Contrastive World Models beats Dreamer without pixel decoder
A new paper replaces the pixel-reconstruction step in Google DeepMind's Dreamer world-model architecture with a contrastive objective, and reports better results when backgrounds are cluttered or filled with video.
A paper titled "Contrastive World Models," posted to arXiv and flagged by the research-summary account askalphaxiv, removes the pixel-reconstruction decoder from the design of Google DeepMind's Dreamer world-model architecture. It replaces reconstruction with a Deep InfoMax-style objective that maximizes mutual information between action sequences and patterns in future frames, the paper says.
https://x.com/askalphaxiv/status/2104142235487101053
The author argues that reconstructing every pixel wastes model capacity on irrelevant background detail rather than the dynamics that matter for planning and control, according to the paper. Dropping the decoder also cuts the compute cost of each training step.
On Dreamer's standard benchmark the contrastive version performs about the same as the original, the paper reports. Adding a simple visual distractor, it clearly outperforms Dreamer, especially on the cheetah-run task. Against a natural video background, it beats both Dreamer and a momentum-prediction baseline across every task tested, the paper says.
The paper does not say whether it has released training code. Askalphaxiv's summary of the work drew traction on X on Saturday, though no lab or outside group has yet reproduced the results independently.