Encodes cause-and-effect links between events and states.
What It Does
Relation.Causal neurons activate on causal relationships: explicit causal markers ('because', 'therefore', 'as a result', 'caused by', 'led to'), implicit causal structures ('smoking damages lung tissue' implies a causal mechanism), and counterfactual causation ('if X had not occurred, Y would not have followed'). They encode the directionality of causation — which event produces which outcome.
How It Behaves
Despite being one of the smallest elements by count, Causal neurons show a distinctive bimodal layer distribution — active in both early layers (parsing causal connectives) and late layers (committing to a causal interpretation of the full context). The late-layer activation is disproportionately strong relative to the neuron count, suggesting Causal neurons have high per-neuron influence on output. They co-activate with Time.Sequence neurons (cause precedes effect) and Boolean.Presence neurons (causes must be present to produce effects).
Research Example
In Mistral 7B, Relation.Causal neurons activate on implicit causation as strongly as explicit causation — 'the ice melted' in a warm-weather context produces similar Causal neuron firing to 'the ice melted because of the heat', even though no causal connective is present. The model has learned to infer causation from contextual plausibility, not just from surface markers. This is why models can reason about causal chains without explicit 'because' statements.