AI
Scoped Semantic Invariance as a Foundation for LLM Reliability
Publication: cs.AI updates on arXiv.org
Author: Rosario Uceda-Sosa, Maria Chang, Karthikeyan Natesan Ramamurthy, Moninder Singh
Source published: 27/10/2026, 04:00:00 UTC
Briefing created: 22/11/2026, 16:39:37 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 28/10/2026, 02:33:25 UTC.
Briefings
Research formalizes semantic invariance as a foundation for LLM reliability, proving that invariance under arbitrary meaning-preserving transformations is undecidable. The study operationalizes this via knowledge graphs, identifying structured inconsistencies concentrated in localized hotspot subgraphs (TREE, STAR, CYCLE patterns) across MITRE D3FEND and eight Wikidata domains. Selective graph grounding improves correctness, while indiscriminate grounding causes regressions. Hotspot structures from consistency probes transfer to compositional reasoning tasks.
Key points
- Semantic invariance is proven undecidable when modeled under arbitrary meaning-preserving transformations.
- Inconsistencies in LLMs are structured and localized into hotspot subgraphs rather than random errors.
Context
This research addresses LLMs' inconsistent answers to semantically equivalent queries by formalizing semantic invariance as a reliability foundation.
Limitations
Evidence covers only eight Wikidata-derived domains and the MITRE D3FEND ontology; findings are preliminary and based on initial consistency probes.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 22/11/2026, 16:39:37 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeigeupq3zuohi76kfhdzcboure6xslwmykh7uwnxlhxpllvzpe34lm