We propose a method to evaluate the extent to which an LLM’s observable input–output behavior aligns with established theories in the humanities and cultural studies. We instantiate the framework on three humanities theories—Davidson’s truth-conditional semantics, Lewis’s truth in fiction, and Iser’s concept of textual gaps—using a top-down, theory-driven black-box framework. Core assumptions of these theories are reconstructed into testable behavioral rules and assessed via controlled classification tasks with systematic prompt comparisons and significance testing. Our experiments show that theory-uninformed classification prompts generally outperform theory-enriched prompts in Lewis and Iser settings, while theory-informed prompts help in the Davidson task. Gemini Flash consistently achieves the highest scores across tasks and corpora, while the Iser gap detection task remains substantially harder than binary truth-conditional judgments. Statistical tests confirm robust prompt effects and the failure of basic prompts. However, model behavior under incremental theory exposure is unstable and architecture-dependent.

Prompt engineering holds the promise for the computational literary studies (CLS) to obtain high quality markup for literary research questions by simply prompting large language models with natural language strings. We test prompt engineering’s validity for two CLS sequence labeling tasks under the following aspects: (i) how generalizable are the results of identical prompts on different dataset splits?, (ii) how robust are performance results when re-formulating the prompts?, and (iii) how generalizable are certain fixed phrases added to the prompts that are generally considered to increase performance. We find that results are sensitive to data splits and prompt formulation, while the addition of fixed phrases does not change performance in most cases, depending on the chosen model.

In this paper, we evaluate two different natural language processing (NLP) approaches to solve a paradigmatic task for computational literary studies (CLS): the recognition of knowledge transfer in literary texts. We focus on the question of how adequately large language models capture the transfer of knowledge about family relations in German drama texts when this transfer is treated as a classification or textual entailment task using in-context learning (ICL). We find that a 13 billion parameter LLAMA 2 model performs best on the former, while GPT-4 performs best on the latter task. However, all models achieve relatively low scores compared to standard NLP benchmark results, struggle from inconsistencies with small changes in prompts and are often not able to make simple inferences beyond the textual surface, which is why an unreflected generic use of ICL in the CLS seems still not advisable.