EchoPrompt's AUROC rises with text length, from an average of 75.8% in the shortest 1-40 word bin to 98.9% in the 321-360 bin on DetectRL Length¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Short texts are challenging for all evaluated zero-shot detectors, but EchoPrompt remains competitive in the shortest bin and its performance rises rapidly with length. → Hongrui Bao 2026

Evidence¶

Hongrui Bao 2026¶

Hongrui Bao, Yubing Ren, Jinhan You, Fang Fang, Shi Wang, Yanan Cao. (2026). ONCE A RESPONSE, ALWAYS A RESPONSE: DETECTING LLM-GENERATED TEXT VIA LATENT PROMPT RESTORATION. arXiv preprint. https://arxiv.org/abs/2608.05741 doi:10.48550/arxiv.2608.05741

q2 · i? · design · r2

Length-binned AUROC analysis (Figure 4) comparing EchoPrompt with DNA-DetectLLM, IRM, Binoculars and Fast-DetectGPT on DetectRL Length across three Llama-family proxies, showing monotonic improvement with text length.

"Short texts are challenging for all detectors because they provide fewer tokens for estimating reliable detection signals. Even in the shortest 1-40 word bin, EchoPrompt remains competitive, with an average AUROC of 75.8% across the three Llama-family proxies. As length increases, its performance rises rapidly to 93.1% in the 81-120 bin and 98.9% in the 321-360 bin."

Discussion¶

Related Claims¶

• EchoPrompt maintains cross-proxy robustness, averaging 87.06% AUROC across seven proxy models and outperforming DNA-DetectLLM by 2.70% AUROC on average — related
• EchoPrompt remains robust across attack settings, obtaining the best AUROC and Best-F1 in four out of five attack groups and improving over IRM by 0.24% AUROC on average — related
• The detection gain comes from restoring the missing prompt-response relation, not merely assigning an assistant role: the full prefix outperforms the empty prompt by 14.73% and 12.57% AUROC on Qwen2.5-3B and 5.63% and 5.33% on Llama-3.1-8B — related
• Detector accuracy declined significantly as text length increased, with both detectors performing best on short texts — reports the opposite
• LLM-generated responses were substantially longer than authentic student submissions, suggesting response length as a signal the detector may exploit — related
• LLM self-detection performance is strongly task-dependent: reliable for programming tasks and longer reflective responses but poor for short-answer questions — a narrower finding that bears on this claim