Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Seeing the Unseen: Rethinking Illicit Promotion Detection with In-Context Learning

Sangyi Wu, Junpu Guo, Xianghang Mi

Abstract

Illicit online promotion is a persistent threat that evolves to evade detection. Existing moderation systems remain tethered to platform-specific supervision and static taxonomies, a reactive paradigm that struggles to generalize across domains or uncover novel threats. This paper presents a systematic study of In-Context Learning (ICL) as a unified framework for illicit promotion detection. Through rigorous analysis, we show that properly configured ICL achieves performance comparable to fine-tuned models using 22x fewer labeled examples. We demonstrate three key capabilities: (1) Generalization to unseen threats: ICL generalizes to new illicit categories without category-specific demonstrations, with a performance drop of less than 6% for most evaluated categories. (2) Autonomous discovery: A novel two-stage pipeline distills 2,900 free-form labels into coherent taxonomies, surfacing eight previously undocumented illicit categories such as usury and illegal immigration. (3) Cross-platform generalization: Deployed on 200,000 real-world samples from search engines and Twitter without adaptation, ICL achieves 92.6% accuracy. Furthermore, 61.8% of its uniquely flagged samples correspond to borderline or obfuscated content missed by existing detectors. Our findings position ICL as a new paradigm for content moderation, combining the precision of specialized classifiers with cross-platform generalization and autonomous threat discovery. By shifting to inference-time reasoning, ICL offers a path toward proactively adaptive moderation systems.

Seeing the Unseen: Rethinking Illicit Promotion Detection with In-Context Learning

Abstract

Illicit online promotion is a persistent threat that evolves to evade detection. Existing moderation systems remain tethered to platform-specific supervision and static taxonomies, a reactive paradigm that struggles to generalize across domains or uncover novel threats. This paper presents a systematic study of In-Context Learning (ICL) as a unified framework for illicit promotion detection. Through rigorous analysis, we show that properly configured ICL achieves performance comparable to fine-tuned models using 22x fewer labeled examples. We demonstrate three key capabilities: (1) Generalization to unseen threats: ICL generalizes to new illicit categories without category-specific demonstrations, with a performance drop of less than 6% for most evaluated categories. (2) Autonomous discovery: A novel two-stage pipeline distills 2,900 free-form labels into coherent taxonomies, surfacing eight previously undocumented illicit categories such as usury and illegal immigration. (3) Cross-platform generalization: Deployed on 200,000 real-world samples from search engines and Twitter without adaptation, ICL achieves 92.6% accuracy. Furthermore, 61.8% of its uniquely flagged samples correspond to borderline or obfuscated content missed by existing detectors. Our findings position ICL as a new paradigm for content moderation, combining the precision of specialized classifiers with cross-platform generalization and autonomous threat discovery. By shifting to inference-time reasoning, ICL offers a path toward proactively adaptive moderation systems.

Paper Structure

This paper contains 27 sections, 10 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Illicit Online Promotion
  4. In-Context Learning
  5. Experimental Setup
  6. Datasets
  7. Models
  8. Analysis of Prompt Configurations
  9. Impact of Demonstration Quantity
  10. Demonstration Selection Strategies
  11. Instruction Tuning and Model Alignment
  12. Label Necessity and Semantic Verbalization
  13. Comparative Efficiency and Robustness Analysis
  14. Data Efficiency: ICL vs. Parameter-Efficient Fine-Tuning
  15. Robustness to Prompt Ordering and Positional Bias
  16. ...and 12 more sections

Figures (10)

  • Figure 1: Impact of Demonstration Quantity. F1-scores of different model families as a function of the number of shots for (a) binary and (b) multiclass classification tasks.
  • Figure 2: Efficacy of Demonstration Selection Strategies. Comparison of Random, Lexical (BM25), and Semantic (Embedding-based) selection strategies across varying shot counts using the Mistral model.
  • Figure 3: Instruction Tuning Impact. Performance comparison between Instruction-Tuned and Base variants across Llama, Mistral, and Qwen families (32-shot, semantic retrieval).
  • Figure 4: Necessity of Explicit Labels. Impact of removing labels from demonstrations compared to Zero-shot and Standard (Labeled) Few-shot baselines on FPR and F1 Score.
  • Figure 5: Sensitivity to Label Verbalization. Comparison of Original (Semantic), Inverted, and Abstract (Symbolic) label mappings for (a) binary and (b) multiclass tasks.
  • ...and 5 more figures