Improving Factuality by Contrastive Decoding with Factual and Hallucination Prompts
Bojie Lv, Ao Feng, Chenlong Xie · 2024
Read Improving Factuality by Contrastive Decoding with Factual and Hallucination Prompts by Bojie Lv, Ao Feng, Chenlong Xie (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/pmc-11548250.txt.
Description
Bojie Lv, Ao Feng, and Chenlong Xie present a decoding method that makes language model answers more factual. Their 2024 paper tackles hallucination, where models produce text that is wrong or off input. Their method uses two prompts: one pushing for facts, one pushing for hallucination. Decoding then contrasts the two outputs. Answers favored by the factual prompt are boosted.
Tests covered multiple choice and text generation tasks. The method improved factual accuracy with no extra training. On a standard truthfulness data set, gains averaged 6.4 percent across four sizes of one model family. The sizes ran from seven to sixty five billion parameters. The authors pitch the approach for high reliability uses. Examples they name include medical diagnosis and legal work, where errors carry real cost.
Catalog record
| CREATOR | Bojie Lv, Ao Feng, Chenlong Xie |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 39 KB |
| VERSION | 1.0 |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |