Improves length generalization in looped language models by aligning each iteration with a specific intermediate state and anchoring recurrent-state scale. On swap-S₅, a model trained on up to 32 moves reaches 95.5% accuracy at 512 moves.
Improves length generalization in looped language models by aligning each iteration with a specific intermediate state and anchoring recurrent-state scale. On swap-S₅, a model trained on up to 32 moves reaches 95.5% accuracy at 512 moves.