One Model, Two Roles: Emergent Specialization in a Shared Recurrent Transformer
Abstract
Can a shared-weight recurrent Transformer---one that repeatedly applies the same Transformer to update evolving latent states---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. Across Sudoku-Extreme and Maze, decoded rollouts reveal a consistent split: zH behaves like a fully committed proposal state, whereas zL retains local uncertainty and shifting intermediate structure. Freeze interventions provide evidence of dynamical dependence between the two states, with responses that vary across tasks, intervention directions, and seeds. Ablations show that to induce specialization, the shared model needs to be able to tell the two update types apart, either from input injection asymmetry or from a separate level token. Mechanistically, attention analysis shows that L-updates are consistently more local than H-updates in both Sudoku and Maze. Together, these results show that, in a two-state recurrent setting, a clear state-identity signal can induce stable, related functional roles inside a shared-parameter recurrent Transformer. Code is available at https://github.com/juchengshen/air.
Related
- Injectivity of specialization maps in critical cohomology
- Revenge of Monosemanticity: Neuron Specialization as a New Form of Feature Learning in MLPs
- Macdonald principal specialization ratios and their generating polynomials
- Local Shellings of Broken-Circuit Complexes via Boundary States and the Chromatic Specialization of the Negami Polynomial
- A Degenerate Clausenian 3F2(1) Summation and Its Arithmetic Specialization
BibTeX
@misc{shen2026model,
title = {{One Model, Two Roles: Emergent Specialization in a Shared Recurrent Transformer}},
author = {Jucheng Shen and Barbara Su and Anastasios Kyrillidis},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2605.17811},
eprint = {2605.17811},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2605.17811},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - One Model, Two Roles: Emergent Specialization in a Shared Recurrent Transformer
AU - Jucheng Shen
AU - Barbara Su
AU - Anastasios Kyrillidis
PY - 2026
DA - 2026/10/28
PB - arXiv
DO - 10.48550/arxiv.2605.17811
UR - https://arxiv.org/abs/2605.17811
AB - Can a shared-weight recurrent Transformer---one that repeatedly applies the same Transformer to update evolving latent states---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. Across Sudoku-Extreme and Maze, decoded rollouts reveal a consistent split: z_H behaves like a fully committed proposal state, whereas z_L retains local uncertainty and shifting intermediate structure. Freeze interventions provide evidence of dynamical dependence between the two states, with responses that vary across tasks, intervention directions, and seeds. Ablations show that to induce specialization, the shared model needs to be able to tell the two update types apart, either from input injection asymmetry or from a separate level token. Mechanistically, attention analysis shows that L-updates are consistently more local than H-updates in both Sudoku and Maze. Together, these results show that, in a two-state recurrent setting, a clear state-identity signal can induce stable, related functional roles inside a shared-parameter recurrent Transformer. Code is available at https://github.com/juchengshen/air.
ER -