Standoff: Benchmarking Representation Learning for Nonverbal Theory of Mind Tasks

We present our design and implementation of Stand-off, an innovative benchmark suite of computational theory of mind tasks, based on the competitive feeding paradigm from comparative psychology. We find that a small convolutional LSTM model without explicit theory of mind mechanisms can reach high levels of accuracy when exposed to the full variety of our task design during training. Such a model faces generalization challenges when exposed to narrower subsets of tasks. Finally, we discuss how this test may be used as a gateway for studying theory of mind skills beyond attribution of seeing and knowing.