A new preprint on arXiv introduces the Graph Dirichlet Energy Causal Inference (GDECI) criterion for distinguishing cause from effect when only two variables are observed. The method is designed for bivariate settings, where standard causal inference often struggles without additional assumptions.
For a candidate direction (X \to Y), the criterion constructs a kernel similarity graph from (X) and treats (Y) in a way that the available abstract does not fully describe. The title indicates the approach is fitting-free, meaning it avoids explicit model fitting when scoring causal directions.
Because only the abstract was provided, no experimental results, baselines, or comparisons are available in this digest. The write-up is limited to the proposal itself, and no claims about performance are made beyond what the source states.