Which feature of PyTorch is particularly appealing to researchers in AI?

Dynamic computation graphs are particularly appealing to researchers in AI because they offer flexibility in model design and experimentation. With dynamic computation graphs, also known as "define-by-run" paradigms, the graph is created on-the-fly as operations are executed. This allows researchers to modify the model architecture easily during runtime, making it simpler to implement new ideas, try out different configurations, and debug models.

This flexibility is vital in research environments where rapid prototyping and iteration are essential. For instance, researchers often need to experiment with varying inputs, batch sizes, and model parameters, and the ability to easily change structures without the need to recompile or reconstruct the entire model helps streamline this process.

The other features listed may have their benefits, but they do not provide the same level of adaptability that is critical for research. Static computation graphs, for example, require defining the entire computation graph beforehand, which can hinder exploration and creativity in developing new models or approaches. High-level abstractions can streamline coding and may improve development speed, but they may also abstract away the control and flexibility that researchers often need. Integration with R, while beneficial for certain data science workflows, does not directly contribute to the core requirements for dynamic experimentation in AI research.