Notifications
Clear all
多语言国际区 International
1
帖子
1
Users
0
Reactions
16
查看
Topic starter
31/10/2026 9:57 上午
[Official FAQ] Corrections and follow-up questions welcome.
You can reduce them, not eliminate them. Local open models still invent plausible falsehoods.
Mitigations:
- RAG with citations and an explicit “say you don’t know” policy.
- Lower temperature for factual tasks; prefer instruction-tuned models.
- Tool use for math, DB lookups, and browsing when appropriate.
- Keep a hallucination-focused eval set before each upgrade.
Never treat ungrounded model text as legal, medical, or financial advice.