Hierarchical Code Embeddings with Multi-Level Attention for Reinforcement Learning State Representation

ICLR 2026 Conference Submission25398 Authors

Keywords: Multi-Level Attention

Abstract:

\begin{abstract}
In this paper, we propose novel state representation and reinforcement learning (RL)
system of encoding the semantics of code hierarchically using multiple
attention mechanisms. Traditional approaches regularly address code embeddings
as flat sequences or to be reliant only on graph-based representations,
which don't capture the complex level of interplay between local and global
code features. The proposed method incorporate token-level,
function-level, and module-level attention using graph-structured
dependencies, to allow the RL agent to reason about code at varying
granularities while maintaining structural relationships
\end{abstract}

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 25398