In addition to the ability to generate fluent text in various languages,
large language models have been successful at tasks that involve basic forms of
logical "reasoning" over their context. Recent work found that selectively
removing certain components from weight matrices in pre-tr