Large-scale language models such as transformers have become the de facto standard for a wide range of natural language processing (NLP) tasks. Despite their apparent linguistic savvy, such sequence models are known to lack a real understanding of the cause and effect of their actions, which can lead to false decisions due to auto-suggestive delusions. In the new paper Shaking the Foundations: Delusions in Sequence Models for Interaction and Control, a DeepMind research team explores the origin of these mismatches and addresses the problem by treating actions as causal interventions. The team shows that a system can learn to condition or intervene on data by training with the use of factual and counterfactual error signals respectively. Sequence models are updated based on collected data, and these updates will differ depending on whether the data was generated by the model itself (i.e.

The recent phenomenal success of language models has reinvigorated machine learning research, and large sequence models such as transformers are being applied to a variety of domains. One important problem class that has remained relatively elusive however is purposeful adaptive behavior. Currently there is a common perception that sequence models "lack the understanding of the cause and effect of their actions" leading them to draw incorrect inferences due to auto-suggestive delusions. In this report we explain where this mismatch originates, and show that it can be resolved by treating actions as causal interventions. Finally, we show that in supervised learning, one can teach a system to condition or intervene on data by training with factual and counterfactual error signals respectively.