Evaluating Synthetic Activations composed of SAE Latents in GPT-2

Nora Petrova,Giorgi Giglemiani,Chatrik Singh Mangat,Stefan Heimersheim,Jett Janiak

提交: 2024-09-28更新: 2024-10-18

TL;DR

We investigate how synthetic activations composed of SAE latents compare to real model-generated activations in GPT-2, revealing insights into the importance of SAE latents relationships and activation plateau characteristics.

摘要

关键词

Mechanistic InterpretabilitySAEsActivationsSAE Latents

评审与讨论

编辑台拒稿- Submission Desk Rejected by Program Chairs

2024-10-18

直接拒稿原因

Wrong formatting template.