Mean-difference concept vector

anchor

Average the embeddings of examples that have a property and subtract the average of examples that do not. The resulting direction is treated as the model's internal representation of that property. Shows that a direction correlates with the property; it does not show the model uses it. The anchor is Turner et al., the activation-steering paper that the TerraMind concept vector report cites for this construction.

Findings

• TM-001A mean-difference vector between before and after image pairs acts as a transferable concept vector [primary]
• TM-002The mean-difference concept vector scores higher than trained classifiers under cosine similarity [primary]