ChAIRO : Contextual Hierarchical Analogical Induction and Reasoning Optimization for LLMs
Haotian Lu [email protected], Yuchen Mou [email protected], Bingzhe Wu [email protected] · 2026
Read ChAIRO : Contextual Hierarchical Analogical Induction and Reasoning Optimization for LLMs by Haotian Lu [email protected], Yuchen Mou [email protected], Bingzhe Wu [email protected] (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-10502.txt.
Introduction
Lu, Mou, and Wu police online moderation with analogy driven rules that adapt on the fly. Fixed rules and plain tuned models strain as user posts grow stranger and shift. Their CHAIRO frame tunes example search, rule making, and verdict making end to end. Fresh analogies let rules bend to new cases instead of freezing on old ones. Tests beat tuned and static retrieval rivals on verdict quality and rule clarity.
Catalog record
| CREATOR | Haotian Lu [email protected], Yuchen Mou [email protected], Bingzhe Wu [email protected] |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 36069 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.10502
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
ChAIRO : Contextual Hierarchical Analogical Induction and Reasoning Optimization for LLMs Thanks: This work was conducted while Haotian Lu and Yuchen Mou were interning at the National Engineering Laboratory for Big Data System Computing Technology under the supervision of Bingzhe Wu. Haotian Lu [email protected], Yuchen Mou [email protected], Bingzhe Wu [email protected] Abstract Warning: This paper may contain content that could be disturbing or offensive. Content moderation in online platforms faces persistent challenges due to the evolving complexity of user-generated content and the limitations of traditional rule-based and machine learning approaches. While recent advances in large language models (LLMs) have enabled more sophisticated moderation via direct prompting or fine-tuning, these approaches often exhibit limited generalization, interpretability, and adaptability to unseen or ambiguous cases. In this work, we propose a novel moderation framework that leverages analogical examples to enhance rule induction and decision reliability. Our approach integrates end-to-end optimization of analogical retrieval, rule generation, and moderation classification, enabling the dynamic adaptation of moderation rules to diverse content scenarios. Through comprehensive experiments, we demonstrate that our method significantly outperforms both rule-injected fine-tuning baselines and multi-stage static RAG pipelines in terms of moderation accuracy and rule quality. Further evaluations, including human assessments and external model generalization tests, confirm that our framework produces rules with better clarity, interpretability, and applicability. These findings show that analogical example-driven methods can advance robust, explainable, and generalizable content moderation in real-world applications. Keywords content moderation \cdot large language models \cdot analogical reasoning \cdot rule induction 1 Introduction The exponential growth of online content has made content moderation an indispensable component in maintaining healthy, safe, and compliant digital environments ( 37 ) . Automated content moderation systems are increasingly relied upon to filter out harmful, illegal, or inappropriate material on social media platforms, forums, and other user-driven services ( 38 ) . As large language models (LLMs) have demonstrated remarkable progress across various natural language understanding tasks, deploying LLMs for content moderation has become a promising direction ( 16 ; 24 ; 34 ) . Recent studies have explored the application of LLMs to content moderation tasks through diverse strategies, including post-training ( 25 ; 28 ; 15 ; 19 ; 20 ) and prompt engineering ( 27 ; 26 ; 16 ; 1 ; 3 ) , yielding promising progress in both moderation accuracy and reasoning abilities ( 17 ; 32 ) . However, despite their impressive capabilities, even state-of-the-art LLMs often struggle in scenarios characterized by contextual ambiguity or vague moderation criteria ( 21 ; 11 ; 14 ) . For example, when moderation rules are implicit, incomplete, or open to interpretation, LLMs may produce inconsistent or erroneous judgments, undermining the reliability of automated moderation systems. As shown in Figure 1 , Chain of thought (CoT) relies solely on explicit standards such as the absence of insults, incitement, or attacks on specific groups, and fails to recognize the underlying discriminatory logic of the statement “low scores equal low ability.” As a result, it incorrectly classifies the statement as "Safe." This approach lacks the deep semantic understanding needed to address finer-grained, metaphorical, or indirect discriminatory content. Conversely, when explicit and precise