Extending Minimal Pairs with Ordinal Surprisal Curves and Entropy Across Applied Domains

Standard language model evaluation assigns scores to single predicted answers, rewarding high-confidence responses regardless of how residual probability mass is distributed over alternative options. This creates a systematic pressure toward overconfident guessing: under accuracy-based schemes, a model maximises its expected score by always committing to an answer rather than abstaining, even when its uncertainty is high. While penalty-based approaches partially address this by raising the confidence threshold for strategic guessing, they still treat all sub-threshold responses identically, ignoring a fundamental distinction in how models can express uncertainty - for example between hedging toward incorrect answers versus hedging toward "I don't know" responses. This paper introduces a novel evaluation metric to solve this problem of not considering a model's entire probability distribution over answer choices. The metric naturally distinguishes between harmful overconfidence in wrong answers and uncertainty expressed through abstention, providing scores in an interpretable default range. Through theoretical analysis and illustrative examples, the metric is shown to offer a more nuanced and aligned evaluation paradigm that incentivises models to express genuine uncertainty rather than guessing. Adapting 12 existing evaluation benchmarks to the metric's variants and measuring performance on six language models shows that for half of the tested benchmarks scores are negative across all tested models, indicating significant tendencies towards hallucination.

While language models provide powerful mechanisms for evaluating textual predictability, it remains unclear whether the relative extents of LM surprisal remain stable across models. This study presents empirical evidence of the invariance of ranking based on cross-entropy related measures (e.g. mean, variance of each data-point), across four distinct small scale transformer-based models, for a given dataset of text sources. We consider two stylistic environments: a corpus of contemporary poetry and a corpus of academic research papers. We demonstrate that mean and variance of cross-entropy maintain robust cross-model rank correlation ($ρ> 0.90$) across both regimes. To test whether this agreement merely reflects generic differences in textual complexity, we compare it against six parameter-free compression baselines. The extent of agreement among language models is found to be substantially higher than that among model-baselines, for most of the measures. Geometric analysis of the structurally constrained academic regime reveals that higher-order moments are sensitive to specific models, although strong agreement persists. The findings suggest that relative extents of textual unpredictability may be independent of the language model they are computed against.

Surprisal Theory is often characterized as a computational-level explanation per (Marr, 1982). We argue in this work that, even though a computational level narrative has been used to support "representation-agnostic research" within computational psycholinguistics, the movement toward black box systems embodied by large language models (LLMs) does not exempt modelers using the surprisal metric from the representational decisions required by computational-level characterizations. In fact, we argue that the uncritical use of LLM-surprisal obfuscates the representational and algorithmic-level commitments of different models. In three analyses, we show that the choice of algorithm and model architecture play significant roles in the computation of language model probabilities. We advise that researchers who wish to test Surprisal Theory re-evaluate the practice of treating large language model probabilities as interchangeable