Mitigating Fabrication in Multi-Stage LLM Pipelines for Hiring: An Empirical Evaluation of Prompt Guardrails and Human-in-the-Loop Checkpoints

Research indicates that multi-stage Large Language Model (LLM) pipelines used in hiring processes are highly prone to fabricating credentials, inflating qualifications, and inventing experience. An empirical evaluation of mitigation strategies found that while prompt guardrails significantly reduce the density of fabrications, they do not eliminate them. The study suggests that human-in-the-loop checkpoints are crucial for effective mitigation.

Why it matters

The findings highlight a significant risk of 'hallucination' or fabrication when deploying LLMs in critical, decision-making processes such as recruitment. This directly impacts the integrity of data used for talent acquisition and necessitates robust controls to prevent erroneous or misleading information from influencing human capital decisions.

Key insights

• Fully automated multi-stage LLM hiring pipelines produced at least one unsupported claim in 96.7% of outputs, with an average of 6.80 fabrications per output.
• Implementing prompt guardrails reduced the density of fabrications by 86% (from 6.80 to 0.92 per output).
• Despite prompt guardrails, 50.0% of outputs still contained at least one fabrication, indicating prompt-level mitigation alone is insufficient.
• The study implies that human-in-the-loop checkpoints are necessary to address residual fabrication rates in LLM-driven hiring processes.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.26171