← 研踪 PaperTrailarXiv:2605.17811 · v2
One Model, Two Roles: Emergent Specialization in a Shared Recurrent Transformer
首次提交 2026-05-18 11:36
版本更新 2026-10-29 04:53
雷达观测 2026-10-30 11:25
摘要
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.
作者备注：Accepted at NeurIPS 2026. 29 pages, 18 figures, 13 tables
观测到的版本
相邻记录可能跳过中间版本；首次发现 v2 及以上也会标为版本更新。
v2 · 2026-10-29 04:53 · 最新观测版本
One Model, Two Roles: Emergent Specialization in a Shared Recurrent Transformer
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.
查看此版本解读 →查看此版本原文 ↗