[KongchangAI]
PaperLABEL研究 / LABEL论文

LABEL

A paper that studies the asymmetry between drug indications and contraindications as recorded in publicly available text, and examines whether language models inherit this bias.

Timeline (last 90 days)

Nov 17

A study named LABEL analyzed how 574 drugs are described in Wikipedia and PubMed abstracts, finding that public text records contraindications far less than indications.

Unverified50%
Nov 17

In Wikipedia articles and PubMed abstracts, drug benefits are mentioned about twice as often as safety restrictions, with a within-drug odds ratio of 2.19.

Unverified50%
Nov 17

When the researchers provided the model with a retrieved passage that explicitly stated the answer, the model's utilization efficiency of safety facts was comparable to that of benefit facts.

Unverified50%
Nov 17

The root cause of the model's failures on drug safety lies in the scarcity of safety facts in free-text corpora, not in defects in the model's reasoning ability.

Unverified50%
Nov 17

The research team has released the code and data of the LABEL study for subsequent research to reproduce and extend.

Unverified50%

All Facts (5)

Source Articles