# Reducing Hallucination — Large Language Models

Source: https://www.geekswithgeeks.com/en/llms/e-hallucination

> Apply grounding, constraints and verification.

## Make the truth easy to find and the lie easy to catch

Hallucinations cannot be removed entirely, only reduced and caught. Practical measures: **ground** answers in supplied sources (RAG) and require quotes or citations; instruct the model to **abstain** when the context is insufficient; use **low temperature** for factual tasks; **constrain output** to schemas or allowed values; **verify** with code or a second pass (check that cited passages really contain the claim, run generated code and tests, validate numbers); and **show uncertainty and sources** in the UI so users can check. For high-stakes areas such as medicine, law or finance keep a human in the loop.

**Quiz:** Which measure helps catch a hallucinated citation?

- [ ] Removing the system prompt
- [ ] Raising the temperature
- [ ] Using shorter words
- [x] Checking in code that the cited passage contains the claim

*Answer:* Checking in code that the cited passage contains the claim. Automated verification against the source catches unsupported claims.