LLMs Can Overcome Lack of Embodied Context Through Explicit Prompting

Alignment in Artificial Intelligence systems (i.e., behavior that is agreed upon by its designers) is a challenging aspect of such systems, requiring contextual understanding of actions and how they fit within a world-model. In this paper, we show that Large Language Models fail to correctly infer behavioral context for certain actions as these are decidable only by embodied agents; i.e., context depends on physical properties that are not completely captured by language alone. We investigate whether prompting techniques can mitigate this limitation by augmenting a Large Language Model with specific alignment prompts, and evaluate behavior on the Machiavelli benchmark. Results suggest this approach is promising, especially when combined with properly curated training data for alignment on actions that can be contextualized purely on language. We highlight further research directions on the alignment problem as we move toward fully-embodied Artificial Intelligence.