幻觉(Hallucination)是模型生成了流畅但缺乏事实依据或与事实冲突的内容。要识别它,必须把“像真”与“有证据”拆成两条独立判断。A hallucination is fluent model output that lacks factual support or
conflicts with facts. Detecting it requires separating plausibility
from evidence.
01 · 先从日常问题开始01 · Start with an everyday question
同事讲得特别顺,却拿不出关键文件——你会因为语气自信就签字吗?A colleague sounds completely confident but cannot produce the key
document. Would you approve on tone alone?
∅
02移走关键证据,盯住不断线的流畅度Remove Key Evidence and Watch Fluency Stay Intact
拖动保留证据数drag evidence count
语言流畅度Fluency
—
style score
锚定主张Anchored claims
—
of 3 claims
无锚点警报No-anchor alerts
—
claim-level
03四个零件解释“为什么仍像真的”Four Parts Explain Why It Still Sounds Plausible
把刚才的变化拆开Map the change to parts
01 · EVIDENCE
证据片段Evidence snippets
三片来源各自支持一个可核查主张;移走后,来源本身不会被模型自动补回。Three source snippets each support one checkable claim. Removing
one does not make the source magically return.
02 · CLAIM
可核查主张Checkable claims
回答里的日期、数量和归因需要逐项检查,而不是给整段文字一个绿色勾。Dates, quantities, and attributions need claim-level checks, not
one green tick for an entire paragraph.
03 · FLUENCY
语言流畅度Fluency
句子是否顺滑与是否有依据是两条轴;流畅度可以在证据断开后保持很高。Smooth language and factual support are separate axes. Fluency
may stay high after evidence disappears.
04 · ALARM
无锚点警报No-anchor alarm
只要主张没有来源连接,就明确降级,不用语气自信替代证据。Any claim without a source link is downgraded explicitly;
confident tone never substitutes for evidence.
04最小机制:流畅度不参与证据计数Minimal Mechanism: Fluency Is Not Evidence
anchor rate = linked / 3
anchor rate = linked claims / 3
本页固定三个主张。每移走一片关键证据,就断开一条主张—来源连接;流畅度分数故意保持不变。This lab fixes three claims. Each removed source breaks one
claim-to-source link while the fluency score deliberately stays
unchanged.
核对主张Check each claim逐项寻找来源look for a source per claim
断线就报警Warn on broken link不把语气当证据never treat tone as evidence
05边界实验:证据清空后仍给整段绿灯Boundary Test: Keep the Paragraph Green with No Evidence
主动制造失败Create a failure
制造一个“流畅但零锚点”的回答Create a fluent answer with zero anchors
一键移走三片证据。若系统仍只看语言质量,这段回答会显得完全正常;逐主张检查则会亮出三个警报。Remove all three sources. A style-only check still sees a normal
answer, while claim-level checking raises three warnings.
尚未执行。先在上方改变主变量。Not run yet. Change the main variable above first.
✗ 写得顺,所以大概率是真的。It reads smoothly, so it is probably true.
→ ✓ 流畅度与证据支持分开测。Measure fluency and support separately.
✗ 整段有一个引用就算有依据。One citation supports the whole paragraph.
→ ✓ 每个可核查主张都要有对应来源。Every checkable claim needs its own source.
✗ 删掉证据后继续显示绿色。Keep the claim green after its source is removed.
→ ✓ 连接断开就降为“未证实”。Downgrade to unverified when the link breaks.
06 · 一句话带走06 · One line to keep
幻觉之所以危险,是因为关键证据被移走后回答仍可能很流畅;可靠系统必须逐主张检查锚点,断线就报警。Hallucinations are dangerous because an answer may stay fluent
after key evidence is removed. Reliable systems check anchors
claim by claim and warn on broken links.