1 paper · 1 filter
Qilin Ye, Deqing Fu, Robin Jia +1
Transformers often fail to learn generalizable algorithms, instead relying on brittle heuristics. Using graph connectivity as a testbed, we explain this phenomenon both theoretical…