ManimAgent: Self-Evolving Multimodal Agents for Visual Education

Wenjia Jiang 1 , Zongyuan Cai 2 , Yuanhang Shao 2 , Chenru Wang 3 , Boyan Han 2 , Zhixue Song 4 , Keyu Chen 5 , Shengwei An 3 , Xu Yang 2 , and Zhou Yang 1 1 University of Alberta 2 Southeast University 3 Virginia Tech 4 Emotion Machine Inc. 5 Vivavia Inc · 2026

Read ManimAgent: Self-Evolving Multimodal Agents for Visual Education by Wenjia Jiang 1 , Zongyuan Cai 2 , Yuanhang Shao 2 , Chenru Wang 3 , Boyan Han 2 , Zhixue Song 4 , Keyu Chen 5 , Shengwei An 3 , Xu Yang 2 , and Zhou Yang 1 1 University of Alberta 2 Southeast University 3 Virginia Tech 4 Emotion Machine Inc. 5 Vivavia Inc (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-2606-30296.txt.

Description

Wenjia Jiang and nine colleagues build an agent that learns across tasks while making math animations. The agent writes Python in the Manim library to animate a paper section. Reflection within one task fixes errors, but lessons die with the task. Their system, ManimAgent, carries experience forward in memory. No model weights change and no human examples seed it.

Once an animation converges, a vision-language model grades the rendered keyframes. Scores feed a dual-channel Episodic Memory Bank grown from the agent's own work. The positive channel stores success rationales as soft reference examples. The negative channel stores validated failure patterns as known pitfalls. Fixed-probe tests compare against no-memory and retrieval baselines. Blind human first-try success rises as memory grows. Reflection rounds needed per task fall at the same time.

Catalog record

CREATORWenjia Jiang 1 , Zongyuan Cai 2 , Yuanhang Shao 2 , Chenru Wang 3 , Boyan Han 2 , Zhixue Song 4 , Keyu Chen 5 , Shengwei An 3 , Xu Yang 2 , and Zhou Yang 1 1 University of Alberta 2 Southeast University 3 Virginia Tech 4 Emotion Machine Inc. 5 Vivavia Inc
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE79 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.30296

Browse the full library · About the Tower of Babel