SV-Detect: AI-generated Text Detection with Steering Vectors
About
Detecting machine-generated text is especially difficult under distribution shift, such as transfer across domains, source models, and editing attacks. We propose a fake-text detector based on steering vectors extracted from the hidden representations of a frozen language model. At each layer, we construct a direction that separates human-written from machine-generated text, and represent each input by its layer-wise alignment with these directions. A lightweight classifier trained on these projection features yields the final detection score. Our method achieves strong performance both in-distribution and under distribution shift, including across domains, source models, and machine-editing transformations such as polishing and rewriting. Interpretation analyses show that the learned directions align with recognizable stylistic cues while capturing substantial additional signal beyond surface features. These results position fake-text detection as a representation-space probing problem and show that steering vectors provide a simple and effective solution.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Machine-generated text detection | COLING MGT English 2025 | Macro F184.8 | 41 | |
| AI-generated text detection | DetectRL Multi-LLM | -- | 27 | |
| AI-generated text detection | DetectRL Multi-Domain | -- | 27 | |
| Machine-generated text detection | DetectRL GPT-3.5 | AUROC99.99 | 16 | |
| Machine-generated text detection | DetectRL Llama-2 | AUROC99.99 | 16 | |
| Machine-generated text detection | DetectRL Direct Attack | AUROC99.99 | 16 | |
| Machine-generated text detection | DetectRL Prompt Attack | AUROC99.96 | 16 | |
| Machine-generated text detection | DetectRL Paraphrase Attack | AUROC99.98 | 16 | |
| Machine-generated text detection | DetectRL Perturbation Attack | AUROC99.99 | 16 | |
| Machine-generated text detection | DetectRL Mixing Attack | AUROC (%)99.83 | 16 |