Tool Use Reduces Depth-Induced Collapse in OOD Reasoning
A benchmark in which a model progressively solves a Boolean circuit over $GF(2)$ from data is introduced, isolating OOD reasoning from these confounding factors, and general, as any computable function can be represented as a sufficiently large $GF(2)$ polynomial.
D. Koplow, T. Galanti, Tomaso A. Poggio
· 0 citations