Gradient Attribution
A class of gradient-based sample influence analysis methods that approximate the influence of training samples on model outputs through the gradient of the loss function with respect to model parameters. Its computational complexity grows linearly with data size.
Timeline (last 90 days)
In single-latent ablation experiments, the proportion of cases where the sign predicted by gradients matched the true sign reached 91%–97%, whereas the direction of change suggested by probes was correct only 17%–35% of the time.
Small-step steering experiments show that attribution estimates closely match actual steering effects.
All Facts (2)
In single-latent ablation experiments, the proportion of cases where the sign predicted by gradients matched the true sign reached 91%–97%, whereas the direction of change suggested by probes was correct only 17%–35% of the time.
50%UnverifiedSmall-step steering experiments show that attribution estimates closely match actual steering effects.
50%