Reducing Hallucinations in LLM-based Scientific Literature Analysis Using Peer Context Outlier Detection

Daniel Xie, Maxwell J. Jacobson, Adil Wazeer, Haiyan Wang, Xinghang Zhang, Yexiang XuearXiv 2026 · arXiv

TL;DR

Large Language Models (LLMs) are powerful tools for parsing scientific literature but are notoriously prone to generating plausible-sounding lies, known as hallucinations. This paper introduces Peer Context Outlier Detection (P-COD), a method that flags these errors by cross-referencing an LLM's extractions against semantically similar "peer" papers. By turning cross-document validation into a prioritized ranking tool for human review, P-COD achieves up to 98% precision in spotting data anomalies across multiple scientific domains.

What Problem Does This Paper Address?

When researchers use LLMs to automate data extraction from massive scientific corpora, they run into a major roadblock: hallucinations. Because LLMs predict text based on statistical patterns rather than verified facts, they frequently extract incorrect numerical values.

Current mitigation strategies—such as prompt engineering, Chain-of-Thought (CoT) prompting, and Retrieval-Augmented Generation (RAG)—all share a fundamental flaw: they evaluate documents in complete isolation. They fail to exploit corpus-level trends or inter-document relationships. In scientific research, papers utilizing similar methodologies and experimental setups should generally yield similar results. Without a mechanism to cross-validate these findings across the literature, incorrect data goes unnoticed, forcing researchers to manually verify every single extraction to prevent errors from propagating downstream.

Methods

The authors propose Peer Context Outlier Detection (P-COD), a human-in-the-loop framework that validates extracted data against its semantic peers. The workflow consists of three main steps:

  1. Semantic Text Embedding: Instead of relying on simple keyword matching (like TF-IDF), P-COD uses OpenAI’s text-embedding-3-large model to map papers into a high-dimensional vector space. Papers with similar topics, methods, and terminology naturally cluster together, forming distinct "peer groups."
  2. Surprising Score Calculation: For any extracted numerical value, P-COD calculates a "Surprising Score" ($S_p$). This score mathematically combines the semantic similarity of neighboring papers (using standard cosine similarity) with the normalized experimental deviation of the extracted value from those peers.
  3. Threshold-Based Flagging: If a paper's Surprising Score exceeds a tunable threshold ($T$), it is flagged as a potential hallucination. Rather than acting as an autonomous filter, P-COD outputs a ranked list, allowing human experts to focus their manual verification efforts where uncertainty is highest.

Datasets

To evaluate P-COD, the authors curated datasets using seed papers from the established scientific papers dataset (sourced from arXiv and PubMed):

  • Multi-Domain Dataset: A collection of 168 synthetic papers spanning 6 distinct fields: Computer Science, Physics, Biology, Chemistry, Materials Science, and Environmental Science. To create a reliable ground truth for validation, the authors strategically corrupted 25% of the papers by introducing values 3–5× outside their normal domain ranges.
  • Large-Scale Single-Domain Dataset: A 200-paper Computer Science dataset divided into 8 specialized sub-fields (e.g., Deep Learning Optimization, Reinforcement Learning, Blockchain, Data Mining) to test within-domain precision and scalability.

Key Insights and Ideas

  • The Power of Contextual Outliers: Traditional outlier detection looks at statistical deviations across an entire dataset (like global Z-scores). P-COD's core "aha moment" is realizing that a numeric outlier globally might be perfectly normal within its specific niche, and vice versa. Contextualizing outliers within semantic peer groups drastically reduces false positives.
  • Prioritization Over Autonomous Filtering: Instead of trying to completely replace humans—which risks accidentally deleting valid, groundbreaking anomalies that legitimately buck the trend—P-COD acts as an intelligent prioritization layer. It ensures that the most suspicious, "surprising" data points get human eyes first.

Key Results

  • High Precision Across Domains: P-COD achieved an average precision of 88.7% across all tested domains. It hit 98.0% precision in the large-scale CS dataset (correctly identifying 49 out of 50 corrupted papers) and achieved a perfect 100% precision in both the Chemistry and Environmental Science benchmarks.
  • Baseline Domination: P-COD significantly outperformed standard isolated LLM approaches. Across all domains, standard prompt engineering averaged a meager 55.9% precision, while Chain-of-Thought (CoT) prompting averaged 66.9% precision.

Limitations and Future Work

  • Methodological Heterogeneity: P-COD assumes that semantic similarity implies methodological similarity. In highly diverse or fragmented fields (like Physics, where P-COD's precision dropped to 66.7%), papers might use completely different experimental setups despite using similar terminology. This causes legitimate data variations to be falsely flagged as errors.
  • Future Directions: The authors plan to explore domain-specific embedding models (like SciBERT variants) and dynamic thresholding. They also aim to introduce multi-modal integration to extract and validate data directly from figures and tables, alongside temporal analysis to help distinguish genuine scientific breakthroughs from LLM errors.

Summary updated

Abstract

Reducing hallucinations in Large Language Models (LLMs) is essential for accurate data extraction from large text corpora. Current methods, like prompt engineering and chain-of-thought prompting, focus on individual documents and fail to consider relationships across a corpus. This paper introduces Peer Context Outlier Detection (P-COD), which uses inter-document relationships to improve extraction accuracy in scientific literature summarization, where papers with similar experiment settings should draw similar conclusions. By comparing extracted data to validated peer information within the corpus, we adjust confidence scores and flag low-confidence results for expert review. Our experiments demonstrate up to 98% precision in outlier detection across 6 scientific domains, reducing hallucinations and letting researchers focus on genuinely ambiguous cases.
BibTeX
@misc{xie2026reducing,
  author = {Daniel Xie and Maxwell J. Jacobson and Adil Wazeer and Haiyan Wang and Xinghang Zhang and Yexiang Xue},
  title = {{Reducing Hallucinations in LLM-based Scientific Literature Analysis Using Peer Context Outlier Detection}},
  year = {2026},
  eprint = {2604.01461},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.01461},
}
Figure