Research on Hallucination Mitigation in Large Language Models

Abstract

The proliferation of large language models has brought the issue of hallucinations, which are outputs that deviate from inputs or fabricate information, to the forefront of concerns regarding reliability and deployment safety. This paper provides a systematic review of existing mitigation strategies and evaluation frameworks for hallucinations. Among mitigation approaches, supervised fine-tuning and reinforcement learning from human feedback have demonstrated moderate success; however, their heavy reliance on extensive, high-quality human annotations renders them costly and difficult to scale. On the evaluation side, this study examines several multimodal benchmarks, including POPE, MMHal-Bench, MM-Vet, and MMBench, which enable comprehensive assessment of model outputs through a blend of automated metrics and human judgment across multiple dimensions. Overall, substantial progress has been achieved in hallucination evaluation frameworks, yet scalable and cost-effective mitigation techniques remain elusive, particularly for hallucinations in complex multimodal settings, where further breakthroughs are urgently needed.