Learning Local Rules Without Global Invariants Out-of-Distribution Shortcut Learning in Autoregressive Transformers

This work studies out-of-distribution generalization, shortcut learning, and internal circuit routing in autoregressive Transformers using synthetic state-space benchmarks. By showing both behaviorally and mechanistically that sequence models can achieve 100% in-distribution accuracy and emit low-perplexity Chain-of-Thought rollouts while violating conservation invariants on out-of-distribution inputs, our results highlight the importance of causal circuit auditing and out-of-distribution verification in reliability-critical machine learning systems.