) Ablative results to show the contribution of 2

Figure 1 shows the additional experiments. Meanwhile, we provide the results of non-factorial method in eq.3 as well as comparison with using Singh [2019] requires resettable environment). We will try our best to implement these prior methods and add all results in the new version. A2: We do not use state splitting after using the one-hot embedding. Due to space limitations, we cannot include the derivation here; we will add it into the new version.