BenchmarkLRC / latent rollout consistency
Latent Rollout Consistency
An evaluation metric measuring the quality of action-conditioned latent predictions. It examines whether an entire trajectory from multi-step autoregressive rollout in latent space is consistent with the true environment transitions, decoupling model faithfulness from task performance.
Timeline (last 90 days)
Oct 28
The study proposes the Latent Rollout Consistency (LRC) metric to test whether a model maintains consistency with environment transitions when performing multi-step rollouts in latent space.
Unverified50%