Causal Writability: Core Checkpoints

Research checkpoints for A Chosen Future Can Still Be Rewritten: Causal Writability in Video Models.

Project page | Code and reproduction instructions

Task Checkpoints Purpose
Spring Short and Long, seeds 3407/3408/3409, 50K Core behavior, interventions, transfer and history comparison
Free Fall hist32, 100K Gravity controller and causal-access replication
Pendulum Short and Long, seed 3407, 50K Core controller and history comparison
Pretrained Wan Direct and Neutral-first at 2K/5K/10K; neutral-only 10K start Main pretrained adaptation and controller results

These sixteen files deliberately exclude the 15-seed x 6-checkpoint Spring longitudinal ensemble and additional pretrained intermediate checkpoints used in appendix timing analyses. They require the paper's task-specific code and configuration; they are not generic text-to-video models.

The frozen Wan2.1 VAE is available from its upstream repository: Wan2.1 VAE. It is not duplicated here.

Controllers

  • controllers/spring/3407, 3408, 3409: frozen Short-model phase controllers at zero-based blocks 3, 6 and 4, respectively. Each contains model.json and 20 top-four basis arrays, about 401 MB per seed.
  • artifacts/reproduction/freefall/: the frozen train-only PCA basis, split arrays and archived decoded frames. Coefficients and commands are included in the code repository.
  • Pendulum's dense PCA basis is not hosted here. Its code provides the fit-only reconstruction and coefficient-fitting workflow; model weights alone are not a ready-to-run Pendulum controller.

From the code repository's reproduction/spring directory, for example:

hf download xingyun-24/causal-writability \
  --include 'controllers/spring/3409/*' --local-dir weights

Use weights/controllers/spring/3409 with the matching Short seed-3409 checkpoint and block 4. See the task README for the complete generation and intervention command. Raw activation banks are not needed for this download.

downloads.json lists the weights. The files are publicly downloadable without logging in. Original research code uses MIT; the code's license does not automatically relicense upstream dependencies or pretrained-model derivatives. The seven adapted Wan checkpoints retain the upstream model's Apache-2.0 terms; see LICENSE-WAN.txt and the upstream model repository.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support