RetrieverGuard: Empowering Information Retrieval to Combat LLM-Generated Misinformation

Chuwen Chen, Shuai ZhangNAACL 2025 · DOI

Introduction

Large Language Models (LLMs) have revolutionized information retrieval (IR) by enabling systems to understand complex queries with unprecedented nuance. However, these models are a double-edged sword: they are equally capable of generating "hallucinations"—text that is stylistically indistinguishable from human writing but factually incorrect. When these deceptive documents enter a corpus, standard retrieval models often fail to distinguish between authentic evidence and AI-generated misinformation, compromising the integrity of the entire information pipeline.

The Challenge of AI-Generated Misinformation

The primary difficulty in modern IR is that LLM-generated fakes are designed to mimic the logical structure and style of real documents. In dense retrieval systems, which rely on semantic embeddings, these "hard negatives" reside in extremely close proximity to authentic documents.

T-SNE visualization of real vs. fake corpora showing significant semantic overlap.

As shown in the T-SNE visualization above, real and fake texts overlap almost entirely in standard embedding spaces. This "semantic overlap trap" explains why traditional retrievers are easily fooled; to the underlying mathematics, a sophisticated lie looks identical to the truth. Experiments indicate that the introduction of LLM-generated fakes causes an average 11% decrease in NDCG scores for standard retrieval models.

The RetrieverGuard Framework

To address this vulnerability, we introduce RetrieverGuard, a self-supervised fine-tuning framework designed to train retrievers to recognize the subtle "textual rules" that characterize LLM-generated misinformation. RetrieverGuard builds upon unsupervised contrastive learning, specifically utilizing a dual-negative sampling strategy to sharpen the model's discriminatory boundaries.

The RetrieverGuard training pipeline featuring dual-negative sampling.

The framework operates through three core mechanisms:

  1. Data Augmentation with Dropout Noise: The model processes the same text twice with different random dropout masks to create "positive pairs," learning that these internal representations should remain consistent despite minor noise.
  2. Dual Negative Sampling: The model is exposed to "Soft Negatives" (random documents from the batch) and, crucially, "Hard Negatives." These hard negatives are deceptive versions of real documents generated by an LLM (e.g., GPT-3.5-turbo). By distorting key facts while maintaining style, these fakes force the retriever to learn fine-grained distinctions.
  3. Layered Contrastive Loss: A modified loss function maximizes the similarity between original text and its noisy version while simultaneously pushing away both random and deceptive negatives.

Experimental Evaluation

RetrieverGuard was evaluated using the BEIR (Benchmarking Information Retrieval) benchmark across four diverse domains: scientific fact-checking (SciFact), multi-hop QA (HotpotQA), medical retrieval (NFCorpus), and climate change fact-checking (Climate-FEVER).

To ensure robust testing, two separate deceptive corpora were created: Fake1, used as hard negatives during training, and Fake2, an unseen validation set used to test the model's ability to generalize its detection capabilities to novel misinformation styles.

Key Results and Insights

RetrieverGuard consistently outperformed vanilla models in the presence of misinformation. For instance, on the Climate-FEVER dataset, where the original Stella-1.5B model's accuracy dropped from 34.7 to 25.2 upon the introduction of fakes, RetrieverGuard maintained a significantly higher score of 30.2.

Key takeaways from the study include:

  • Fighting Fire with Fire: Using AI to generate training adversaries is the most effective way to defend against AI-generated misinformation. By observing how LLMs distort facts, the retriever learns the "fingerprints" of deception.
  • Self-Supervision for Scalability: Because misinformation evolves rapidly, RetrieverGuard’s reliance on self-supervision rather than human-labeled data makes it practical for real-world, large-scale deployment.
  • The Scale Effect: Larger models (e.g., 1.5B parameters) showed a greater latent capacity for spotting logical inconsistencies compared to smaller 400M parameter versions.

Conclusion and Future Work

RetrieverGuard demonstrates that IR systems can be hardened against the threat of LLM hallucinations through targeted contrastive training. While the computational cost of training with hard negatives is higher, the resulting resilience is essential for maintaining trust in automated information systems. Future research will focus on mitigating overfitting to specific LLM styles and optimizing the framework for low-resource environments, potentially expanding its application to social media moderation and fraud prevention.

Summary updated

BibTeX
@inproceedings{chen2025retrieverguard,
  author = {Chuwen Chen and Shuai Zhang},
  title = {{RetrieverGuard: Empowering Information Retrieval to Combat LLM-Generated Misinformation}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
  doi = {10.18653/v1/2025.findings-naacl.249},
}
Figure