Survey of Hallucination in Natural Language Generation

Summary

A comprehensive survey of hallucination in natural-language generation — defining the phenomenon, distinguishing intrinsic (contradicts the source) from extrinsic (unverifiable against the source) hallucination, mapping causes (training data, modeling, inference, decoding), and cataloging detection and mitigation techniques task by task (summarization, dialogue, MT, QA, data-to-text).

Contribution

The reference taxonomy that subsequent work in factuality, faithfulness, and grounded generation has built on. By organizing hallucination type × cause × task, it makes interventions comparable and exposes that "hallucination" is not a single phenomenon.

Method

Literature survey covering work through 2022. Categorizes papers by task and intervention class; tabulates detection metrics; identifies open problems.

Relevance to RISE

A RISE pipeline that drafts papers, retrieves literature, or generates citations is inherently exposed to both intrinsic and extrinsic hallucination — and many of the catalog's projects (paper-qa, open-scholar, storm) cite citation-grounding and retraction-awareness as core selling points precisely because of the failure modes catalogued here. The taxonomy is also useful for evaluating RISE systems: distinguishing fabricated results from unfaithful summaries from outdated facts requires the type × task breakdown.

Critique / open questions

• Predates the agentic turn — does not address compounding hallucination across multi-agent loops, or hallucination injected via tool outputs.
• The survey treats hallucination as undesirable; the framing does not engage with cases (e.g., hypothesis generation) where generating beyond the source is the point — relevant for RISE ideation stages.