AI RESEARCH
One Model, Two Roles: Emergent Specialization in a Shared Recurrent Transformer
arXiv CS.LG
•
ArXi:2605.17811v1 Announce Type: new Can a shared-weight recurrent Transformer develop distinct internal roles without being partitioned into separate modules? We study this in Asymmetric Input Recurrence (AIR), a minimal two-state reasoning architecture in which the same Transformer model is reused for both updates (per literature, L and H) and the only built-in difference in the update rule is that the encoded input is injected during L-updates but not H-updates.