Stabilizing Extrapolation in Looped Transformers via Learned Stochastic Stopping
For researchers working on length generalization in Transformers, this work identifies a key source of brittleness and proposes a training-time solution, though improvements are incremental and task-specific.
Looped Transformers exhibit brittle length generalization due to spurious correlation between sequence length and number of loops. Introducing stochasticity in the number of loops during training, particularly via RL-Halting, reduces out-of-distribution variance and improves the accuracy-stability trade-off across algorithmic tasks.
Looped Transformers, which repeatedly apply a shared transformer block, are an architecturally natural fit for variable-length algorithmic tasks. Although they can exhibit strong length generalization beyond the length of training sequences, this behavior is brittle, yielding high out-of-distribution (OOD) variance, even across well-performing in-distribution solutions. We trace this variance to the spurious correlation in simple algorithmic tasks between sequence length and number of loops. Introducing stochasticity into the number of loops during training sharply reduces OOD variance and stabilizes predictions across inference-time loop counts. To improve upon heuristic randomization schemes, we further analyze RL-Halting as a learned stochastic schedule and find that it generally improves the accuracy-stability trade-off. Across binary addition, Dyck-1, Unique Set, and Copy, learned stochastic stopping often improves this trade-off but can also stabilize a suboptimal computation. Our work suggests that "when to stop" should be treated as a training-time design choice, not merely an inference-time computation-allocation rule.