A controlled experiment comparing an LSTM baseline and a deterministic symbolic executor on synthetic Boolean expressions.
01 / 1Saved run: neural accuracy falls from 87.1% to 71.7% under structural shift; the deterministic executor reaches 100% with clean inputs.
Select an image to view it full size.
My contribution
Designed a compact comparison of compositional generalization using shallow training expressions and deeper evaluation expressions.
The approach
Generate structured Boolean expressions and explicit evaluation splits.
Compare neural predictions with a parser-based symbolic pipeline.
Inspect generalization under deeper nesting and composition.
Scope & perspective
Explicit structure can help systematic generalization. The symbolic result depends on clean tokens and exact parsing assumptions; it is not a general claim of superiority.
