Detecting Contextual Hallucinations in Large Language Models with Frequency-Aware Attention

Hallucination detection is critical for ensuring the reliability of large language models (LLMs) in context-based generation. Prior work has explored intrinsic signals available during generation, among which attention offers a direct view of grounding behavior. However, existing approaches typically rely on coarse summaries that fail to capture fine-grained instabilities in attention. Inspired by signal processing, we introduce a frequency-aware perspective on attention by analyzing its variation during generation. We model attention distributions as discrete signals and extract high-frequency components that reflect rapid local changes in attention. Our analysis reveals that hallucinated tokens are associated with high-frequency attention energy, reflecting fragmented and unstable grounding behavior. Based on this insight, we develop a lightweight hallucination detector using high-frequency attention features. Experiments on the RAGTruth and HalluRAG benchmarks show that our approach achieves performance gains over verification-based, internal-representation-based, and attention-based methods across models and tasks.

Large language models sometimes confidently state facts that aren't actually supported by the documents they were given to read, a problem known as "hallucination." Catching these errors automatically is essential before we can trust AI. When a language model writes each word, it "pays attention" to different parts of the input text, somewhat like a reader's eyes darting between sentences on a page. Previous detection methods asked where the model looked or how widely it scanned, but they missed something more subtle. Inspired by signal processing, we treated this attention as a signal, similar to a sound recording, and checked whether the signal was smooth or jittery. We found that when a model is making things up, its attention signal becomes oscillates rapidly, even when its overall focus looks reasonable. By measuring this jitter with classic tools like the Fourier transform, we built a small, fast detector that flags fabricated words as they are generated. Our detector outperforms previous methods across multiple models and benchmarks, suggesting that a model's trustworthiness shows not only in where it looks, but in how steadily it looks there.