pyvene-interventions

Activation patching swaps representations between prompts to measure which layers, positions, and components change model predictions. You compare two prompts, patch a target layer, and determine whether attention, MLP, or block outputs drive the final answer.

Causal tracing restores clean activations into corrupted runs to locate where factual associations or decisions are stored. It maps causal effects by intervening on internal activations, testing causal hypotheses instead of relying on surface-level inspection.

Yes, you can learn low-rank or rotated interventions for distributed alignment search, steering, or counterfactual behavior. Trainable interventions allow you to steer model behavior by applying learned representations across target layers and components.

Yes, it applies to causal tracing, activation patching, interchange intervention training, and model steering across transformers and other PyTorch architectures. It requires declarative intervention configuration, layer and component targeting, and reproducible source-to-base activation swapping.

The best way is to map causal effects by intervening on internal activations using reproducible source-to-base activation swapping. This identifies which internal model activations cause specific behaviors, making it easier to test causal hypotheses instead of relying on guesswork.

Targeting specific layers and components is required for declarative intervention configuration and reproducible source-to-base activation swapping. It allows you to precisely measure whether attention outputs, MLP outputs, or block outputs are responsible for the final answer.