SIRF: A Spec-Internalized Risk Foundation Model for... | AI Research

SIRF: A Spec-Internalized Risk Foundation Model for Industrial Content Risk Control introduces a new approach to industrial content moderation that prioritizes high-precision, low-latency decision-making. In high-stakes environments like social platforms, the primary goal is to minimize the unfair penalization of good users while effectively catching risky content. The authors argue that existing methods—such as providing long policy documents to a model at runtime or relying on standard classification—struggle to meet the dual demands of complex policy application and second-level response times. SIRF solves this by "internalizing" platform policies directly into the model's weights, allowing it to apply rules instantly without needing to retrieve external information for every query.

How SIRF Works

The core of the SIRF approach is a specialized training pipeline that moves policy knowledge from external documents into the model’s internal memory. The team uses an automated synthesis process to create a training corpus without requiring new human labels. This process involves three main techniques:

• EntiGraph: Extracting key entities and relations from policies to define boundaries between acceptable and prohibited behavior.

• MAGA Rewriting: Rephrasing policies for different audiences and genres to ensure the model understands the rules in various contexts.

• Account-level CoT Distillation: Using a teacher model to generate "chain-of-thought" reasoning chains that link account features to specific policy clauses and final verdicts. The same ai evaluation question is explored in Discriminative World Models for Web Agents, which adds a research perspective.
  Once this knowledge is internalized through continued pretraining (CPT), the model is fine-tuned for specific risk domains. During deployment, the model operates in a "verdict-only" mode, meaning it outputs only the final decision without needing to generate a reasoning trace, which keeps latency extremely low.

Key Results

In a controlled comparison against a baseline model of the same size, SIRF demonstrated significant improvements in identifying risky content at high precision. Specifically, SIRF-8B-SFT achieved a 71.3% Black Recall@P95, which is a 15.1 percentage point increase over the baseline. The researchers found that the model’s advantage becomes more pronounced as the precision requirements become stricter, meaning it is particularly effective at identifying high-confidence risks. Furthermore, the model achieved these gains using only about 70 million training tokens—a tiny fraction of its original pretraining data—without degrading its general knowledge or reasoning abilities. The same ai systems question is explored in Efficient Test-Time Adaptation through Human-AI Interaction, which adds a research perspective.

Deployment and Efficiency

Beyond its performance, SIRF offers practical benefits for industrial deployment. Because the policy knowledge is internalized, the system requires shorter prompts, which reduces the memory footprint and increases the number of requests the model can handle per second. The model also proves to be transferable; with light fine-tuning, the same base model can be adapted to different risk domains, resulting in a roughly 70% relative reduction in mis-penalized samples.

Important Considerations

While SIRF shows strong results, the authors note that the model’s performance is sensitive to the quality of the synthesized training data. The "no new human labels" claim refers to the fact that the training corpus is generated automatically from existing policy documents and account features, rather than through manual annotation of new data. Additionally, the researchers emphasize that their core findings rely on the comparison between models that provide log-probability outputs, which allows for the precise thresholding necessary for high-stakes risk control. Models that do not provide these outputs cannot be evaluated using the same metrics, which is an interface limitation rather than a reflection of their underlying capabilities. The same ai evaluation question is explored in Xiaomi-TabLDM, which adds a research perspective. as detailed in the full paper on Arxiv