Me LLMs ke hallucinate, da yaya za a iya rage shi?

Question

Accepted Answer

LLM **hallucinate** a lokacin da ya samar da rubutu wanda ke bayyana da karfin yake, amma ba daidai ba ne ko ya yi ni aiki. Don farahinta, dole ne ka fahimta abin da modelar take yi a gaske.

## Yadda LLM ke aiki (kadan kawai)

A wurin da aka nuna, LLM shine **next-token predictor**. Bayan rubutu da ya kasance, yana hasashen wadda wadda waje mai yuwuwa (kalma/ratsin kashi) dangane da **statistical patterns da aka koyo a lokacin training** — ba ta hanyar binciken gida ba.

```text
input: "The capital of Australia is"
model: P(next token) → "Canberra" 0.71, "Sydney" 0.18, ...
→ samples a token, appends it, repeats
```

## Me yasa hallucinations suka faru

- **Yana samar, ba karbare ba.** Modelar ta samar da *wacce mai yuwuwa* rubutu, ba *tabbatacciyar* gida. Fluency da gaskiya suna yawa.
- **Babu jiyya gaskiya a ciki.** Babu komai a cikin modelar wanda ya damje sakonnin zubar ne tare da ainihi.
- **Ragi suka cika da karfin.** Lokacin da bayanai na training suka rami, da damuwa, ko sakaci (rare APIs, abubuwan jiya-jiya, mutane masu abin koyarwa), modelar tana bayyana statistically likely continuation — wanda zai iya zama rashin gida.
- **Karfin ba gaskiya ba.** Kasua wanda aka yi waje yayi daidai da gida na ainihi.

## Yadda za a rage shi

- **Ground tare da RAG / sources.** Karbe bayanai masu alaka da saka su a cikin prompt, don modelar ta amsa *daga* tunani na gida jiya.
- **Bugi don citations** da jiyya — kasua da aka yi waje ya bayyana hallucination.
- **Rage zafin jiki** don ayyukan gaskiya, don modelar ta zaɓi high-probability (fiye da karfi) tokens.
- **Ba da izini "Ba na sani."** Ka ba umarnin modelar in tabbata cewa in ba shi da jiya, ba da damuwa ta yi aiki.
- **Tabbatar tare da kayan aiki.** Aiwatar da code, bugi calculator/search/database, ko kira API jiya ga asircewa da modelar.
- **Ware scope.** Takaice, da aka iyakance prompts suna rage hallucinate fiye da bure-bufe.

## Me yasa ya matukar mahimmanci

LLMs suna kasua a cikin aiki wanda gaske masu amfani suke asircewa. Saboda modelar **ta samar da rubutu mai yuwuwa jiya karbare gida**, hallucination shine yanayin asali, ba occasional bug ba — don haka engineers dole ne su yi tsara kewaye. Sanin *me yasa* ya faru (next-token prediction tare da babu jiyya gaskiya) ya gaya maka *yadda* za a rage: grounding, citations, rage temperature, izini don rashin aiki, da waje tabbatar. Damuwa da LLM sakonnin azaman draft wanda a duba — ba oracle ba — shine bambanci tsakanin tamkar aiki da confident liability.