DAIN: Dynamic Agent-Based Interaction Network for Efficient and Collaborative Multimodal Reasoning

Xinxin Chen, Yuchen Li, Zihan Wang, Ruixin Liu, Mingyuan Zhao · 2026

Read DAIN: Dynamic Agent-Based Interaction Network for Efficient and Collaborative Multimodal Reasoning by Xinxin Chen, Yuchen Li, Zihan Wang, Ruixin Liu, Mingyuan Zhao (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-2606-30189.txt.

Description

Xinxin Chen, Yuchen Li, Zihan Wang, Ruixin Liu, and Mingyuan Zhao rebuild multimodal fusion as teamwork among agents. Static mixture-of-experts systems fuse images, text, and other inputs rigidly. Complex real-world tasks need adaptive, efficient collaboration instead. Their network, DAIN, lets specialized interaction agents cooperate dynamically. A context-aware meta-controller schedules which agents activate for each sample.

Sparse activation keeps computation low while agents build consensus. Compressed communication passes between the active agents during reasoning. A multi-objective loss balances accuracy, agent specialization, and cost jointly. Five diverse benchmarks test the framework, from medical data to sentiment. DAIN sets a new state of the art across them. Accuracy on the ADNI medical benchmark gains noticeably over prior work. Ablations confirm dynamic scheduling and communication both matter. The design also exposes agent roles, making them easier to read.

Catalog record

CREATORXinxin Chen, Yuchen Li, Zihan Wang, Ruixin Liu, Mingyuan Zhao
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE33 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.30189

Browse the full library · About the Tower of Babel