One Model, Two Roles: Emergent Specialization in a Shared Recurrent Transformer
cs.LG, cs.AI, math.OC
Submitted: 2026-05-18
Updated: 2026-10-28
Comments: Accepted at NeurIPS 2026. 29 pages, 18 figures, 13 tables
Code: https://github.com/juchengshen/air
License: http://creativecommons.org/licenses/by/4.0/
The gist: 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…
Terminology
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 https://github.com/juchengshen/air.
Sources
- A Mechanistic Analysis of Looped Reasoning Language Models
- Hierarchical Reasoning Models: Perspectives and Misconceptions
- Adaptive Computation Time for Recurrent Neural Networks
- Less is More: Recursive Reasoning with Tiny Networks
- Show Your Work: Scratchpads for Intermediate Computation with Language Models
- Emergence of Functionally Differentiated Structures via Mutual Information Minimization in Recurrent Neural Networks
- Hierarchical Reasoning Model
Related papers
- A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics
- FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis