How to Reduce Hallucinations with Better Prompts

Assembling a literature review on urban noise policy means reconciling 62 papers whose numbers disagree with each other, and the dangerous failure is not an answer you can spot as wrong. It is a fluent sentence about night-time noise limits that cites an author, a journal, and a publication year that were never written. A model asked to sound complete will happily supply all three.

This workflow does not beg the model to be careful. It restructures the task so invention becomes expensive and checking becomes cheap: inventory every claim, tag each one with its evidence status, run an adversarial audit against itself, then demand a locator you can verify in ten seconds. The running example is a review of how cities regulate road and rail noise near housing, led by a researcher named Dr. Priya Raghavan, but the pattern fits any synthesis where a false citation can sink the work.

What a Finished Draft Looks Like

A trustworthy review draft has three properties you can inspect. Every factual sentence carries a tag stating where it came from. Every number is traceable to a specific table or paragraph, not just a paper title. And the draft flags its own weak spots instead of smoothing them over with confident prose.

Concretely, you want a document where a claim like “night-time road noise above 55 dB is linked to a 9 percent rise in self-reported sleep disturbance” arrives with a source tag, a confidence label, and a note that says which of the 62 papers it came from. If any of those three is missing, the sentence is a liability, not a finding.

There is a second reason to front-load structure. Hallucinations cluster around specifics, and a review is almost entirely specifics: thresholds, percentages, city names, directive numbers. If you design the prompt so that each specific is a separate field with a required status, the model has far fewer places to hide a confident guess.

If you already work with structured summaries, the habits in our research summary prompt settings pair well with the audit steps below.

Step 1: Build a Claim Inventory Before You Ask for Prose

Never ask for a polished paragraph first. Ask for the raw claims, separated from the writing, so you can see how many there are before you dress them up.

You are a research assistant supporting a literature review on urban noise policy.
From the source material I will paste, extract every factual claim that could appear in the review.
For each claim output one line with these fields:
- claim_id
- claim_text
- claim_type (statistic, policy_fact, causal_relation, definition)
- source_present (yes/no)
- locator (page, section, or table if the source states it; otherwise "MISSING")
Do not paraphrase into elegant prose. Plain lines only. If a claim has no apparent source, still list it and mark source_present as no.

You get a flat list of 20 to 40 atomic claims instead of a page of text. The value is visibility: a claim with no locator is stripped of its paragraph disguise and stands alone as something you must either source or delete.

Step 2: Force an Evidence Tag onto Every Sentence

Write the draft only after the inventory exists, and require the model to keep the tags attached while it writes.

Rewrite the claim inventory into review prose, one short paragraph per theme.
Rules:
1. After every factual sentence, append the claim_id in square brackets.
2. If a sentence combines two claims, give both ids.
3. Ban any sentence that has no claim_id behind it. If you feel an unsupported sentence is needed, write it and mark it [UNSOURCED] so I can see it.
4. Keep a plain academic tone. No filler transitions like "it is important to note".

You get a draft where every assertion is chained to the inventory. The [UNSOURCED] tag is the whole point: it turns an invisible invention into a marked item you can delete or chase down.

Step 3: Run an Adversarial Self-Audit

Now make the model attack its own draft. A single prompt asking it to critique its work catches generic problems; a structured attack catches specifics.

Act as a hostile peer reviewer whose only job is to find invented or unsupported content in the draft below.
Work through it sentence by sentence and list:
- any statistic with no locator,
- any named study, author, or policy document that might not exist,
- any causal claim stated more strongly than the evidence allows,
- any number that is suspiciously round or suspiciously precise.
For each problem give the sentence, the reason you doubt it, and a safer rewrite that removes the unsupported detail. Do not defend the draft.

You get a list of twenty-plus specific doubts rather than a vague “looks fine”. The instruction to remove unsupported detail matters: a rewrite that keeps the invented number but softens the verb is still a hallucination.

One rule keeps this honest: let the model attack, but never let it repair. Give the audit its own turn, read it, and only then decide what survives. If you ask it to critique and rewrite in a single pass, it tends to keep the claim and rephrase the sentence, which is exactly the failure you are trying to catch.

Step 4: Demand a Locator You Can Check by Hand

A citation is not proof. Push the model to quote the exact sentence it relied on, so you can open the paper and confirm it in one search.

