On the Salience of Low-Probability Tokens for AI-Generated Text Detection: A Multiscale Uncertainty Perspective

Abstract: AI-generated text increasingly blends with human writing, raising practical risks such as misinformation, academic misuse, and corpora contamination. While statistical detectors are appealing for efficiency and generalization, they suffer from two key limitations. (i) Boilerplate dominance, boilerplate tokens shared across human and LLM writing can overwhelm discriminative signals. (ii) Brittle point estimates, relying on a single probability score yields unstable decisions under adversarial manipulations. To address these issues, we propose Uncertainty, a multiscale uncertainty estimator that focuses on informative low-probability tokens, which more clearly expose distributional discrepancies. Locally, it alleviates boilerplate dominance by averaging the log-probabilities of low-probability tokens; globally, it reduces brittleness by capturing the distributional shape of this low-probability region via Rényi entropy. We further extend the detector to Uncertainty++ via conditional independent sampling, yielding a more stable uncertainty estimation. Experiments across seven datasets and sixteen LLMs demonstrate high effectiveness, generalization, and robustness. Our code is available at https://github.com/guoyikai2000/Uncertainty-AIGT.

• The paper demonstrates that focusing on low-probability token regions significantly improves the discrimination of AI-generated text.
• It uses multiscale uncertainty measures, combining local percentile-based log-probabilities with global Rényi entropy, to enhance robustness.
• Empirical results show improved AUROC and stability across various datasets and adversarial conditions, supporting its practical integration.

Multiscale Low-Probability Token Salience for Robust AI-Generated Text Detection

Introduction

Distinguishing AI-generated text from human-written content remains a critical challenge, exacerbated by the increasing fluency and adaptability of LLMs. Traditional statistical detection frameworks often aggregate token probability signals across the entire text, making them susceptible to two core structural weaknesses: dominance of high-probability "boilerplate" tokens that dilute discriminative signal, and brittleness arising from reliance on pointwise probability estimates that are highly sensitive to adversarial perturbations. This work, "On the Salience of Low-Probability Tokens for AI-Generated Text Detection: A Multiscale Uncertainty Perspective" (2606.02158), proposes a principled approach—Uncertainty and Uncertainty++—that prioritizes low-probability token regions and incorporates multiscale uncertainty measures to overcome these detection pathologies.

Methodological Framework

Low-Probability Token Hypothesis

The central empirical and theoretical hypothesis underpinning the framework is that low-probability tokens—positions where the model finds observed token realizations most surprising—carry substantially more discriminative evidence for separating AI-generated from human-written text than do high-probability, boilerplate sequences. Quantitative analysis (e.g., on XSum, with LLaMA3-8B as the source and GPT-J as proxy, $p=0.15$) shows a 3.5× increase in Human-AI LogRank gap at low-probability positions over their high-probability counterparts, and an AI/Human probability ratio of nearly 10× at the low-probability tail.

Multiscale Uncertainty: Local and Global Components

• Local (Percentile-Based) Uncertainty: Instead of uniform aggregation, tokens are ranked per-sequence by log-probability under a (proxy) LLM. The detector computes the mean log-probability over a bottom $p$ percentile (typically $p \simeq 15\%$), producing a local uncertainty statistic that magnifies discriminability while suppressing the influence of boilerplate.
• Jensen Discrepancy and Sampling-Based Normalization: To account for length and domain effects, the method further normalizes the tail aggregation via conditional independent sampling—analogous to the Fast-DetectGPT strategy but applied selectively to low-probability tokens. The normalized discrepancy exploits the concavity of the percentile operator, with a strong "Jensen amplification" observed only for AI text.
• Global (Entropy-Based) Uncertainty: For robustness against perturbations (e.g., paraphrasing or decoding changes), the framework incorporates Rényi entropy computed on the conditional token distribution at each low-probability position. This global statistic, averaging Rényi entropy over the selected low-probability set, captures distributional shape information that is more stable under surface text rewrites than are point probability estimates.
• Unified Score Fusion: The Uncertainty detector fuses local and global scores by a learned convex combination; Uncertainty++ uses the normalized local term and fuses with global entropy for greater robustness.

