Mean Difference Vector
The most commonly used direction construction method in refusal direction research. It is obtained by separately computing the means of residual stream activations for harmful and harmless prompts and subtracting them, yielding a vector pointing toward refusal, under the assumption that refusal behavior is maximally separable along that direction.
Timeline (last 90 days)
Refusal behavior is largely mediated by a single direction in the residual stream, typically computed as the mean difference vector between harmful and harmless prompts.
The standard practice aggregates all heterogeneous harmful prompts into a single uniform mean vector, which may conflate distinct refusal-related phenomena.
All Facts (2)
Refusal behavior is largely mediated by a single direction in the residual stream, typically computed as the mean difference vector between harmful and harmless prompts.
50%UnverifiedThe standard practice aggregates all heterogeneous harmful prompts into a single uniform mean vector, which may conflate distinct refusal-related phenomena.
50%