Structured Frequency-Domain Evidence for LLM-Based Time-Series Anomaly Detection
This paper proposes an evidence-augmented zero-shot time-series anomaly detection framework that enhances LLM performance by integrating compact global and local frequency-domain evidence, computed via FFT, to complement existing time-domain inputs and better capture structural anomalies like shifted periodicity.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Time-series data is the rhythm of the modern world, a continuous stream of numbers recording everything from the pulse of a heart to the temperature of a server room. In these streams, the most dangerous signals are often not the loudest spikes, but the subtle shifts in the pattern itself. A machine might begin to vibrate at a slightly different speed, or a financial market might lose its usual daily cycle, long before the numbers themselves crash. Detecting these changes is the job of anomaly detection, a field that has traditionally relied on specialized computer programs trained on vast amounts of historical data. However, a new approach is emerging that asks a different question: can a general-purpose artificial intelligence, one that has never seen this specific data before, learn to spot these irregularities just by looking at the numbers?
For years, researchers have tried to teach these large language models to understand time by feeding them lists of numbers or pictures of graphs. The models are surprisingly good at this, but they often miss the hidden structure of the data. They see the values rising and falling, but they struggle to understand the underlying beat—the periodicity that defines how the data repeats over time. It is as if a listener could hear a song but could not tell if the drummer had suddenly sped up or slowed down, because the music was presented only as a list of notes rather than a rhythm. This gap in understanding limits the ability of artificial intelligence to catch the most insidious types of errors, those that hide in the timing rather than the volume.
A team of researchers at Hanyang University in South Korea has proposed a way to bridge this gap. They developed a method that gives these artificial intelligence models a new kind of sense: the ability to see the frequency of the data. Instead of just showing the model a list of numbers or a graph, they added a compact summary of the data's rhythm, calculated using a mathematical tool called the Fast Fourier Transform. This tool breaks the complex stream of numbers down into its constituent cycles, revealing the dominant beats and how they change over time. The researchers did not ask the model to learn a new skill or train it on thousands of examples. Instead, they simply provided this rhythmic summary as extra evidence alongside the original data, allowing the model to make a decision based on a much clearer picture of what was happening.
The results of their experiments suggest that this extra layer of information is transformative. When the researchers tested their method on a benchmark containing various types of anomalies, the models equipped with this frequency evidence performed significantly better than those without it. The improvement was most dramatic when the anomalies involved changes in the data's rhythm, such as a sudden shift in periodicity or a localized fluctuation that did not look like a simple spike. In these cases, the models were able to pinpoint the exact moment the pattern broke, whereas the models without the frequency summary often missed the change entirely or guessed the wrong location. The researchers found that providing both a global view of the entire sequence's rhythm and a local view of how that rhythm changed in specific windows gave the models the best results.
This approach does not replace the need for the models to analyze the raw numbers, but it acts as a powerful guide. The study shows that when an artificial intelligence is given explicit information about the frequency structure of a time series, it can reason about anomalies with greater accuracy. The models were able to distinguish between normal variations and genuine structural changes, even when those changes were subtle. The researchers tested their method on several different large language models, including some of the most advanced available, and found that the benefit held true across the board. This suggests that the ability to understand frequency is a fundamental piece of the puzzle that was previously missing from how these models interact with time-series data.
The findings also highlight a specific limitation in how these models currently process information. Without the frequency summary, the models struggled to recover the periodic structure of the data from the raw numbers alone. They could see the values, but they could not reliably infer the underlying cycles. By adding the frequency evidence, the researchers effectively gave the models a map of the terrain, allowing them to navigate the data with a much sharper sense of direction. This does not mean the models are now perfect; they still make mistakes, particularly with certain types of anomalies that are defined purely by amplitude rather than rhythm. However, the addition of this structured evidence significantly raised the floor of their performance, making them more reliable tools for real-world monitoring.
In the end, this work demonstrates that the way we present data to artificial intelligence matters as much as the data itself. By translating a time series into a format that includes both its values and its rhythmic structure, researchers can unlock a deeper level of understanding in these models. The study suggests that for tasks involving time, where the pattern of change is often more important than the change itself, providing explicit frequency-domain evidence is a crucial step forward. It is a reminder that even the most advanced artificial intelligence benefits from a little help in seeing the world in the right way, turning a confusing stream of numbers into a comprehensible story of rhythm and change.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.