BibTeX Citation Hallucinations in Scientific Publishing Agents: Evaluation and Mitigation

LLMs with web search often hallucinate BibTeX citations, leading to pervasive field-level errors. This Action Pack guides you to evaluate and mitigate these errors by building specialized benchmarks and implementing validation layers, ensuring academic integrity in AI-assisted scientific writing.

intermediate2 hours5 steps

The play

Acknowledge BibTeX Hallucination Risks
Understand that LLMs integrated with web search frequently generate erroneous BibTeX citations. Recognize this as a critical problem for academic integrity in AI-assisted scientific writing tools.
Identify Evaluation Gaps
Realize that standard LLM evaluation benchmarks are inadequate for assessing citation accuracy when web search is involved. Current methods fail to account for the real-world context where these errors occur.
Design a Context-Aware Benchmark
Develop or adopt a specialized benchmark that incorporates web search capabilities. This benchmark should include a diverse set of papers (e.g., 900+ as suggested by research) to thoroughly test citation generation in realistic scenarios.
Categorize Field-Level Errors
Analyze generated BibTeX entries to identify and categorize specific field-level errors. Focus on common inaccuracies such as incorrect authors, titles, publication years, or journal names within the structured BibTeX data.
Implement Mitigation & Validation Layers
Integrate validation and correction mechanisms into your LLM agent's pipeline. This could involve post-processing BibTeX output against known databases, using rule-based checks, or prompting the LLM for self-correction based on identified error types to reduce or eliminate hallucinations.

Starter code

@article{smith2023ai,
  author = {Smith, J. and Johnson, A.},
  title = {AI in Scientific Publishing: A Review},
  journal = {Journal of AI Research},
  volume = {1},
  number = {1},
  pages = {1-10},
  year = {2024} % <-- Potential hallucination: year should be 2023
}

Source

Paperarxiv.org