PURE: Turning Polysemantic Neurons Into Pure Features by Identifying Relevant Circuits

Open in new window