concept
active
concept:counterfactualCounterfactual
The output value a model produces when an interchange intervention forces certain variables to take values from source inputs.
Neighborhood — ranked by edge-count
Methods (1)
method
- Interchange Interventionassociated_withFundamental operation for causal abstraction analysis; forces neurons to take values from source inputs to create counterfactuals.
Concepts (7)
concept
- Counterfactual Behaviorrelated_toThe behavior that would have occurred had the value of a causal variable been different while everything else remained the same; used as training labels in DAS/MAS.
- Counterfactual Hypothesisrelated_toAbility to entertain competing hypotheses within one inference engine; proposed hallmark of mindful inference
- Counterfactual Staterelated_toThe state a neural network is placed in when its activations are modified via intervention
- Counterfactual Thinkingrelated_toEncoding possible future or alternative states; bioelectric patterns can represent an alternative morphological target even in an intact animal.
- Counterfactual maintenancerelated_toThe mental effort of holding models of how things could/should be different from actuality, contributing to compression stress.
- Counterfactual claimsrelated_toTrue/false claims that a prompt edit shifts a behavior rate by at least 30pp; basis for all evaluation labels.
- Counterfactual Representationrelated_to
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Lens for evaluating explanations by whether they help predict behavior on counterfactual inputs.
- Pre-recorded latent vector encoding the expected causal variable values post-intervention; used as ground truth in the CLMAS auxiliary loss.
- Auxiliary training objective from Grant (2025) that constrains intervened representations to remain near natural distribution
- Computes feature attribution passing ablated pre-activation through JumpReLU to handle inhibition.
- Auxiliary objective combining L2 and cosine losses against pre-recorded CL vectors to improve causal relevance when one model is causally inaccessible.
- Training models to answer Yes/No whether a prompt edit changes their behavior.
- Fine-tuning on reflection turns to implant principles into workspace.
- MAS variant with an auxiliary CL loss objective for cases where one model is causally inaccessible, enabling ANN-BNN comparisons.