Domain-Invariant Latent Lookahead (DILL): learning task-relevant future representations to reduce shortcut learning in vision-language-action models. CoRL 2026.