Sparse Attentive Backtracking: Temporal Credit Assignment Through Reminding

Nan Rosemary Ke, Anirudh Goyal ALIAS PARTH GOYAL, Olexa Bilaniuk, Jonathan Binas, Michael C. Mozer, Chris Pal, Yoshua Bengio

Feb-14-2026, 23:59:56 GMT–Neural Information Processing Systems 

The T = 100, itisclearthatT grows.SABstill tocompleteT = 5000, whereasT = 2000bothv self-attention 1/8 = 12.5%).