the new agent, Δ-IRIS, uses a model-based RL approach with a world model that encodes changes between time steps and predicts future changes. sets new performance records on the Crafter benchmark and trains significantly faster than previous models.

i’ll leave this here…

the new agent, Δ-IRIS, uses a model-based RL approach with a world model that encodes changes between time steps and predicts future changes. sets new performance records on the Crafter benchmark and trains significantly faster than previous models.

i’ll leave this here…

https://arxiv.org/abs/2406.19320

Love this. Do you know of any other similar models you'd recommend checking out?