Tag
This paper investigates how transformers organize causal knowledge, showing that type-level supervision induces a typed routing structure that is functionally decoupled from answer readout, with exact local editability and bit-exact revertibility.
The paper identifies 'temporal credit dilution' in learned dynamics models where global readouts focus on spurious correlates rather than brief physical events. It proposes CREST, a training-free method that re-anchors pooled representations using event core estimates, improving out-of-distribution robustness.