diff --git a/README.md b/README.md index 1a8c3d6..7c6d79b 100644 --- a/README.md +++ b/README.md @@ -14,6 +14,8 @@ **Why use it?** Prompting is fragile. System prompts get ignored. Jailbreaks work. AntiPaSTO trains directly on the model's internal representations, measuring and modifying what the model actually computes rather than what it says it will do. On the DailyDilemmas benchmark, it outperforms prompting on small models (≤4B) and complements arithmetic steering methods on larger ones. +map + ![Bidirectional control](docs/img/fig_bidirectional_demo.svg)