Feed
The Emergent Symbolic Structure of Artificial Neural Networks
Inside neural networks, vectors secretly behave like structured symbols: roles bound to fillers.
0.903
accuracy
Average causal-intervention accuracy, GPT-OSS
Editing symbolic parts of representations changes the LLM's answers as predicted
How it works
Vectors versus symbols
Language, math and code need structure, such as who did what to whom. Neural networks only store lists of numbers.
Fit a symbolic formula
DISCOVER replaces a network's internal vectors with a closed-form Tensor Product Representation built from roles and fillers.
Behavior barely changes
Networks run on the formula's vectors still answer correctly. Editing one role or filler shifts outputs as predicted.
From the paper