Generating Symbolic Reasoning Problems with Transformer GANs

Jens U. Kreber, Christopher Hahn

Keywords: Transformer, GAN, symbolic reasoning, temporal logic

Abstract: Constructing training data for symbolic reasoning domains is challenging: Existing instances are typically hand-crafted and too few to be trained on directly and synthetically generated instances are often hard to evaluate in terms of their meaningfulness. We study the capabilities of GANs and Wasserstein GANs equipped with Transformer encoders to generate sensible and challenging training data for symbolic reasoning domains. We conduct experiments on two problem domains where Transformers have been successfully applied recently: symbolic mathematics and temporal specifications in verification. Even without autoregression, our GAN models produce syntactically correct instances and we show that these can be used as meaningful substitutes for real training data when training a classifier. Using a GAN setting also allows us to alter the target distribution: We show that by adding a classifier uncertainty part to the generator objective, we obtain a dataset that is even harder to solve for a classifier than our original dataset.

One-sentence Summary: Transformer GANs generate useful problem instances for symbolic reasoning.

Community Implementations: [ 2 code implementations](https://www.catalyzex.com/paper/arxiv:2110.10054/code)