The Truncation Blind Spot: How Decoding Strategies Systematically Exclude Human-Like Token Choices

Abstract

Standard decoding strategies for text generation, including top-k, nucleus sampling, and contrastive search, select tokens based on likelihood, restricting selection to high-probability regions. Human language production operates differently: tokens are chosen for communicative appropriateness rather than statistical frequency. This mismatch creates a truncation blind spot: contextually appropriate but statistically rare tokens remain accessible to humans yet unreachable by likelihood-based decoding. We hypothesize this contributes to the detectability of machine-generated text. Analyzing over 1.8 million texts across eight language models, five decoding strategies, and 53 hyperparameter configurations, we find that 8-18% of human-selected tokens fall outside typical truncation boundaries. Simple classifiers trained on predictability and lexical diversity achieve remarkable detection rates. Crucially, neither model scale nor architecture correlates strongly with detectability; truncation parameters account for most variance. Configurations achieving low detectability often produce incoherent text, indicating that evading detection and producing natural text are distinct objectives. These findings suggest detectability is enhanced by likelihood-based token selection, not merely a matter of model capability.

Figures (10)

• Figure 1: Distribution of predictability (also referred to as coherence su2022contrastivegarces-arias-etal-2024-adaptiveding-etal-2025-guard) and lexical diversity scores for human-authored text (blue) and machine-generated text from Falcon 2 - 11B (red) across multiple decoding configurations and datasets. The separation in this feature space is consistent with the hypothesis that likelihood-based truncation produces systematically different token distributions than human language production.
• Figure 2: Token ranks (log scale) of human-written WikiNews text as a function of token position. Red highlights mark human-selected tokens that would be excluded under top-$k$ truncation ($k=10$), with exclusion rates reported per model. Top: GPT2-XL (1.5B). Bottom: Qwen2-7B. Sequence lengths differ due to model-specific tokenization.
• Figure 3: Binary classification performance for proprietary models GPT-3.5-turbo and Claude-3-Haiku. Despite advanced training procedures and alignment, both remain highly detectable using only predictability and diversity features.
• Figure 4: Effect of architecture and scale on AUC-ROC, comparing Transformer- and Non-Transformer-based architectures at different sizes.
• Figure 5: AUC-ROC heatmaps for contrastive search showing interactions between $k$ and $\alpha$ parameters. Darker regions indicate lower detectability.