An arXiv paper (2610.27439) identifies a core bottleneck in graph reasoning for language reasoning models (LRMs): representing a graph as a serialized description causes the input length to scale with the number of edges. This makes large graphs expensive or impractical to feed into a model.
The proposed solution, indicated by the title, is to use learnable graph positional encodings instead of relying solely on serialized graph descriptions. The abstract text cuts off before describing the method in detail or presenting experimental results, so the source supports only the problem statement and the general direction of the approach.
No evaluation claims are made in the available text, and the paper's findings cannot be assessed from this excerpt alone.