← Back to all news
arXiv Machine Learning By Suwan Wu, Yumeng Lin, Pengcheng Yuan, Xiaolong Jiang

SIRF: A Spec-Internalized Risk Foundation Model for Industrial Content Risk Control

Read the original on arXiv Machine Learning →

The paper introduces SIRF, a Spec-Internalized Risk Foundation Model designed for industrial content risk control. SIRF internalizes complex platform policies through EntiGraph, MAGA rewriting, and account-level chain-of-thought, then embeds them into the model weights via continued pretraining, enabling high‑precision, ultra‑low‑latency verdict‑only deployment. In controlled comparisons, SIRF‑8B‑SFT achieves a 71.3% Black Recall@P95—15.1 percentage points higher than a baseline—using only ~70 M CPT tokens and improves mis‑penalization recovery by 20% in a tree‑model adjudication layer.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Machine Learning.

Related stories

arXiv AI
Oct 28

CASEL: Consequence-Aware Selective Execution for Reliable Long-Horizon Tool-Using Agents

CASEL (Consequence-Aware Selective Execution Layer) is a runtime allocation policy for tool‑using language agents that chooses among direct execution, verification, sandboxing, clarification, approval, and abstention for each proposed tool call. It combines calibrated failure probability, state‑dependent consequence, reversibility, ambiguity, and evidence freshness to minimize a per‑mode cost. In controlled simulations, CASEL cuts harmful‑side‑effect trajectories from 14.75 % to 7.02 % and reduces mean normalized harm from 0.184 to 0.123 compared to an uncertainty‑ranked gate.

By Yanyan Zhang, Fei Shen, Ting Yan, Hao Zhang
agents
More like this →
arXiv Machine Learning
Oct 15

Calibration and Risk-Budgeted Decisions with System One Models

The paper investigates how large language model (LLM) workflows and agents make binary decisions—such as allowing or blocking an input—when human deferral is not possible. It proposes using System One decision scores that default to a conservative fallback in uncertain cases and control false omission risk within the subset of inputs that receive the utility‑maximizing action. Experiments on six datasets spanning moderation, safety, and reranking show that naive threshold baselines often exceed target risk, whereas a Learn‑Then‑Test approach achieves only 7.8% risk, demonstrating the limitations of probability‑like interfaces and the benefits of risk‑budgeted selection.

By Soundouss Messoudi, Youness Mansar
llmsragagentsfine-tuningsafety
More like this →
arXiv Computation and Language
Oct 1

Safety Monitors Mostly Catch What the Model Already Refuses

The paper evaluates safety monitors by measuring recall only on prompts that the target model actually answers, rather than on all harmful prompts. Across several guard systems, recall at a 1% false‑positive rate drops sharply when focusing on answered prompts, with monitors catching refused requests 1.1–6.4 times more often than answered ones. Rewriting prompts to be less explicit dramatically increases compliance and reveals that many harmful requests slip past monitors, especially when phrasing is softened. Fine‑tuning guards on these rewritten prompts improves recall from 0.24 to 0.89 on answered requests and generalizes to unseen benchmarks.

By Sripad Karne
fine-tuningbenchmarkssafety
More like this →