Can Test-Time Memory Evolution Outperform Test-Time Training for Scientific Discovery?

Test-Time Memory Evolution (TTME) improves scientific discovery efficiency by distilling experience into external memory instead of performing online weight updates. This plug-and-play adaptation method outperforms test-time training (TTT) frameworks by bypassing the computational overhead required for gradient-based model updates during inference.