Paper review · November 4, 2026
Research trends in agent-based computer graphics
Advances in Agentic Computer Graphics
The authors organized more than 300 studies that use LLM agents for graphics tasks with a unified four-axis taxonomy and an agent-based literature analysis system, and presented research trends and open problems.
Problem
Research on LLM agents that generate graphics code and revise it based on feedback is growing quickly. The papers the authors collected alone number more than 300 as of September 2026. Manual literature review struggles to keep up with this pace. In addition, fields such as CAD, animation and simulation use different names for similar components (feedback modules, planning modules and so on), which makes it hard to compare studies.
Method
The authors first define the scope. It covers systems in which (1) an LLM decides at least part of the actions, (2) actual tasks are executed in a graphics environment, and (3) feedback on intermediate results changes the next action. These systems are classified along four axes: Specification, foundation model, harness (planning and execution, feedback, memory), and graphics interfaces and tools. Each axis has defined options. For example, planning is divided into Hard-Coded, Hybrid and Non Hard-Coded, and the verifier is divided into self-verification, Critic, program verification, environment-based verification and human verification. On top of this, the authors built an agent-based literature analysis system that searches for and reads papers and then annotates them with this taxonomy. The system records classification labels together with supporting sentences, and later runs can add new papers or revise existing annotations.
Results
- The authors collected more than 300 papers published through September 2026 and showed that the number of studies is growing quickly.
- The authors proposed a common terminology made up of four axes (specification, foundation model, harness and graphics interface) and their detailed options. They applied it consistently to analyze 2D, static 3D and dynamic 3D applications.
- Based on the analysis, the authors identified four open problems: coordinating agents and tools across tasks, designing interfaces and representations for human-agent collaboration, maintaining planning and revision over long tasks, and reliable evaluation of interactive graphics systems.
Limitations
As a survey paper, it has no quantitative performance comparison of a new method. Within the provided text, results verifying the accuracy of the annotations made by the agent-based literature analysis system could not be found. By the scope definition, it does not cover dedicated networks trained from scratch, LLM approaches that generate in one pass without feedback, or approaches that use feedback only for training or final evaluation.
Why it matters
For practitioners who want to attach LLM agents to graphics tools handled through code, such as Blender, CAD and SVG, it shows the design options at a glance. The options include what to receive feedback from, how fixed the plan should be, and whether to output a program or a rendering. The approach of using agents to survey a fast-growing field is also worth noting.
Key terms
- Harness
- The execution layer that connects the foundation model's decisions to actual tool execution. It passes execution results, evaluations and memory back to the model so that the next action responds to the results.
- Agent-based graphics system
- A system in which an LLM operates graphics tools and repeatedly revises its work based on feedback such as rendering results or execution results.
- Critic
- A separate agent that evaluates outputs it did not create. It may use the same model as the acting agent.
Original
Authors: Hengyu Liu, Jihyeon Je, Ruiqi Wang, Mingxuan Wu, Haven Feng, Qixing Huang, Georgios Pavlakos, Hao Zhang, et al. · Posted 2026-11-01 · License CC BY 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.