LeWorldModel action sensitivity diverges from simulators at small scales
Original title: Does LeWorldModel Preserve Local Action Geometry?
Across nearly a thousand Reacher and OGBench Cube states, the most sensitive action directions differ by median angles of about 36 and 42 degrees.
A study tests how closely pretrained LeWorldModel reproduces the effects of small action changes, comparing matched simulator and model responses with Gram matrices at three perturbation scales. On Reacher and OGBench Cube, sensitive action directions diverge most at the smallest scale, and a tenfold larger perturbation cuts median normalized Gram discrepancy by roughly 70% and 47%. Pixel and embedding geometry varies more across scales than physical-state geometry. On held-out directions, environment Grams beat an isotropic reference, while LeWM's errors are higher; effects on control remain open.
Unconfirmed: Whether the measured action-sensitivity mismatch affects control is left open.
- Nearly a thousand states came from Reacher and OGBench Cube.
- Median direction angles differ by about 36 and 42 degrees at the smallest scale.
- A tenfold perturbation cut Gram discrepancy by roughly 70% and 47%.
- Held-out embedding errors were 17% and 28% for environment Grams versus 26% and 38% for an isotropic reference.
From the original article
1 min read at huggingface.co
We examine how closely pretrained LeWorldModel (LeWM) reproduces the effects of small action changes. Using matched simulator and model responses, we measure action sensitivity with Gram matrices at three perturbation scales. Across nearly a thousand states from Reacher and OGBench Cube, the most sensitive action directions differ by median angles of about 36^circ and 42^circ, respectively, at the smallest scale.
By Vasil Vasilev, Siddharth Nayak · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.