Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

When AI Settles Down: Late-Stage Stability as a Signature of AI-Generated Text Detection

Ke Sun, Guangsheng Bao, Han Cui, Yue Zhang

TL;DR

This work tackles AI-generated text detection in zero-shot settings by revealing a temporal signature of autoregressive generation: Late-Stage Volatility Decay, where AI-generated text shows rapidly stabilizing log-probability fluctuations in the later portions of text. The authors introduce two simple, training-free features, Derivative Dispersion and Local Volatility, computed from the second half of sequences, and combine them into a Temporal Stability Detection (TSD) score that achieves state-of-the-art performance on EvoBench and MAGE without perturbation sampling. The method is shown to be complementary to existing global detectors, and fusion with a global detector further boosts robustness, especially for longer texts and diverse model families. The findings highlight the importance of leveraging temporal dynamics in autoregressive generation for robust AI-text detection and suggest practical, scalable detection that generalizes across surrogate models and frontier models. Limitations include reduced effectiveness on very short texts and simple fusion strategies that could be enhanced with adaptive weighting.

Abstract

Zero-shot detection methods for AI-generated text typically aggregate token-level statistics across entire sequences, overlooking the temporal dynamics inherent to autoregressive generation. We analyze over 120k text samples and reveal Late-Stage Volatility Decay: AI-generated text exhibits rapidly stabilizing log probability fluctuations as generation progresses, while human writing maintains higher variability throughout. This divergence peaks in the second half of sequences, where AI-generated text shows 24--32\% lower volatility. Based on this finding, we propose two simple features: Derivative Dispersion and Local Volatility, which computed exclusively from late-stage statistics. Without perturbation sampling or additional model access, our method achieves state-of-the-art performance on EvoBench and MAGE benchmarks and demonstrates strong complementarity with existing global methods.

When AI Settles Down: Late-Stage Stability as a Signature of AI-Generated Text Detection

TL;DR

This work tackles AI-generated text detection in zero-shot settings by revealing a temporal signature of autoregressive generation: Late-Stage Volatility Decay, where AI-generated text shows rapidly stabilizing log-probability fluctuations in the later portions of text. The authors introduce two simple, training-free features, Derivative Dispersion and Local Volatility, computed from the second half of sequences, and combine them into a Temporal Stability Detection (TSD) score that achieves state-of-the-art performance on EvoBench and MAGE without perturbation sampling. The method is shown to be complementary to existing global detectors, and fusion with a global detector further boosts robustness, especially for longer texts and diverse model families. The findings highlight the importance of leveraging temporal dynamics in autoregressive generation for robust AI-text detection and suggest practical, scalable detection that generalizes across surrogate models and frontier models. Limitations include reduced effectiveness on very short texts and simple fusion strategies that could be enhanced with adaptive weighting.

Abstract

Zero-shot detection methods for AI-generated text typically aggregate token-level statistics across entire sequences, overlooking the temporal dynamics inherent to autoregressive generation. We analyze over 120k text samples and reveal Late-Stage Volatility Decay: AI-generated text exhibits rapidly stabilizing log probability fluctuations as generation progresses, while human writing maintains higher variability throughout. This divergence peaks in the second half of sequences, where AI-generated text shows 24--32\% lower volatility. Based on this finding, we propose two simple features: Derivative Dispersion and Local Volatility, which computed exclusively from late-stage statistics. Without perturbation sampling or additional model access, our method achieves state-of-the-art performance on EvoBench and MAGE benchmarks and demonstrates strong complementarity with existing global methods.
Paper Structure (47 sections, 6 equations, 10 figures, 9 tables)

This paper contains 47 sections, 6 equations, 10 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Temporal Dynamics of Zero-Shot Detection Metrics
  3. Analysis Setup
  4. Datasets and Models.
  5. Zero-Shot Detection Metrics.
  6. Temporal Features.
  7. Comparative Metric Sensitivity
  8. Not All Metrics Capture Temporal Dynamics.
  9. Late-Stage Volatility Decay
  10. Quantitative Characters.
  11. Interpretation.
  12. Summary of Findings
  13. Method
  14. Late-Stage Stability Metrics
  15. Derivative Dispersion (DD).
  16. ...and 32 more sections

Figures (10)

  • Figure 1: Temporal dynamics of log probability features for human vs. AI text on EvoBench (top) and MAGE (bottom). Left: Absolute derivative (change rate); Right: Local standard deviation (volatility). Shaded regions show the human-AI gap, with $\Delta$ indicating mean difference (AI $-$ Human) and relative percentage. Yellow regions (50--100%) highlight larger divergence in the second half.
  • Figure 2: Temporal dynamics of six base metrics on EvoBench. Each row shows a metric's raw value (left), absolute derivative (middle), and local standard deviation (right) across normalized token positions (0--100%). Log probability and Sampling Discrepancy (alias Conditional Probability Curvature) display clear human-AI separation that widens in the second half, while entropy and top-$k$ concentration show minimal divergence.
  • Figure 3: TSD performance (AUROC) versus average token length across EvoBench sources. Each point represents a dataset from one of seven model families. The dashed line shows linear regression fit with Pearson correlation $r = 0.54$.
  • Figure 4: Relationship between average token length and temporal detection performance (AUROC) across MAGE generator families. The dashed line shows linear regression fit with Pearson correlation $r = 0.66$
  • Figure 5: TSD performance versus relative start position. The midpoint (50%) achieves optimal performance on both benchmarks.
  • ...and 5 more figures