Robust Detection of LLM-Generated Text under Contamination
Abstract
We study the detection of LLM-generated text under editing and contamination. Modeling human and machine text as finite-order Markov processes with Huber contamination, we characterize an exact boundary for reliable detection under our assumptions. Detection is impossible when contamination is sufficiently large relative to clean-source separation. Below this boundary, a collection of clipped likelihood-ratio tests achieves vanishing worst-case errors. This construction motivates clipping as a simple modification of existing statistical detectors. For a broad class of additive scores, we identify conditions under which the clipped test is consistent while the raw test's worst-case power tends to zero. We evaluate seven detectors across three datasets and three generation models, and on the RAID benchmark. Clipping improves robustness in both studies, with gains varying across detectors and contamination settings. For example, at a target false-positive rate of 5\%, clipping improves the log-likelihood--log-rank ratio (LRR) detector's true-positive rate by a median of 8.3 percentage points in the controlled study and 2.1 and 4.3 points in rate- and attack-specific RAID evaluations, respectively.
Related
- Embedded Conditional Independence Tests for Large Language Model Generated Text with an Application to German Parliament Speeches
- Robust Dual-Regularized Variable Selection under Outlier Contamination
- Robust dimension-free estimation of simple random tensors: optimal guarantees under heavy tails and adversarial contamination
- Generalised Robust Bayes for Joint Inference of Model and Contamination
- Robust Active Learning for Few-Shot Example Selection in Text-to-SQL
BibTeX
@misc{li2026robust,
title = {{Robust Detection of LLM-Generated Text under Contamination}},
author = {Jiaxun Li and Saptarshi Chakraborty and Ambuj Tewari},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.29935},
eprint = {2609.29935},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.29935},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Robust Detection of LLM-Generated Text under Contamination
AU - Jiaxun Li
AU - Saptarshi Chakraborty
AU - Ambuj Tewari
PY - 2026
DA - 2026/09/24
PB - arXiv
DO - 10.48550/arxiv.2609.29935
UR - https://arxiv.org/abs/2609.29935
AB - We study the detection of LLM-generated text under editing and contamination. Modeling human and machine text as finite-order Markov processes with Huber contamination, we characterize an exact boundary for reliable detection under our assumptions. Detection is impossible when contamination is sufficiently large relative to clean-source separation. Below this boundary, a collection of clipped likelihood-ratio tests achieves vanishing worst-case errors. This construction motivates clipping as a simple modification of existing statistical detectors. For a broad class of additive scores, we identify conditions under which the clipped test is consistent while the raw test's worst-case power tends to zero. We evaluate seven detectors across three datasets and three generation models, and on the RAID benchmark. Clipping improves robustness in both studies, with gains varying across detectors and contamination settings. For example, at a target false-positive rate of 5\%, clipping improves the log-likelihood--log-rank ratio (LRR) detector's true-positive rate by a median of 8.3 percentage points in the controlled study and 2.1 and 4.3 points in rate- and attack-specific RAID evaluations, respectively.
ER -