DWT-Fusion: Treating Token Probabilities as Signals — What Wavelet Transforms Mean for AI-Generated Text Detection
> Paper: DWT-Fusion: A Signal-Based Framework for Training-Free LLM-Generated Text Detection > Authors: Mehmet Batuhan Özdaş, Murat Osmanoğlu > Institution: Ankara University, Cyber Security Vocational School > arXiv: 2607.22026 (July 24, 2025)
> TL;DR: This article analyzes the core findings and engineering implications of DWT-Fusion, which treats token probabilities as a signal and uses wavelet transforms to detect AI-generated text.
The Scenario
You receive a student's essay and want to know whether the student wrote it or ChatGPT did.
You have an open-source language model (e.g., GPT-2 or LLaMA) that can compute the log probability of every token in the text. How do you use those probabilities to decide whether the text is AI-written?
The most intuitive approach: compute the average log probability of the whole text. AI-generated text usually has higher probability (more "fluent"); human-written text usually has lower probability (less "fluent").
But this approach has a fatal flaw: a fluent human writer also produces high probabilities, and AI text deliberately made "less fluent" to mimic humans may not score high. The global average masks local differences.
DWT-Fusion's core insight: don't look at the global average — look at local, multi-scale fluctuations in the probabilities.
Turning Text into a Signal
The first step is converting text into a 1-D signal.
Concretely: use a proxy language model (e.g., GPT-2 Medium) to compute the conditional log probability of each token. That sequence of numbers is a one-dimensional signal — like an audio waveform.
Then apply the Discrete Wavelet Transform (DWT) to this signal.
Why Wavelets?
The wavelet transform is a classic signal-processing tool. Its core idea: decompose a signal into "details" at different scales.
Analogy: looking at a photo. Global statistics (average brightness, variance) tell you about overall tone but reveal no detail. The Fourier transform tells you which frequency components exist but not where they occur in the image. The wavelet transform tells you both "what frequency" and "where" — it has time-domain (position) and frequency-domain (scale) information simultaneously.
DWT-Fusion uses wavelets because AI-vs-human differences may not be global but local and multi-scale:
- Local differences: AI may be especially "fluent" (high, stable probability) in some paragraphs but deliberately "less fluent" in others to imitate humans. A global average cancels these out.
- Multi-scale differences: AI probability fluctuations may be small at short scales (a few tokens) but show regular patterns at long scales (tens of tokens). Human text fluctuates differently.
- HC3: Chinese AI-text detection, relatively easy
- M4: multi-generator, multi-domain, medium difficulty
- MAGE: multi-generator, multi-domain, multilingual, hardest
- AUROC of only 0.7471 on MAGE. In practice this implies a fairly high false-positive rate — insufficient for high-precision use cases like academic integrity enforcement.
- Proxy-model dependence. The method relies on an open-source model for token probabilities. For text generated by GPT-4, GPT-2 probabilities may not discriminate well. The paper includes a proxy-model sensitivity analysis but did not test GPT-4-generated text.
- Adversarial robustness unverified. If an attacker knows the detector uses wavelets, they could generate text with "human-like multi-scale fluctuations." Such adversarial scenarios were not tested.
- Gap vs. supervised SOTA. The best supervised detectors (RADAR, Ghostbuster) reach AUROC 0.85+ on M4; DWT-Fusion's 0.8477 approaches but doesn't exceed them. Whether the training-free advantage compensates depends on the deployment scenario.
The wavelet transform decomposes the signal into multi-level "approximation coefficients" (low-frequency, global trends) and "detail coefficients" (high-frequency, local changes), capturing both kinds of difference.
Three Wavelet-Domain Scores
DWT-Fusion defines three scalar scores:
1. First-level detail energy: intensity of the finest-scale local fluctuations. AI text usually fluctuates less at this scale (more "uniform"). 2. Multilevel detail energy: total detail energy across all scales, capturing overall multi-scale fluctuation strength. 3. Window-energy variability: split the signal into windows and measure energy variation across them, capturing the "unevenness" of local fluctuations.
These scores are independent and can be used alone or combined.
Calibration-Guided Voting Fusion
Beyond single-score detection, DWT-Fusion introduces Calibration-Guided Voting Fusion.
The idea: compute scores under multiple wavelet configurations (different wavelet families, different decomposition depths), let each configuration cast a "vote," then weight the votes with calibration weights.
It's like ensemble learning — combining weak detectors into a strong one. Calibration weights ensure better-performing configurations contribute more.
Experimental Results
The authors tested on three datasets:
Best single-score results:
| Dataset | AUROC | |---------|-------| | HC3 | 0.9872 | | M4 | 0.8185 | | MAGE | 0.7138 |
After calibration-guided voting fusion:
| Dataset | AUROC | |---------|-------| | HC3 | 0.9919 | | M4 | 0.8477 | | MAGE | 0.7471 |
Near-perfect on HC3 (0.99+) with significant gains on M4 and MAGE. Notably, these results were achieved training-free — no labeled data, no classifier training, just an open-source proxy model.
Comparison with a DFT Baseline
A key ablation: replace DWT with the Discrete Fourier Transform (DFT).
DFT provides frequency-domain information only, with no time-domain information. The result: the DFT baseline underperforms DWT-Fusion on all datasets. This directly proves the importance of "local + multi-scale" information — signal processing per se isn't what helps; it's the specific time-frequency joint structure of wavelets.
Engineering Significance
1. Zero training cost. No labeled data, no classifier training — just an open-source proxy model. Extremely low deployment cost. 2. Interpretability. All three scores have clear physical meaning (local fluctuation, multi-scale fluctuation, window variability) — not black boxes. You can analyze why a passage was flagged as AI-generated. 3. Configurability. Wavelet family and decomposition depth are tunable for different scenarios.
An Honest Assessment
Limitations of this paper:
One-Sentence Summary
DWT-Fusion treats token probabilities as a one-dimensional signal and uses the wavelet transform to capture both local and multi-scale probability fluctuations — no training, no labels, just an open-source model — approaching the performance of supervised detectors.
---
Paper: https://arxiv.org/abs/2607.22026 HTML version: https://arxiv.org/html/2607.22026v1 Open-source code: not yet available