共享权重循环 Transformer 的涌现分工：Asymmetric Input Recurrence 研究

View PDF HTML (experimental)

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: $z_H$ 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 this https URL.

Comments:	Accepted at NeurIPS 2026. 29 pages, 18 figures, 13 tables
Subjects:	Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Optimization and Control (math.OC)
Cite as:	arXiv:2605.17811 [cs.LG]
	(or arXiv:2605.17811v2 [cs.LG] for this version)
	https://doi.org/10.48550/arXiv.2605.17811 arXiv-issued DOI via DataCite

Submission history

From: Jucheng Shen [view email]
[v1] Mon, 18 May 2026 03:36:58 UTC (5,329 KB)
[v2] Wed, 28 Oct 2026 20:53:46 UTC (2,818 KB)