6 Supplementary Material 6.1 Original CLUTRR evaluation

The models are given as input " [story] [query] " and asked to generate the proof and answer. In our case, the entity names are important to evaluate systematic generalization. Train and test overlap percentages between all building blocks are in Table 5. The models are given as input " [story] [query] " and asked to The models are given as input " Figure 6: Evaluation of models trained on levels 2, 4, 6 only. Additional experiments in Section 6.4 were run with The models are given as input " [story] Models are trained on levels 2, 4, 6 only.