A SURVEY OF HALLUCINATION MITIGATION STRATEGIES IN LARGE LANGUAGE MODELS: FROM DECODING-TIME FIXES TO AGENTIC RETRIEVAL AND VERIFICATION

Abstract

Hallucination – the generation of fluent but factually unsupported or fabricated content – remains one of the most persistent obstacles to deploying Large Language Models (LLMs) in high-stakes domains such as healthcare, finance, and law. A wide range of mitigation strategies has been proposed, spanning several distinct points of intervention: constraining or calibrating the decoding process itself, fine-tuning or aligning models to prefer grounded outputs, augmenting generation with external retrieval, and verifying or correcting outputs after they are produced. This paper surveys hallucination mitigation across these four intervention points rather than focusing narrowly on any single pipeline. We discuss the underlying causes of hallucination identified in prior work, review representative techniques at each intervention point – including confidence based decoding constraints, instruction-tuning for faithfulness, hybrid sparse-dense retrieval with rank fusion, cross-encoder re-ranking, and Natural Language Inference (NLI)-based claim verification within agentic architectures – and compare the costs and guarantees each category offers. We further examine evaluation practices and benchmarks used to measure hallucination and identify a set of open problems, including the absence of standardized crosscategory benchmarks, the tension between validation strictness and unnecessary refusal, and the limited maturity of hallucination mitigation for multi-hop and cross document reasoning. We conclude with directions for future work that combines lightweight decoding-time safeguards with heavier, verification-based agentic pipelines depending on application risk.