ViCo: Visual-oriented Coding with Self-Reflection for Chart Replication

ViCo tackles a very specific but increasingly relevant problem for game and tools teams: getting AI coding agents to produce visuals that actually match a reference, not just something technically valid. In this case the target is academic chart replication, but the underlying issue is the same one that shows up in UI generation, telemetry dashboards, editor tooling, and any workflow where layout fidelity matters as much as code correctness.

The framework adds iterative self-reflection to the coding loop so the model can revise its output toward a reference image step by step. To make that work, the system uses a self-supervised warm-up stage with Monte Carlo Tree Search and consistency-based pruning to build better reflection trajectories. That is paired with a multi-step RL setup that assigns credit separately to reflection and action steps, which helps with the sparse reward problem that usually makes visual refinement hard to train.

ViCo also introduces an automated evaluation pipeline that scores charts across style, layout, and semantic consistency using a hierarchical heterogeneous layout graph. That matters because visual quality is expensive to judge at scale, and manual review does not fit large training runs. The authors report results on three public benchmarks, with an 8B model reaching performance close to proprietary LLMs when reflection is available.

For developers, the practical takeaway is that visual reasoning is becoming a first-class constraint for coding agents. If these techniques generalize, they could improve AI-assisted generation of graphs,...