There Is No Free Loop: When Does Looping a Trained Model Pay for Its Inference Compute?

In 30 Sekunden

Was passiert ist

Researchers found that adding computational loops to pretrained models only improves performance for small models, not larger ones, despite keeping parameters constant.

Warum es zählt

Matters for engineers optimizing inference costs on fixed parameter budgets, especially when considering architectural modifications to deployed models.

Achtung

Results tested only on Qwen3 and Qwen3.5 models; generalization to other architectures and whether loops learned during pretraining behave differently remains unclear.