moderation rules are incorporated into the context, models demonstrate significantly improved accuracy and interpretability, as Figure 1 illustrates. This observation highlights the importance of well-defined moderation rules, which improve both moderation precision and transparency of the model’s decision-making process ( 30 ; 17 ; 35 ) . Nevertheless, identifying or constructing the most appropriate moderation rule for a given content instance remains a challenging problem. Existing solutions typically fall into two categories: (1) manually defined high-level rules, such as those targeting broad categories like "sexual content." While effective to some extent, such rules often fail to account for the nuanced differences among fine-grained instances or across diverse application scenarios, making it virtually impossible to exhaustively enumerate all necessary rules ( 2 ; 9 ) . (2) LLM-driven adaptive rule discovery, which leverages the model’s world knowledge and prompt engineering to synthesize rules on the fly ( 17 ) . However, these approaches frequently overlook domain-specific expertise and the rich experience accumulated by human moderators, relying instead on generic or coarse-grained priors. To address these challenges, we propose leveraging the inductive power of LLMs to generalize rules from analogous instances within the same moderation context. Our key insight is that by systematically analyzing similar content samples and their corresponding moderation outcomes, models can distill more robust and contextually relevant rules that generalize better across instances. A straightforward approach to realizing this high-level idea is to first retrieve samples similar to the current instance from an existing database and then employ an auxiliary LLM to induce and analyze the relevant rules based on these retrieved examples. However, this pipeline separates the processes of rule generation and content moderation, and thus may lose fine-grained cues most pertinent to the current sample during rule induction. To address this limitation, we introduce ChAIRO , a c ontextual h ierarchical a nalogical i nduction and r easoning o ptimization framework, which jointly optimizes the case retrieval and rule induction process. By jointly optimizing these components, our method ensures that the induced rules are grounded in highly relevant examples and better tailored to each moderation instance. This end-to-end approach allows the model to more effectively utilize the annotated data and uncover hidden expert knowledge from human moderators. Concretely, our post-training framework is designed to endow the model with an integrated three-stage reasoning capability: example generation, rule induction, and final moderation. Our framework consists of three critical steps: we first fine-tune the base LLM with a chain-of-analogy approach on the training set, enabling the model to autonomously generate the most relevant analogical cases for each content sample. Second, we employ an auxiliary rule-generation module that synthesizes explicit moderation rules by analyzing the commonalities between the original and analogous cases. Finally, these generated rules are injected back into the LLM’s moderation reasoning chain during a second round of fine-tuning, equipping the model with both exemplar-based and rule-inductive reasoning capabilities. Through comprehensive experiments, we demonstrate that our framework leads to more accurate and robust moderation outcomes while enhancing the interpretability and adaptability of LLM-based moderation systems across diverse, real-world scenarios. Figure 1: Comparison of the Chain of Thought (CoT) method and the ChAIRO framework for content moderation. CoT fails to identify the implicit bias and classifies it as ’Safe,’ while ChAIRO , leveraging analogical reasoning and explicit rule induction, correctly identifies it as ’Unsafe’ by recognizing the underlying discriminatory logic. 2 Method Figure 2: The workflow of the ChAIRO framework. Our proposed framework, ChAIRO , leverages a systematic three-stage pipeline to enhance the reasoning capabilities of LLMs in content moderation tasks. Specifically, our approach comprises the following stages as shown in Figure 2 : First, we introduce a self-augmented analogical chain-of-thought generation strategy, enriching training data to empower the model’s capability of adaptively retrieving relevant analogical moderation cases. Second, an auxiliary reasoning LLM performs inductive reasoning over these retrieved examples, extracting explicit moderation rules tailored to specific moderation contexts. Finally, we inject these induced moderation rules back into the reasoning chains and fine-tune the model again, integrating exemplar-based and rule-inductive reasoning capabilities. This structured approach effectively leverages annotated moderation