CS639 · University of Wisconsin–Madison · 2026

INSCONE: Unknown-Aware Detection of LLM-Generated Text via Informed Wild Data

Informed SCONE for zero-shot generalization to unseen LLM families

Mark Stanley, Masa Abboud, Saira Khatoon, Fairoz Khan, Samad Syed
University of Wisconsin–Madison
Paper Code RAID+
Energy distributions by split

Energy distributions under the trained baseline. ID and covariate LLMs cluster at low energy; human text sits near the OOD margin. Zero-shot unseen models track the covariate seen distribution — the geometry INSCONE exploits.

Abstract

AI-generated text is increasingly difficult to distinguish from human writing, creating risks in academic integrity, medical misinformation, and social media disinformation. While recent work has reframed MGT detection as an out-of-distribution problem with strong in-distribution and zero-shot results, generalization to LLMs unseen during training remains underexplored.

We propose INSCONE (Informed SCONE), which adapts the SCONE wild-data framework to the text domain by exploiting curated wild data with known mixing proportions (πid, πc, πs) to stabilize the energy geometry around seen and unseen LLM families. INSCONE achieves a 6.1-point FPR95 improvement over a competitive baseline on the RAID benchmark. We additionally release RAID+, an extended evaluation set regenerating RAID prompts with contemporary frontier models.

Code and data: github.com/markstanl/INSCONE · huggingface.co/datasets/markstanl/RAID-Plus

Key Insight

Standard SCONE uniformly pushes all wild samples toward high energy. This is a reasonable assumption in vision where covariate shift is pixel-level. In text, covariate shift means an entirely different model family. The push term overwhelms the Lipschitz drag and incorrectly drives unseen LLM embeddings toward the human text region.

⬅️

Proximal anchor

Wild samples below the quantile threshold τprox are estimated as covariate OOD and pulled toward low energy, anchoring unseen LLMs near the ID region.

➡️

Distal anchor

Wild samples above τdist are estimated as semantic OOD (human text) and pushed toward high energy. A buffer δ around the boundary receives no gradient.

Method

Energy-Based Baseline

We build on HTAO's energy-based detector, which uses a frozen SimCSE-RoBERTa encoder to map text into an embedding space where LLM-generated text clusters tightly at low energy and human text diffuses toward high energy. A lightweight classification head over known LLM families sharpens this geometry during training.

INSCONE Wild Loss

INSCONE instead utilizes the known composition of the wild batch to split samples into two groups: those likely to be machine-generated (pulled toward low energy) and those likely to be human (pushed toward high energy). A small tunable buffer around the boundary receives no gradient, preventing noisy updates from ambiguous samples.

Wild energy quantile curve

Empirical wild batch energy distribution. Wild ratios set to 10% ID, 60% OOD Covariate, and 30% OOD Semantic. Green = proximal (covariate) region anchored at low energy; red = distal (human) region pushed high; gray = buffer receiving no gradient. ~99.2% of proximal samples are machine text; ~99.6% of distal samples are human.

Proximal region

The lowest-energy wild samples are estimated as covariate OOD (unseen LLMs) and pulled toward the ID energy region.

Distal region

The highest-energy wild samples are estimated as human text and pushed further toward high energy.

Dead zone

Samples near the composition boundary receive no gradient — preventing spurious updates from ambiguous wild samples.

Results

Evaluated on RAID temporal split.

Method AUROC ↑ Wild FPR95 ↓ Zero-Shot FPR95 ↓
Standard SCONE 0.9298 0.4702 0.4486
Fair Ablation (20k labeled) 0.9371 0.1552 0.3166
Baseline (10k) 0.9520 0.1646 0.2862
INSCONE (ours) 0.9506 0.1764 0.2252

INSCONE achieves the best zero-shot FPR95, improving 6.1 points over the naive baseline and 20.3 points over standard SCONE, at near-identical AUROC.

RAID+ Dataset

We regenerate RAID prompts using frontier models absent from the original benchmark, providing an evaluation set for testing detector behavior against unseen contemporary LLMs not present in any existing MGT benchmark.

ModelSamples
Gemini-3.1-Pro2,000
DeepSeek-V32,000
Gemma-3-27B2,000
LLaMA-3.3-70B2,000
Total8,000

Available on 🤗 at markstanl/RAID-Plus.

BibTeX

@misc{stanley2025inscone,
  title   = {INSCONE: Unknown-Aware Detection of LLM-Generated Text
             via Informed Wild Data},
  author  = {Stanley, Mark and Abboud, Masa and Khan, Fairoz and
             Khatoon, Saira and Syed, Samad},
  year    = {2025},
  url     = {https://github.com/markstanl/INSCONE}
}