English

H-Node Attack and Defense in Large Language Models

Machine Learning 2026-03-30 v1 Artificial Intelligence Computation and Language Neural and Evolutionary Computing

Abstract

We present H-Node Adversarial Noise Cancellation (H-Node ANC), a mechanistic framework that identifies, exploits, and defends hallucination representations in transformer-based large language models (LLMs) at the level of individual hidden-state dimensions. A logistic regression probe trained on last-token hidden states localizes hallucination signal to a small set of high-variance dimensions -- termed Hallucination Nodes (H-Nodes) -- with probe AUC reaching 0.90 across four architectures. A white-box adversarial attack amplifies these dimensions at inference time via a real-time forward hook, achieving a selectivity of 3.02x with less than 10% visibility to the defender. Adaptive ANC defense suppresses H-Node excess in-pass using confidence-weighted cancellation, reducing grounded activation drift by 33-42% over static cancellation. A dynamic iterative extension that re-ranks cancellation targets across successive passes recovers up to 0.69 robustness from a single-pass baseline of 8%. All contributions are validated on OPT-125M, Phi-3-mini-4k-instruct, LLaMA-3-8B-Instruct, and Mistral-7B-Instruct-v0.3 (125M-8B parameters). Perplexity impact is surgical (<5%) and MMLU degradation is at most 3%, confirming that the defense does not impair general reasoning capability.

Cite

@article{arxiv.2603.26045,
  title  = {H-Node Attack and Defense in Large Language Models},
  author = {Eric Yocam and Varghese Vaidyan and Yong Wang},
  journal= {arXiv preprint arXiv:2603.26045},
  year   = {2026}
}

Comments

17 pages, 7 figures, 6 tables

R2 v1 2026-07-01T11:40:10.445Z