Low-Rank Hidden State Patches Enable Counterfactual Rollouts in Learned World Models

A recent preprint on arXiv (2608.15156) explores the potential of minor, directly addressable modifications to the hidden state of a recurrent world model to influence its trajectory towards a counterfactual outcome. Utilizing a controlled environment featuring two objects in two dimensions and a 192-dimensional hidden state, the study demonstrates that the model can inherently represent and project edited futures for a limited set of local velocity adjustments. Researchers create candidate low-rank carriers from the factual-to-counterfactual hidden state differences observed during training and develop a mapping from the factual state and desired edit to these carrier coefficients. The smallest rank tested that meets the full development-panel criteria is rank 4. A single rank-4 patch at the anchor effectively enables the desired counterfactual rollout, enhancing the interpretability of world models by pinpointing which hidden state components influence predictions. The paper can be found on arXiv with the identifier 2608.15156.