Researchers built a natural-language-inference system that classifies sentence pairs without ever letting its classifier see the sentences.
The pipeline first breaks each sentence into atomic propositions, then converts those into ConceptNet triples using constrained decoding. Premise, hypothesis, and a retrieved ConceptNet subgraph are each turned into a graph, and only those three graphs get fed into a fine-tuned 0.8-billion-parameter model. On SNLI, the graph-only system hit 89.7% accuracy, just 1.9 points behind an identically trained text-based model. On ANLI, it matched published RoBERTa-large numbers on the harder rounds - 48.0% vs. 48.9% on R2, 44.9% vs. 44.4% on R3 - but fell 16 points behind on R1. Measured separately against the researchers' own text-based counterpart model, the system's overall ANLI gap came out to 9 to 14 points, a different baseline than the round-by-round RoBERTa-large comparison. Combining graphs with text pushed SNLI accuracy to 92.1%, ahead of either approach alone.
The pitch here isn't raw accuracy, it's a paper trail. Every decision the classifier makes can be traced back to specific extracted propositions and graph edges, instead of buried somewhere in a transformer's attention weights. That matters for any setting where someone needs to check why a model called two statements contradictory, not just accept the verdict.
Call it the price of interpretability, which is what the researchers themselves call it: legible reasoning costs a few accuracy points on easy cases and a lot more on the adversarial ones. R1's 16-point shortfall is the tell - that round is designed to trip up simple pattern-matching, and it's exactly where compressing a sentence into propositions and graph triples throws away the nuance a full-text model would have caught.