Hallucination Control in Large Language Models: A Systematic Literature Review and Research Framework

This project is the research outcome of a graduate course in the first semester of 2026 at the School of Electrical and Computer Engineering, University of Campinas (Brazil). It conducts a systematic literature review of over 300 studies on LLM hallucination control from 2022 to 2025, constructing a complete knowledge framework across six dimensions: hallucination classification, cause analysis, detection techniques, mitigation strategies, evaluation benchmarks, and future challenges. It aims to address hallucination issues in LLM applications in high-risk fields such as healthcare and law, providing comprehensive references for researchers and practitioners.

Research Background

LLMs have made significant progress in recent years, but hallucination issues restrict their applications in high-risk fields like healthcare and law. Incorrect outputs may lead to severe consequences, making it a key challenge for AI safety.

Core Concept Distinction

• Hallucination: Generated content is irrelevant to input context (rootless)
• Factuality: Generated content is inconsistent with verifiable world knowledge (false facts)

RAG can reduce hallucinations but has limited improvement on factuality errors (if the retrieved documents themselves are incorrect).

Hallucination Classification

• Factual Hallucination: Conflicts with world knowledge (verifiable/unverifiable)
• Faithfulness Hallucination: Inconsistent with input context (input conflict/context conflict/logical conflict)

Cause Analysis

• Data Level: Errors in pre-training corpus/repeated reinforcement/knowledge timeliness
• Training Level: Maximum likelihood estimation prefers plausibility over truth/alignment tax of RLHF/unstable knowledge editing
• Inference Level: Attention limitations/long-sequence information loss/sampling randomness

Detection Techniques

• Uncertainty Estimation: Semantic entropy/self-consistency check/confidence calibration
• External Validation: Fact-checking/RAGAS framework/FACTSCORE
• Internal State: HalluShift/attention visualization

Mitigation Strategies

1. Training Optimization: SFT/RLHF/knowledge editing
2. Architecture Improvements: RAG/attention enhancements/multimodal fusion
3. Prompt Engineering: Chain of Thought/self-consistency decoding/few-shot learning
4. Post-Generation Control: External verification/LLM-as-a-Judge/post-editing
5. Interpretability: Uncertainty quantification/confidence calibration/internal state analysis
6. Agent Systems: Multi-agent collaboration/reflexive RAG/self-refinement

• TruthfulQA (2022): 817 adversarial questions, focusing on factuality
• HaluEval2.0 (2024): 8770 questions covering 5 domains, highly comprehensive
• FaithBench (2024): Manually annotated, evaluating hallucinations in summarization tasks
• HalluLens (2025): Dynamic benchmark, strictly distinguishing hallucinations from factuality
• FACTSCORE/RAGAS: Fine-grained claim detection, no manual annotation required

Literature Research Methods

1. Retrieval: Covering AI conferences and journals from 2022 to 2025
2. Classification: Categorizing over 300 studies into six dimensions
3. Comparison: Evaluating trade-offs like computational cost and applicable scenarios of methods
4. Critical Review: Summarizing achievements and pointing out limitations

Experimental Expansion Plan

• Compare at least two mitigation strategies on HaluEval2.0
• Conduct quantitative evaluation using open-source LLMs (Llama/Qwen)
• Use metrics like AUROC/accuracy/hallucination rate

1. Layered Defense: Combine data cleaning, prompt optimization, RAG enhancement, and post-generation verification
2. Domain Adaptation: High-risk fields (e.g., healthcare) require mandatory fact-checking and manual review
3. Continuous Monitoring: Track real-world performance after deployment and correct issues promptly
4. User Education: Clearly inform users of AI output limitations, especially in high-risk scenarios

Hallucination is a core bottleneck for the widespread application of LLMs, and this review provides a comprehensive knowledge map. Future breakthroughs are needed in:

• Developing reliable dynamic benchmarks (e.g., HalluLens)
• Deeply understanding the neural mechanisms of hallucinations
• Designing alignment methods that avoid alignment tax
• Building interpretable uncertainty quantification frameworks

Only by systematically solving hallucination issues can LLMs become trustworthy AI assistants.