Research map / Large language models / Language, safety and interpretability

Alignment Imprint: Zero-Shot AI-Generated Text Detection via Provable Preference Discrepancy

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: Zero-shot machine-text detection · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LAPD, a zero-shot AI-generated text detector that standardizes an information-weighted log-likelihood ratio between an aligned LLM and its base model, with theoretical dominance claims over Fast-DetectGPT.

Problem

Likelihood-based detection of AI-generated text is confounded by content complexity and performs unstably, especially in high-entropy regions and at low false-positive rates.

Approach

Derives the log-likelihood ratio between an aligned model and its base model (the Alignment Imprint, RAI) as decomposing into SFT instructional bias and preference reward terms; weights per-token imprints by self-information and standardizes them using conditional-independence perturbation statistics (LAPD).

Key results

• LAPD reaches 92.37% average AUROC over four benchmarks versus 86.12% for the strongest baseline (DNA-DetectLLM) and 82.26% for Fast-DetectGPT, which the paper reports as a 45.82% relative gain under its (new-old)/(1-old) metric.
• LAPD averages 99.49% AUROC across XSum, WritingPrompts, and Arxiv for three source LLMs, versus 95.36% for Fast-DetectGPT.
• At 0.5% FPR, LAPD averages 92.27% TPR versus 66.67% for Binoculars and 59.56% for DNA-DetectLLM (averaged over the three generated-text datasets).

Surprising findings

• The raw alignment imprint (RAI) alone scores 96.19% AUROC on DetectRL Multi-Domain, slightly above LAPD's 96.11%.
• Likelihood-based scores from the base or aligned model alone perform near chance on XSum (about 70%), yet the log-ratio of the two is much stronger.

Limitations

• Theoretical guarantees rely on strong asymptotic assumptions (indistinguishability, prescriptiveness, diversity) and the alignment process is abstracted as two constrained optimization steps.
• Requires access to both a base and an aligned model of the same family, and the paper notes performance varies across model pairs (e.g., Falcon-7B pair 89.01% average AUROC).
• Code is stated as 'coming soon' and no data release is described.

Experimental setup

Datasets: M4 DetectRL RAID RealDet XSum WritingPrompts Arxiv

Benchmarks: M4 DetectRL Multi-LLM DetectRL Multi-Domain RAID RealDet XSum WritingPrompts Arxiv

Metrics: AUROC TPR at 0.5% FPR time per sample

Baselines: Entropy Likelihood LogRank DetectGPT Fast-DetectGPT Lastde++ Binoculars DNA-DetectLLM ReMoDetect ImBD

Builds on: Fast-DetectGPT DPO RLHF DetectGPT

Closest papers by meaning

• ReMoDetect: Reward Models Recognize Aligned LLM's Generations (NeurIPS 2024 · similarity 0.89)
• Exploring Intrinsic Alignments Within Text Corpus (AAAI 2025 · similarity 0.88)
• Noise Contrastive Alignment of Language Models with Explicit Rewards (NeurIPS 2024 · similarity 0.88)
• Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text via Conditional Probability Curvature (ICLR 2024 · similarity 0.88)
• Imitate Before Detect: Aligning Machine Stylistic Preference for Machine-Revised Text Detection (AAAI 2025 · similarity 0.88)
• ComPO: Preference Alignment via Comparison Oracles (NeurIPS 2025 · similarity 0.88)