Factuality detection encompasses computational approaches to identify when text contains false, inaccurate, or unsupported claims. This spans:
Hallucination detection in language models: Identifying when LLMs generate plausible-sounding but false information, including claims not supported by training data or contradicting known facts.
Long-form text verification: Assessing factuality of longer generated responses (QA answers, essays, code explanations, scientific synthesis) rather than single-sentence claims.
Tool-augmented verification: Using external tools (search engines, calculators, code interpreters, databases) to gather evidence and verify specific claim types.
Multi-granularity assessment: Detecting errors at different levels—word-level, sentence-level, statement-level, or response-level—enabling fine-grained error localization.
Key papers in this wiki
- 2023-chern-factool — Framework for detecting factual errors in LLM outputs across knowledge-based QA, code generation, math reasoning, and scientific writing; leverages external tools (Google Search, code interpreters, Python, Google Scholar) for evidence collection and verification