bag100/action-atlas-xvla

Action Atlas: X-VLA sparse autoencoders and concepts

Sparse autoencoders (SAEs) and identified concepts for X-VLA, part of the Action Atlas release accompanying the paper on cross-task activation injection in vision-language-action models. The interactive explorer is at https://action-atlas.com.

What is here

• —saes/ TopK SAEs (k=64, 8x expansion) over the X-VLA (soft-prompted Florence-Large), 24 layers, 1024-dim. Arms present: per-token 48, mean-pool 48. Per-token is the primary release. Mean-pool SAEs are included for transparency.
• —concepts/ the concept-to-feature index (1 file(s)): per (pathway, layer), the SAE feature indices that score for each manipulation concept, with Cohen's d and frequency.
• —videos/ the trajectory-to-video linkage: 50 curated baseline rollouts and 56335 concept-ablation rollouts, each pointing to its public Tigris URL with suite, task, and success.
• —manifest.jsonl one fully labelled row per artifact (type, model, pooling, pathway, layer, dims, k, metrics, sha256), and loader.py a reference loader.

Pathways: single. Environments: LIBERO and SimplerEnv.

Loading

from safetensors.torch import load_file, safe_open
sae = load_file("saes/per_token/sae_layer0.safetensors")  # encoder/decoder weights and biases
with safe_open("saes/per_token/sae_layer0.safetensors", framework="pt") as f:
    meta = f.metadata()  # model, pooling, pathway, layer, d_in, d_sae, k

See loader.py and REPRODUCIBILITY.md for the activation hook placement and training configuration.

Counts

Per-token SAEs released: 48. All SAE arms: per-token 48, mean-pool 48.

License and attribution

The SAE weights and derived artifacts in this repository are released under CC-BY-4.0 as interpretability tools. The underlying model weights (X-VLA weights (per upstream release terms)) retain their own upstream licenses and are not redistributed here.

Citation

@inproceedings{actionatlas2026,
  title     = {Action Atlas: mechanistic interpretability of vision-language-action models},
  booktitle = {NeurIPS},
  year      = {2026}
}