Answers are grounded in this paper and cite their sources.
Demonstrates a new metric for model consistency in language models, suggesting improved input processing.
We introduce the Noema Consistency Index (NCI), a metric that measures whether a language model has formed a stable conceptual representation of its input by computing cosine similarity between hidden states at layers 3 and 28. On Apertus-8B, NCI distinguishes meaningful inputs from noise (0.88 vs 0.53), correlates with input type, requires only 2 of 32 layers (94% reduction), and needs no token generation or external knowledge. A complementary activation norm at layer 3 provides additional signal. Applications include input validation, early exit, and hallucination detection.
No takes yet. Share an insight, caveat, or question.
László Popovics (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: