Humans can learn and generalize novel concepts from sparse data by expressing knowledge in rich structural formats. The survey proposes that programs are a strong candidate for a universal representation of concepts.
It reviews various computational models of concept learning that utilize programs as their form of concept representation. The evaluation focuses on how these models contribute toward creating a universal language for concepts.
This approach matters for engineers running models or agents because it could enable more flexible and generalizable knowledge systems. Understanding these models helps in designing systems that better mimic human concept learning.
Source: https://arxiv.org/abs/2609.04528