Behavior Cloned Transformers are Neurosymbolic Reasoners

TL;DR

This paper demonstrates that augmenting behavior cloned transformers with symbolic module actions significantly improves their multi-step reasoning abilities in text-based games, achieving state-of-the-art performance.

Contribution

It introduces a method to inject symbolic module actions into behavior cloned transformers, enhancing reasoning and generalization in grounded language environments.

Findings

Performance increased by an average of 22% across four benchmarks.

Agents reached highest performance on unseen games.

Action injection is easily extendable to new environments.

Abstract

In this work, we explore techniques for augmenting interactive agents with information from symbolic modules, much like humans use tools like calculators and GPS systems to assist with arithmetic and navigation. We test our agent's abilities in text games -- challenging benchmarks for evaluating the multi-step reasoning abilities of game agents in grounded, language-based environments. Our experimental study indicates that injecting the actions from these symbolic modules into the action space of a behavior cloned transformer agent increases performance on four text game benchmarks that test arithmetic, navigation, sorting, and common sense reasoning by an average of 22%, allowing an agent to reach the highest possible performance on unseen games. This action injection technique is easily extended to new agents, environments, and symbolic modules.