leoluo25933/forge-benchmark

FORGE: Fake Online Recommendations in Generative Environments

FORGE is a benchmark for measuring whether search-augmented large language models recommend synthetic fake brands when their retrieval evidence is poisoned. It contains 225 Chinese product queries across 15 categories, evaluation results for 12 production LLMs, and rebuildable evidence-bundle indexes.

This dataset accompanies the paper One Polluted Page Is Enough: Evaluating Web Content Pollution in LLM Recommenders (Luo & Chen, 2026). Accepted to EMNLP 2026 Findings.

• Code and full documentation: leoluolol/forge-benchmark
• Source snapshot used for this release: 6d5ec82461ed288f11280fd3f287f73122978ca3
• Paper: arXiv:2606.13610
• DOI: 10.48550/arXiv.2606.13610

Dataset configurations

The Hub repository exposes three configurations, each with a test split:

Configuration	Rows	Description
queries	225	Product queries and the 15-category taxonomy
results	225	Query-level fooled/top-1 indicators for 12 evaluated models
evidence_bundles	225	URL, anchor, and short-excerpt indexes used to rebuild clean retrieval evidence

from datasets import load_dataset

queries = load_dataset(
    "leoluo25933/forge-benchmark", "queries", split="test"
)
results = load_dataset(
    "leoluo25933/forge-benchmark", "results", split="test"
)
bundles = load_dataset(
    "leoluo25933/forge-benchmark", "evidence_bundles", split="test"
)

The queries configuration is the default:

queries = load_dataset("leoluo25933/forge-benchmark", split="test")

Data fields

queries

• query_id: stable query identifier
• scenario: high-level product scenario
• category: normalized product category
• product: Chinese product name
• query: Chinese recommendation query

results

Each row corresponds to one query. Besides query_id, product, category, and attack, the table contains two binary fields per evaluated model:

• <model>__fooled: whether the response recommends the synthetic fake brand
• <model>__top1: whether the synthetic fake brand is ranked first

The evaluated models are Gemini 3 Flash, GPT-5.4, o4-mini, Gemini 3.1 Pro, Claude Opus 4.7, Claude Sonnet 4.6, Qwen3.6-27B, Qwen3.6-35B-A3B, Qwen3.5-9B, DeepSeek V4 Pro, GLM-4.6V-Flash, and Ministral 3R.

evidence_bundles

• query_id: stable query identifier
• query: Chinese recommendation query
• retrieved_at: retrieval month
• recipe_version: evidence/pollution recipe version
• docs: ranked documents, including URL, title, short excerpt, and—where applicable—the annotated real-brand anchor

FORGE distributes a rebuildable index rather than full third-party page bodies. Use python -m forge.fetch from the GitHub repository to re-fetch and extract source text, with a Wayback fallback.

Benchmark metric

The primary metric is the fooled rate, the mean of the binary recommendation indicator Rec(t, r). Rec = 1 when the fake-brand string or its prefix is a case-insensitive substring of the response. The metric uses no LLM judge. Higher fooled rate means greater vulnerability.

The evaluation harness and pollution recipe live in the GitHub repository:

git clone https://github.com/leoluolol/forge-benchmark
cd forge-benchmark
pip install -e .
python -m forge.eval leaderboard

Intended use and safety

FORGE is intended for reproducible, defensive research on retrieval poisoning, generative recommendation robustness, and mitigations. It should not be used to manipulate live recommendation systems, deceive users, impersonate real brands, or publish poisoned pages. Fake brands used in experiments should be synthetic and collision-free with real entities.

The benchmark covers Chinese product-recommendation queries and one documented entity-replacement attack recipe. Its results should not be interpreted as a general measure of model quality or security outside this setting. Source pages can change or disappear over time, so rebuilt evidence may differ from the original retrieval snapshot.

License

The FORGE-authored data layer—queries, taxonomy, results, and anchor annotations—is released under CC BY-NC-SA 4.0. Evidence bundles contain URLs and short display excerpts rather than full page bodies. Copyright and terms for the underlying third-party pages and excerpts remain with their respective owners. The separately distributed evaluation code in the GitHub repository is licensed under MIT.

Citation

@article{luo2026forge,
  title   = {One Polluted Page Is Enough: Evaluating Web Content Pollution in Generative Recommenders},
  author  = {Luo, Minghao and Chen, Liang},
  journal = {arXiv preprint arXiv:2606.13610},
  year    = {2026}
}