pengcheng Yuan, Xiaolong Jiang
Abstract
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).