Hallucination Mitigation

Hallucination mitigation reduces unsupported claims; it does not make a model know truth in the abstract. The strongest controls combine grounding, citations, constrained answer formats, abstention, and targeted evaluation. In production, the goal is usually not “never be wrong”; it is “make unsupported claims detectable and less likely to reach the user.”

Mitigation across the pipeline

A practical pipeline retrieves evidence, instructs the model to answer only from that evidence, extracts claims, and checks each claim against cited spans. LLM-as-judge can help label semantic support, but deterministic checks should verify that cited source IDs were actually retrieved and that every required claim has a citation.

Mitigation can happen before generation through better retrieval, during generation through source-grounded prompts and schemas, and after generation through citation validation or abstention. Guardrails decide what happens when support is missing: revise, ask for more evidence, route to a human, or say there is not enough information.

Mitigation levers

Lever	What it reduces	Example
Better retrieval	missing or stale evidence	retrieve current policy version before answering.
Context discipline	evidence overload and distractors	pack only the top cited chunks into context construction.
Constrained format	unsupported free-form claims	require each answer bullet to cite a source ID.
Abstention policy	confident guesses	answer “not enough evidence” when no source supports the claim.
Claim checking	subtle contradictions	compare generated claims against cited spans.
Evaluation slices	hidden domain failures	test by policy version, language, jurisdiction, and topic.

An answer policy with checks

{
  "answer_policy": "If no cited source supports a claim, say 'I do not have enough evidence.'",
  "checks": ["source_id_seen", "claim_has_citation", "citation_support"],
  "fail_action": "revise_or_abstain",
  "unsupported_claim": {
    "claim": "Refunds are approved in two days.",
    "cited_source": "policy-9",
    "source_says": "up to five business days"
  }
}

The artifact separates policy, checks, and failure handling. source_id_seen prevents fabricated citations, claim_has_citation prevents unsupported assertions from passing silently, and citation_support checks whether the cited passage actually entails the claim. In the example, the generated claim says “two days” while the source says “up to five business days”, so the correct action is revision or abstention.

Realistic failure analysis

Suppose a support assistant says:

Enterprise refunds above 5000 EUR are automatically approved within two days.

The cited policy says:

Enterprise refunds above 5000 EUR require finance approval and may take up to five business days.

This is not a retrieval miss; the right evidence was present. It is a generation or citation-support failure. The fix is not necessarily a larger retriever. Better interventions are claim-level validation, a stricter answer schema, or a prompt that requires quoting the approval condition before summarizing it.

If the retrieved context instead contained only a general refund overview, the failure would be upstream retrieval or RAG evaluation. Separating these cases prevents teams from tuning the wrong component.

Measuring hallucination mitigation

Measure unsupported-claim rate, citation precision, citation recall for required claims, abstention accuracy, and answer usefulness. A system that abstains on every hard question has low hallucination but poor utility. A system that always answers has high utility but may be unsafe. The product target should define the trade-off.

Caveats

Retrieval can introduce false evidence. A fluent answer with citations can still be wrong if the cited passage is irrelevant, stale, or contradicted elsewhere. Reducing hallucination is therefore a system property, not a prompt trick. For high-stakes domains, unsupported answers should fail closed rather than be softened with vague disclaimers.

References

• Anthropic Claude docs: Reduce hallucinations
• Lewis et al., 2020, Retrieval-Augmented Generation
• OWASP Top 10 for LLM Applications

Nav

Section — Generative AI and Agentic Systems

← Citations RAG Evaluation →