GTAGenerative Tower of AgentsRead the Paper

GTA: Generative Tower of Agents for Scalable Holonic Scene Generation

Zeng Tao1, Yunuo Chen1,2, Xiang Feng1, Yumeng He1, Peihao Li2,3Nanxiang Jiang1, Yaowei Guo1, Fanlian Zeng1,2, Yin Yang4, Chenfanfu Jiang1

1UCLA2Envora3UC Berkeley4University of Utah

First page of the GTA paperRead the Paper

Building Tower of Babel in AI era

GTA unifies scene representation as holon across scales and types, which enables more agents to work in parallel as the scene scales up, making world-scale modeling feasible.

Roaming the world with GTA

How does GTA scale up agentic scene generation?

Unified holon representation

Scene units of any scale and type can be represented as holons. Each agent processes a holon group, and holon groups exchange information through unified interfaces.

Parallel agentic execution

Once a holon group’s dependencies are satisfied, it enters a pending queue. The scheduler assigns it to an available agentic slot, allowing multiple agents to work in parallel. As the scene scales up, more agents can be used.

Stronger layouts. Richer details.
Built for larger scenes.

Compare the overall scene layout, then explore individual objects in 3D.

GPT-6 Astra
GPT-6 Astra powered by GTA

Faster generation, higher throughput

3.36× faster than GPT-6 Astra (Staged)

Wall-clock time per object (s)

Claude Fable 5.2
23.61
GPT-6 Astra (Original)†
19.26
GPT-6 Astra (Staged)†
63.18
GTA (Ours)†
18.80

15.4× the throughput of GPT-6 Astra (Original)

System-level tokens per second (×10³)

Claude Fable 5.2
1.89
GPT-6 Astra (Original)†
1.47
GPT-6 Astra (Staged)†
1.61
GTA (Ours)†
22.67

† GPT-6 Astra (Original), GPT-6 Astra (Staged), and GTA use the same GPT-6 Astra backbone with medium reasoning effort. Original is the original baseline; Staged is a single-agent baseline following the same three-stage workflow as GTA: planning, asset generation, and placement & refinement. GTA results are measured with 32 agentic slots.

GTA for Embodied AI

Citation

@misc{tao2026gta,
  title   = {{GTA}: Generative Tower of Agents for Scalable Holonic Scene Generation},
  author  = {Tao, Zeng and Chen, Yunuo and Feng, Xiang and He, Yumeng and Li, Peihao and Jiang, Nanxiang and Guo, Yaowei and Zeng, Fanlian and Yang, Yin and Jiang, Chenfanfu},
  year    = {2026},
  eprint  = {2611.23727},
  archivePrefix = {arXiv},
  url     = {https://arxiv.org/abs/2611.23727}
}