Mantis: Ma mba- n ative T uning is Efficient for 3D Point Cloud Foundation Models

Zihao Guo, Jihua Zhu, Jian Liu, Ajmal Saeed Mian · 2026

Read Mantis: Ma mba- n ative T uning is Efficient for 3D Point Cloud Foundation Models by Zihao Guo, Jihua Zhu, Jian Liu, Ajmal Saeed Mian (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-2605-03438.txt.

Description

Zihao Guo, Yiding Sun, Jihua Zhu, Haozhe Cheng, Jian Liu, and Ajmal Saeed Mian present Mantis, a tuning method for 3D point cloud foundation models built on Mamba backbones. Full fine-tuning is costly, and existing light tuning tools were designed for transformers, clashing with Mamba's state-level dynamics and degrading accuracy.

A state-aware adapter injects small task signals directly into the selective state updates while the backbone stays frozen. A consistency distillation step also steadies the model across different point orderings. Across benchmarks, Mantis stayed competitive while training under five percent of backbone parameters, the first Mamba-native solution of its kind.

Catalog record

CREATORZihao Guo, Jihua Zhu, Jian Liu, Ajmal Saeed Mian
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE97 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.03438

Browse the full library · About the Tower of Babel