Structural Generalization on SLOG without Hand-Written Rules

Neural Cellular Automaton Cracks Structural Generalization in Semantic Parsing Without Hand-Written Rules A new NCA-based semantic parser with a discrete bottleneck achieves 100% type-exact match on 11 of 17 SLOG structural generalization categories, outperforming AM-Parser on three categories where AM-Parser scores 0-74%, with a stability of 0.2 standard deviation across 10 seeds versus AM-Parser's 4.3. All compositional rules emerge from local iterative learning rather than hand-engineered algebraic constraints. NLP researchers building robust semantic parsers should take note: the remaining 5,539 failures reduce to exactly two failure modes, offering a precise roadmap for closing the gap entirely.