Coherence-Based Defense Against Adversarial Attacks on Large Language Models v1.2
Position paper proposing a coherence-monitoring approach to adversarial attacks on large language models, with an explicit falsification condition. No experimental results are reported. v1.2 corrects metadata and citation errors from v1.0, tightens the scope of the compression-attack signature, operationalizes the context-coupling measure, and states the principal limitation — coherence-preserving attacks — as a named boundary rather than an open question.