Teaching NNs to learn and improve the existing algorithms and exploring the state/action space representation for it.
To understand if NNs can improve on a well-established algorithm and do not fail out-of-distribution.

Source: original paper.
The paper demonstrates two important points:
I think both of the points have been folklore/written about in RL, but this work is a very convincing show-case when applied to the program induction.
<aside> 💡 Neural program induction can be stated as having a neural controller $f$ which, from a particular program execution state $s$ , issues an instruction $a=f(s)$ which has some pre-determined semantics over how it transforms $s$.
</aside>
The policy is factorised as:
$$ p_{\theta}(a_t\mid s_t) = p_\theta(a_{\text{type}}\mid s_t)p_\theta(a_{\text{args}}\mid a_{\text{type}}, s_t), $$
where $p_\theta(a_{\text{args}}\mid a_{\text{type}}, s_t) = p_\theta(a_{\text{arg1}}\mid a_{\text{type}}, s_t)p_\theta(a_{\text{arg2}}\mid a_{arg1}, a_{\text{type}}, s_t)...$
There are three types of arguments: binary (modelled with Bernoulli), integer (multinomial) and pointer (Ponter Networks). MLPs and GNNs are used.
The whole thing is trained using behaviour cloning (NLL objective) or with policy gradients with value as the baseline also adding an additional imitation loss $-\lambda \log{p_\theta(a_t'\mid s_t)}$, where $a'_t$ is the expert's action.
The authors spent a lot of efforts in designing a proper state/action space. They ignore part of the state space and define macro actions such as SwapWithNext or MoveVar. This is a great demonstration of the fact, that having access to the full state space might be disastrous (i.e. try to learn with a full maze input in MiniGrid 😃). A schematic representation of the interface view is on the picture at the beginning of the section.