Advances in Agentic Computer Graphics

Advances in Agentic Computer Graphics. By Hengyu Liu and thirteen others, from the University of Texas at Austin, Stanford University, and three other institutions. Published on the first of November.

Computer graphics is built on code. Large language model agents can now generate graphics code and refine it iteratively through various kinds of feedback. That capability opens the door to what may be a new paradigm for the field, one in which graphics algorithms are increasingly built around agentic systems that automate progressively complex tasks. Over the past years, research on agentic graphics has expanded rapidly, across a diverse range of tasks.

Two problems come with that growth. Manual literature reviews find it increasingly hard to keep pace with new developments. And similar ideas are often described using different terminology from one task to another, which makes a systematic understanding of the field harder to build. The volume is one obstacle; the vocabulary is another.

The report answers with two contributions. The first is a unified taxonomy for characterising agentic systems across diverse computer graphics tasks: a shared frame, so that systems built for different tasks can be described in the same terms. That matters because the same idea, given a different name in each task, is otherwise hard to recognise as the same idea. The second contribution is itself agentic. The authors develop an agentic literature system that can assist human researchers in analysing existing and newly published papers at scale. A field moving too fast for manual review is surveyed, in part, with the same kind of tool it studies.

Building on those contributions, the report analyses emerging research trends and discusses open challenges. Four are named, and each reads as a place where today's agentic systems still strain. Coordinating agents and tools across graphics tasks, where a single problem may involve several of each. Designing interfaces and representations for human-agent collaboration. Sustaining long-horizon planning and refinement over many steps. And reliably evaluating interactive graphics systems.

The two contributions answer the two problems one for one. Inconsistent terminology gets a unified taxonomy. A literature that outpaces manual review gets an agentic system to help read it. Both serve the same end: a systematic understanding of a field where graphics code is increasingly written and refined by agents, and where the open questions now concern coordination, collaboration with people, long horizons, and evaluation.