The Token Before the Value Is the Key: How Hybrid Architectures Organize Induction Circuits

arXiv:2609.15545 · cs.LG · Submitted 2026-09-14 · Read on arXiv

cs.LG

Submitted: 2026-09-14

Updated: 2026-09-15

Comments: 30 pages, including references and appendices

Code: https://github.com/ckpassenger/bind-match-copy

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Terminology

• Transformer — 105× in this paper · explained in 2 episodes
• Trace — 14× in this paper · explained in 2 episodes
• Calibration — 6× in this paper · explained in 7 episodes
• Ablation — 5× in this paper · explained in 1 episode
• Hybrid Architectures — 4× in this paper · explained in 3 episodes
• In-Context Learning — 4× in this paper · explained in 4 episodes
• Embeddings — 3× in this paper · explained in 2 episodes
• Activation Patching — 2× in this paper · explained in 2 episodes
• Abstraction — 1× in this paper · explained in 1 episode
• Aggregation — 1× in this paper · explained in 1 episode
• Consistency — 1× in this paper · explained in 3 episodes
• Hybrid Model — 1× in this paper · explained in 2 episodes
• Interpretability — 1× in this paper · explained in 15 episodes
• Mechanistic interpretability — 1× in this paper · explained in 2 episodes

Sources

• In-Context Language Learning: Architectures and Algorithms
• Predicting the Emergence of Induction Heads in Language Model Pretraining
• Mechanistic evaluation of Transformers and state space models
• Longformer: The Long-Document Transformer
• Understanding the Skill Gap in Recurrent Language Models: The Role of the Gather-and-Aggregate Mechanism
• Short window attention enables long-term memorization
• Expressivity-Efficiency Tradeoffs for Hybrid Sequence Models
• Griffin: Mixing Gated Linear Recurrences with Local Attention for Efficient Language Models
• Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability
• Mamba: Linear-Time Sequence Modeling with Selective State Spaces
• Repeat After Me: Transformers are Better than State Space Models at Copying
• Kimi Linear: An Expressive, Efficient Attention Architecture
• Kimi K3: Open Frontier Intelligence
• Jamba: A Hybrid Transformer-Mamba Language Model
• Olmo Hybrid: From Theory to Practice and Back
• On the Emergence of Induction Heads for In-Context Learning
• Invariant Learning Dynamics of Transformers in Inductive Reasoning Tasks
• Can Mamba Learn How to Learn? A Comparative Study on In-Context Learning Tasks
• Rethinking the Role of Efficient Attention in Hybrid Architectures
• On the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability

Related papers

• Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
• AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
• Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
• Convergence issues in Relational Concept Analysis based on AOC-posets
• Beliefs Beyond Posteriors: Local-Consistency Optimisation for Bayesian Neural Networks
• Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks