Localizing "summarization" with GCM. We show steering results with a variant of difference-in-means steering on Qwen1.5-14B-Chat
Loading image…
We show localization + steering results for a summarization task on Qwen1.5-14B-Chat, where the original responses were 150-300 tokens long and the target responses were 500-600 tokens long. Post-intervention, the model produced long responses when steered on the localized attention heads.