The DEV Community has announced a technical challenge focused on building Transformer-based models for temporal reasoning on dynamic graphs. Participants must create a system that predicts the next state of a graph, which features nodes that appear or disappear over time. The core task involves integrating graph attention with an external memory component, pushing the boundaries of current neural network architectures.

Developers are required to work within strict constraints, including a 10-step temporal context window and a maximum of 100 nodes per graph. The architecture mandates using a Transformer encoder-decoder setup. Success is measured by accuracy, F1-score, and how efficiently the model utilizes its memory. This mirrors industry efforts to make AI models more resource-efficient.

Submissions are open until February 28, 2026, with a total prize pool of $2,250. Cash rewards will be awarded for top performance in accuracy, F1-score, and memory utilization. Participants must share their code and results publicly. This event tests practical skills in advanced graph neural networks, a key area for future AI applications.