For each claim that survived the audit, give me a verification block:
- claim_id
- the exact quoted sentence from the source that supports it, in quotation marks
- the section or table number
- a one-line note on whether the quote fully supports the claim or only partly.
If you cannot produce a quoted sentence, write "NO QUOTE AVAILABLE" and stop. Do not invent a quote.

You get quotes you can paste into a search box. Two outcomes are useful here: a real quote you can locate, or an explicit NO QUOTE AVAILABLE that tells you the claim was never grounded. Both beat a tidy reference list you cannot trace.

For the single-article case, the extraction logic in summarize an article with ChatGPT is a good companion to this step.

Step 5: Cross-Check the Numbers and the Named Things

Statements of a general kind survive on tone. Numbers and proper nouns do not, so treat them as a separate pass.

List every number that appears in the draft in a table with these columns: the number, what it measures, the unit, the claim_id it belongs to, and the source locator.
Then list every proper noun, meaning every author name, study title, policy name, agency, and place, with the same style of checklist.
Mark anything you are not certain exists with a warning. For night-time noise, remember the review only covers road and rail sources, not aircraft.

You get two short checklists that concentrate the risk. A claim that “noise above 55 dB at night raises sleep disturbance” is easy to gloss; cross-checking the 55 and the 9 percent separately is what catches a number that drifted between papers.

Step 6: Label Confidence in the Final Version

The last pass produces the document you actually hand in, with its uncertainty made explicit instead of hidden.

Produce the final review section with a confidence label on each paragraph: HIGH, MEDIUM, or LOW.
HIGH means multiple sourced claims agree. MEDIUM means a single source or a partial quote. LOW means the claim survived only as inference.
End with a short "Open Questions" list of claims you removed or could not source, so the reader knows what was excluded and why.
Keep the tone plain. Preserve the claim_id tags in a version I can hide before publication.

You get a submission-ready section plus an honest appendix. The Open Questions list is the part most writers cut, and it is the part that protects you: it shows a reviewer you know exactly where your own draft is thin.

Final Deliverables Checklist

  • A claim inventory of 20 to 40 atomic claims, each with a type and a locator status.
  • A tagged draft where every factual sentence carries a claim_id.
  • An adversarial audit listing invented or unsupported content.
  • Verification blocks with real quotes or an explicit NO QUOTE AVAILABLE.
  • A number checklist and a proper-noun checklist.
  • A confidence-labeled final section plus an Open Questions appendix.

Where Anti-Hallucination Workflows Fail

Failure 1 — asking for prose before claims. Write the inventory first. Once the model has a beautiful paragraph, it will defend it rather than audit it.

Failure 2 — trusting a reference list. A formatted citation is not evidence. Demand the quoted sentence and the table number, or assume the source is invented.

Failure 3 — accepting soft rewrites. “Evidence suggests” over an unsupported number is still unsupported. The fix is to cut the number, not to weaken the verb.

Failure 4 — skipping the proper-noun pass. Invented authors and invented policy names are the hardest hallucinations to notice because they read exactly like the real ones.

To sharpen the reasoning behind each step, see our guide to the chain-of-thought prompt.

Sources & Notes

The inventory-then-audit sequence is our own framing of standard source-checking practice, reviewed on 2026-10-09. Dr. Priya Raghavan and the 62-paper review are illustrative. The noise thresholds mentioned are widely cited reference points, but always confirm the exact figure against the original document before you cite it. Language models change between versions, so re-run the audit any time you switch models or paste in new source material.

声明:1.本站所有文章,如无特殊说明或标注,均为本站原创发布。本网站中【文本生成/文生图】使用了DeeSseek人工智能生成内容技术, 算法来源说明:本网站未自行开发、训练或部署深度合成算法,所使用的人工智能算法来自:DeepSeek、Chatgpt、Google Gemini。 2.为了网站加载速度优化,本站预览图片进行压缩处理,所以有些不是太清晰,并不影响AI提示词生图、AI设计等内容。3.任何个人或组织,在未征得本站同意时,禁止复制、盗用、采集、发布本站内容到任何网站、书籍等各类媒体平台。如若本站内容侵犯了原著者的合法权益,可联系我们进行处理。--GUUNN.COM