CARO: Analogical Reasoning Chain Optimization Innovates Fuzzy Boundary Recognition in Content Moderation

The CARO (Analogical Reasoning Chain Optimization) framework injects an analogical reasoning mechanism through two-stage training (RAG-guided supervised fine-tuning + customized Direct Preference Optimization). It effectively solves the problem that LLMs are easily misled by decision shortcuts in fuzzy content moderation. Experiments show its average F1 score increases by 24.9% in complex moderation scenarios, outperforming advanced reasoning models like DeepSeek R1 and providing a new solution for fuzzy boundary recognition.

Content moderation has become more complex due to the popularity of generative AI, with an increase in cases with fuzzy boundaries. Existing LLMs tend to rely on surface features (decision shortcuts) and make mistakes when handling gray areas—for example, whether weight loss advice containing drug names is promotion, or whether historical and political discussions incite hatred (cases without obvious violation features). However, human experts make judgments through analogical reasoning, a capability LLMs lack.

The CARO framework uses two-stage training:

1. RAG-guided supervised fine-tuning: Retrieve similar cases to build analogical reasoning chains and teach the model analogical capabilities;
2. Customized DPO: The optimization target focuses on the quality of analogical reasoning, ensuring the process is reasonable rather than only focusing on results. In addition, CARO adopts dynamic analogy generation, which constructs relevant references in real time for the current case, overcoming the limitations of traditional static retrieval.

Experimental evaluations on fuzzy moderation benchmarks show:

1. An average F1 improvement of 24.9% over reasoning models like DeepSeek R1;
2. Better than dedicated moderation models like LLaMA Guard in handling fuzzy cases;
3. Ablation experiments verify that removing the analogy chain or replacing dynamic generation with static retrieval leads to significant performance degradation, proving the necessity of each component.

The core idea of CARO (enhancing boundary judgment through analogical reasoning) can be extended to fuzzy boundary fields such as legal case analysis, medical diagnosis assistance, and educational evaluation. Its interdisciplinary (cognitive psychology + AI training) paradigm provides a path to break through LLMs' bottlenecks.

The success of CARO shows that combining human cognitive mechanisms (analogical reasoning) is the key to building reliable content moderation systems. In the future, we should continue to explore formalizing and injecting human expert capabilities into AI to deal with increasingly complex online content, replacing simple keyword filtering and rule matching.