Self-Consistency for LLM-Based Motion Trajectory Generation and Verification

Jiaju Ma, R. Kenny Jones, Jiajun Wu, Maneesh Agrawala · 2026

Read Self-Consistency for LLM-Based Motion Trajectory Generation and Verification by Jiaju Ma, R. Kenny Jones, Jiajun Wu, Maneesh Agrawala (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-29301.txt.

Description

Ma and co-authors adapt self-consistency to motion graphics made by language models. A prompt asks for a path, such as a circle spiraling. The model samples many candidate trajectories. Consistent ones are grouped by clustering. Two paths count as consistent if an allowed warp maps one to the other. Warps form families, from rigid to affine. An algorithm recovers the right family on its own. Generation accuracy rose by four to six percent. Verification precision beat vision-model baselines by 11 percent.

Catalog record

CREATORJiaju Ma, R. Kenny Jones, Jiajun Wu, Maneesh Agrawala
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE74885
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.29301

Browse the full library · About the Tower of Babel