LLM很强大,但有显著的局限——幻觉、知识截止、缺乏真正的推理、非确定性等。理解这些局限以及如何缓解它们对负责任和有效地使用LLM至关重要。
关键局限
✗ HALLUCINATION → generate plausible-sounding but FALSE information (confidently wrong) →
the most critical limitation (they predict likely text, not verified truth)
✗ KNOWLEDGE CUTOFF → only know up to their training date → no current info
✗ NO TRUE UNDERSTANDING/REASONING → pattern-match text; can fail at logic, math, novel
reasoning; no real comprehension
✗ NON-DETERMINISTIC → same prompt → different outputs (variability)
✗ NO REAL-TIME/external knowledge by default; limited CONTEXT window (can't hold unlimited info)
✗ Can be BIASED; can be manipulated (prompt injection); inconsistent
缓解局限
✓ HALLUCINATION → use RAG (ground in retrieved facts); verify outputs; cite sources;
don't trust blindly for facts; human review for important decisions
✓ KNOWLEDGE CUTOFF → RAG/tools for current info; provide context in the prompt
✓ REASONING limits → chain-of-thought prompting; break down tasks; use tools (calculators,
code) for math/logic; verify
✓ NON-DETERMINISM → design for variability; structured output + validation; lower temperature
✓ Validate/sanitize outputs; guardrails; human-in-the-loop for critical uses
