Maze-solving transformers store a causal, steerable map of maze connections in sparse features, and activating these features is more effective than removing them.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2024 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Transformers Use Causal World Models in Maze-Solving Tasks
Maze-solving transformers store a causal, steerable map of maze connections in sparse features, and activating these features is more effective than removing them.