Well, still no. I used to be a huge fan of latent stuff, due to my background on signal processing and spectral methods. But the problem with latent space is that it is too brittle and not really intrinsically interpretable. What’s worse — you cannot reuse existing powerful foundation models already out there…
I am betting on neural symbolic world models 
@girish_432 Well, symbols are interpretable. Once things are interpretable, foundation models can be heavily reused! Two examples here: TraceGen: https://arxiv.org/abs/2511.21690 Momagraph: https://arxiv.org/abs/2512.16909
Well, still no. I used to be a huge fan of latent stuff, due to my background on signal processing and spectral methods. But the problem with latent space is that it is too brittle and not really intrinsically interpretable. What’s worse — you cannot reuse existing powerful foundation models already out there…
I am betting on neural symbolic world models