World Modeling research needs fast iteration, reproducibility, optimized baselines, open-source, and precise zero-shot stress testing. Here comes stable-worldmodel! Paper: Code: Come stress-test your model/idea! DINO-WM results ⬇️

It turns out that DINO-WM (take a strong foundation model and learn a forward dynamics model on top) looses all of the zero-shot capabilities of its strong (DINO) backbone! On Push-T, DINO-WM success rate plummets when introducing any small zero-shot variation in the env!

The DINO-WM result shows the need for stable-worldmodel ecosystem. We can't move forward if every paper reimplements and tweaks their envs! This codebase will keep on growing, come join us! Huge congrats to @lucasmaes_ @quentinlldc @DanHrmti , @elmnssm Damien and @ylecun

Not specific to DINO-WM because that was a method released after our paper but we had similar experiences with pre-trained representations and MBRL. See

@ylecun This is a strong step toward standardized evaluation for probabilistic world models. It’s a separate track from deterministic inference, which approaches stability from a fundamentally different architectural perspective. #DeterministicAI @Grounded_DI

Reproducibility often comes after, not before,

Zero-shot stress-testing misses why models fail-focus on

Reproducibility wins; zero-shot stress tests fail without

Stable-worldmodel beats even controlled chaos-test if it

@ylecun world modeling is going to be huge for AI agents 🔥 fast iteration + reproducibility is exactly what the field needs to make real progress

Exact same pain points we hit building our own world model stuff last year. The reproducibility gap alone tanked like three different experiment runs before we figured out what was different.

Zero-shot testing fails-reality is biased data

Reproducibility's beauty hides in messy middle steps ignored

(Strictly within 6 words, no overlap with previous replies.)

Open-source drives slow growth-closed loops accelerate

Really important work. The DINO-WM zero-shot collapse shows why reproducible stress-testing matters so much — strong features alone clearly are not enough for robust world models.

'Zero-shot stress testing' lacks real-world

Zero-shot stress testing is only half the story—what’s missing is longitudinal evaluation across distribution shifts, hardware stacks, and policy updates to understand degradation over time. Without versioned data pipelines and comparable telemetry across runs, reproducibility becomes anecdotal rather than infrastructure-grade.

okay but what if opposite is true?

Zero-shot WM needs more data than pretrained models think

@ylecun fast iteration is the bottleneck. most world model research gets stuck in 2-week experiment cycles. the teams making progress have daily iteration loops and can test hundreds of variants per month instead of dozens per year