Empirical Evaluation

Benchmarks and Experimental Regimes

The evaluation spans seven datasets (XSum, SQuAD, WritingPrompts, Reddit, FiQA, Wiki-csai, arXiv) and sixteen contemporary LLMs, including high-profile closed- and open-source models (GPT-5, Gemini-3-Flash, LLaMA3, etc.) for both black-box (proxy) and white-box (self-evaluated) settings. Nine statistical baselines—probability-based (Likelihood, LogRank, DetectLRR, Lastde) and sampling-based (DetectGPT, DetectNPR, DNA-GPT, Fast-DetectGPT, Lastde++)—serve as comparators. Primary evaluation is via AUROC, augmented by low FPR metrics (TPR@1%FPR, TPR@5%FPR).

Main Results

• Detection Performance: Uncertainty++ achieves the highest mean AUROC in both settings, with black-box average at 93.24%, exceeding the strongest baseline (Lastde++) by 1.73 points and consistently outperforming across all domains and model families.
• Ablation Study: Removal of low-probability aggregation causes significant AUROC drop (e.g., $-6.74$ points on Reddit), establishing the necessity of tail-focused aggregation. Excluding the log-probability local term results in catastrophic failure ($-31.80$ AUROC), while removing global Rényi entropy produces only minor loss, demonstrating that entropy acts as a stabilizing rather than primary signal.
• Generalization: Cross-domain evaluations confirm the framework's domain-transferability, with Uncertainty++ besting all likelihood-based and sampling baselines on FiQA, Wiki-csai, and arXiv. Detection accuracy trails in the highly specialized FiQA domain, which exposes remaining limitations of language-model proxies for out-of-domain text statistics.

Robustness

• Decoding and Paraphrasing Attacks: Both Uncertainty and Uncertainty++ retain superior and stable AUROC under strong decoding variations (temperature, top-p, frequency penalty), where most baselines degrade substantially. Under adversarial paraphrasing, although all methods incur performance loss, Uncertainty++ consistently outperforms sampling and likelihood baselines, but high-quality paraphrasers (GPT-5) still erode detection accuracy—an open challenge.

Efficiency

The Uncertainty detector maintains high throughput among all probability-based methods (27.27 samples/sec, 12.7 GB VRAM on a single A100), and Uncertainty++ is competitive among sampling-based frameworks, supporting practical deployment at scale.

Generalization of Low-Probability Token Strategy

The low-probability selection paradigm generalizes: applying low-probability aggregation to traditional detectors (Likelihood, LogRank, Fast-DetectGPT) unambiguously improves their AUROC, lifting even the sub-optimal full-sequence Likelihood scores to supermajority (>90%) discrimination. Visualization of score distributions confirms drastically improved separation when only the low-probability tail is considered.

Theoretical Significance and Implications

This research substantiates, both empirically and analytically, the hypothesis that discriminative signal for AI-generated text is sharply concentrated in regions of low model confidence—an observation that aligns with emerging trends in related tasks such as membership inference for LLMs, where low-probability evidence offers higher statistical leverage [see also: Min-K%++ strategies]. Formally, the authors provide proofs showing that the percentile aggregation operator is concave (enabling favorable Jensen amplification specifically for AI output) and that Rényi entropy at low-probability positions is robust to adversarial token redistributions.

Practical and Theoretical Implications

• Practical: The framework is effective for zero-shot, label-free AI-generated text detection across architectures, text genres, and models, including recent commercial LLMs. Robustness to perturbation and high throughput make it suitable for integration into real-world content moderation, academic integrity, and dataset hygiene systems. However, detection on very short or highly domain-specific texts remains problematic unless proxy distribution calibration is improved.
• Theoretical: The focus on low-probability tokens and multiscale uncertainty represents a paradigm shift away from single-point estimates and uniform aggregation, with implications for broader detection strategies in both generative and forensic NLP. The concavity-induced signal amplification and the entropy-based stability lend themselves to principled extensions, e.g., adaptive tail selection, more sophisticated fusion with domain-awareness, or mutual information regularization to further suppress adversarial vulnerability.

Future Directions

Potential directions include refined uncertainty calibration to further remediate signal sparsity in short texts, domain-adaptive proxy modeling to address domain shift, and cross-task applications to other forms of model-generated data provenance. Moreover, extending the multiscale uncertainty approach and tail selection strategy to non-text modalities or joint multi-modal generative detection is a promising avenue.

Conclusion

The Uncertainty/Uncertainty++ framework rigorously establishes the salience of low-probability tokens for robust, generalizable, and efficient AI-generated text detection, providing a unified multiscale uncertainty perspective that addresses fundamental limitations of prior statistical detectors. The work’s methodological and empirical contributions delineate a clear path for future advances in robust generative content forensics and forensic AI.