# CARO: Chain-of-Analogy Reasoning Optimization for Robust Content Moderation

Source: [CARO: Chain-of-Analogy Reasoning Optimization for Robust Content Moderation](<https://www.sota2.com/research/paper/69ddca117d0172140ec799b6>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-04-15T05:00:11.244Z

Authors: Bingzhe Wu, Haotian Lu, Yuchen Mou

Published: 2026-04-12T07:46:44.000Z

## Abstract

Current large language models (LLMs), even those explicitly trained for reasoning, often struggle with ambiguous content moderation cases due to misleading "decision shortcuts" embedded in context. Inspired by cognitive psychology insights into expert moderation, we introduce \\caro (Chain-of-Analogy Reasoning Optimization), a novel two-stage training framework to induce robust analogical reasoning in LLMs. First, \\caro bootstraps analogical reasoning chains via retrieval-augmented generation (RAG) on moderation data and performs supervised fine-tuning (SFT). Second, we propose a customized direct preference optimization (DPO) approach to reinforce analogical reasoning behaviors explicitly. Unlike static retrieval methods, \\caro dynamically generates tailored analogical references during inference, effectively mitigating harmful decision shortcuts. Extensive experiments demonstrate that \\caro substantially outperforms state-of-the-art reasoning models (DeepSeek R1, QwQ), specialized moderation models (LLaMA Guard), and advanced fine-tuning and retrieval-augmented methods, achieving an average F1 score improvement of 24.9\\% on challenging ambiguous moderation benchmarks.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Content Moderation on Multi-category Chinese Content Moderation Dataset 1.0 (test)](<https://www.sota2.com/research/sota/content-moderation-on-multi-category-chinese-content-moderation-dataset-1-0-test.md>): Content Moderation; Multi-category Chinese content moderation dataset 1.0 (test); Politics Accuracy: 89.7.

- [Content Moderation on Aegis In-Distribution](<https://www.sota2.com/research/sota/content-moderation-on-aegis-in-distribution.md>): Content Moderation; Aegis In-Distribution; Pornography Score: 75.

- [Content Moderation on OpenAI Out-of-Distribution](<https://www.sota2.com/research/sota/content-moderation-on-openai-out-of-distribution.md>): Content Moderation; OpenAI Out-of-Distribution; Pornography Score: 82.6.

- [Content Moderation on Toxic-Chat Out-of-Distribution](<https://www.sota2.com/research/sota/content-moderation-on-toxic-chat-out-of-distribution.md>): Content Moderation; Toxic-Chat Out-of-Distribution; Pornography Score: 60.8.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2604.10504>)

[PDF](<https://arxiv.org/pdf/2604.10504.pdf>)
