| Summary: | People are incredibly flexible and efficient inductive reasoners. On the other hand, current approaches in program synthesis show strong domain-specific performance, but are both less sample-efficient and less flexible. Large language models improve upon this sample-efficiency and domain-generality, but lack robustness and still fall far short of people and traditional approaches on difficult induction tasks. In this thesis, we propose two hypotheses for how people seemingly overcome this trade-off between flexibility and efficiency. In the first, we propose that people may operate over an incredibly vast language which is made tractable via a strong, bottom-up proposal model. In the second, we propose that, alternatively, people may relax the necessity of such a strong proposal model by learning task-specific reasoning languages through experience. We build models operationalizing both hypotheses and show that they can improve the generality and efficiency of previous models.
|
|---|