EVADE-Bench: Multimodal Benchmark for Evasive Content Detection in E-Commerce Applications

TL;DR

This work tackles evasive content detection in e-commerce by introducing EVADE, a Chinese, expert-annotated multimodal benchmark with 2,833 text samples and 13,961 images across six product categories, designed to test short- and long-context regulatory reasoning via Single-Violation and All-in-One tasks. It presents a thorough data collection, annotation, and filtering pipeline, plus a large-scale evaluation of 26 LLMs and VLMs, revealing substantial gaps in current models, especially under complex, policy-driven prompts. The All-in-One setting demonstrates that clearer, non-overlapping rule definitions can significantly improve alignment and reduce the performance gap more than simply increasing prompt length, with smaller models benefiting the most. The paper further shows that retriever-augmented generation can boost robustness in ambiguous cases, and provides public access to EVADE to catalyze safer, more transparent e-commerce moderation and multimodal reasoning research.

Abstract

E-commerce platforms increasingly rely on Large Language Models (LLMs) and Vision-Language Models (VLMs) to detect illicit or misleading product content. However, these models remain vulnerable to evasive content: inputs (text or images) that superficially comply with platform policies while covertly conveying prohibited claims. Unlike traditional adversarial attacks that induce overt failures, evasive content exploits ambiguity and context, making it far harder to detect. Existing robustness benchmarks provide little guidance for this demanding, real-world challenge. We introduce EVADE, the first expert-curated, Chinese, multimodal benchmark specifically designed to evaluate foundation models on evasive content detection in e-commerce. The dataset contains 2,833 annotated text samples and 13,961 images spanning six demanding product categories, including body shaping, height growth, and health supplements. Two complementary tasks assess distinct capabilities: Single-Violation, which probes fine-grained reasoning under short prompts, and All-in-One, which tests long-context reasoning by merging overlapping policy rules into unified instructions. Notably, the All-in-One setting significantly narrows the performance gap between partial and full-match accuracy, suggesting that clearer rule definitions improve alignment between human and model judgment. We benchmark 26 mainstream LLMs and VLMs and observe substantial performance gaps: even state-of-the-art models frequently misclassify evasive samples. By releasing EVADE and strong baselines, we provide the first rigorous standard for evaluating evasive-content detection, expose fundamental limitations in current multimodal reasoning, and lay the groundwork for safer and more transparent content moderation systems in e-commerce. The dataset is publicly available at https://huggingface.co/datasets/koenshen/EVADE-Bench.

Figures (11)

• Figure 1: Illustration of an EVADE sample.
• Figure 2: Comparison of LLMs and VLMs before and after the introduction of RAG. Here, L- denotes the Llama-3.1 model, Q- denotes the Qwen2.5 model, DS- denotes the Deepseek model, Q3- denotes the Qwen3 model, and IVL- denotes the InternVL3 model.
• Figure 3: All LLMs’ performance on the Single-Violation task.
• Figure 4: All VLMs’ performance on the Single-Violation task.
• Figure 5: Qwen3 series models’ performance on the Single-Violation task.