Handshape and orientation carry much of the meaning in sign language, but many sign-language encoders treat video as a purely visual appearance problem. A new arXiv preprint argues that this leaves out the geometric structure of the hands, and introduces SignDEG (DINO Eats Geometry), a temporal sign encoder designed to use articulated hand geometry as an input signal.

According to the abstract, SignDEG is a temporal encoder—meaning it processes sign sequences over time—and it explicitly incorporates hand geometry rather than relying on appearance alone. The authors position this as a corrective to encoders that learn from appearance only, though the abstract does not specify how the geometry is extracted or fused.

The preprint is listed as arXiv:2609.06296v2, a replacement, but the provided abstract is truncated. No quantitative results or comparisons are included in the available text, so the claims should be read as a proposal rather than a demonstrated improvement.