Tag
IBM researchers formalize LLM consistency as scoped semantic invariance over knowledge-graph-defined semantic domains, showing inconsistencies cluster in stable hotspot subgraphs and that selective, layered KG grounding improves cross-model agreement and reliability.
FormInv proposes a measurement protocol for evaluating semantic invariance in mathematical reasoning benchmarks, revealing that model rankings reverse across paraphrase families and that standard accuracy metrics conceal large gaps in semantic consistency.