bots-bench / refusals

When a model refuses to investigate

A task can end three ways without a solve: the model refuses it, the model abstains after investigating, or the model answers wrong. This page counts refusals and keeps them apart from honest abstentions. Wrong answers live on the leaderboards.

Only aggregate results are public. Prompts, answers, and traces stay private.

Reading the results

Three different ways to not answer

The counts below depend on this distinction. A refusal and an abstention look similar in a score column, but they mean opposite things about the model.

Every measured run

Every published run

Click a heading to re-sort. Runs that kept no record stay at the bottom either way.