Hallucinations refer to instances where large language models generate text that appears coherent and confident but contains false, fabricated, or unsupported claims. Hallucinations are a critical limitation of LLMs in high-stakes applications:
Scope of hallucinations:
- Factual errors: claims contradicting known facts or training data
- Fabrication: entirely invented citations, statistics, or references
- Semantic inconsistency: claims contradicting earlier statements in the same response
- Temporal confusion: outdated information presented as current
Why hallucinations occur:
- Language models learn statistical patterns from training data without explicit knowledge representations
- Models optimize for next-token prediction rather than factual accuracy
- Confidence during generation does not correlate with truthfulness
- Long-form generation amplifies error accumulation
Detection and mitigation strategies:
- Retrieval augmentation: grounding generation in external sources
- Uncertainty estimation: teaching models to express confidence calibration
- Tool use: integrating external tools for verification
- Fine-tuning on factuality-annotated data
Key papers in this wiki
- 2023-chern-factool — Addresses hallucinations in LLM-generated text across multiple task domains; shows GPT-4 self-checking baselines miss 30-40% of factual errors compared to tool-augmented framework