KoLing
KoLing: A bias-controlled adversarial benchmark for Korean LLM-text detection
KoLing is a 50.9k-instance benchmark for detecting LLM-generated and modified Korean text across essays, abstracts, and editorials, controlling for surface shortcuts.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info koling
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
11 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Binoculars | 66.4 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| NPR | 62.7 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Fast-DetectGPT | 62.2 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| DetectGPT | 60.3 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| LLM-Det | 58.7 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Log-Rank | 56.2 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Log-Likelihood | 55.8 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Rank | 54.8 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| LRR | 54.0 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| GECScore | 53.4 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Entropy | 53.0 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
KoLing comprises 50.9k instances of human-written, LLM-modified, and LLM-generated Korean text across three domains. It employs strict shortcut control to eliminate length and punctuation biases.
The benchmark evaluates binary and ternary detection performance. Experiments show that existing detectors rely on superficial cues, while the proposed KoLingProbe achieves robust performance using linguistic features.