Structured World Representations in Maze-Solving Transformers
This work provides insights into transformer interpretability, which is an incremental step for researchers aiming to demystify AI systems.
The researchers tackled the problem of understanding transformer models' internal behavior by studying small transformers solving mazes, finding evidence that these models form structured internal representations of maze topology and paths, such as linearly decoding a single token's residual stream to reconstruct the entire maze.
Transformer models underpin many recent advances in practical machine learning applications, yet understanding their internal behavior continues to elude researchers. Given the size and complexity of these models, forming a comprehensive picture of their inner workings remains a significant challenge. To this end, we set out to understand small transformer models in a more tractable setting: that of solving mazes. In this work, we focus on the abstractions formed by these models and find evidence for the consistent emergence of structured internal representations of maze topology and valid paths. We demonstrate this by showing that the residual stream of only a single token can be linearly decoded to faithfully reconstruct the entire maze. We also find that the learned embeddings of individual tokens have spatial structure. Furthermore, we take steps towards deciphering the circuity of path-following by identifying attention heads (dubbed $\textit{adjacency heads}$), which are implicated in finding valid subsequent tokens.