Understanding as Holding a Concept Together and the Testing of Language Models (Towards a Research Methodology)

The article poses the question of whether it is legitimate to infer that language models understand a given subject domain from the fact that they solve difficult test items. Such an inference is shown to be illegitimate — which is not to claim that models do not understand — and the error is named: the middle term of the syllogism is undistributed. The error goes unnoticed for a reason that has nothing to do with logic: in S.S. Averintsev’s analysis of Byzantine rhetoric, the author shows that a canonical form of exposition confers credibility upon a conclusion before its content has been verified. A high percentage in a table of language-model evaluation results is precisely such a contemporary form. The logical failure points to a substantive one: benchmarks test knowledge by assessing the correctness of judgements, whereas the conclusion is drawn about understanding, whose unit is not the judgement but the connection. To object on the merits, no agreement about the nature of understanding is required: it suffices to specify an observable consequence that must follow from the attribution of understanding. Such a consequence exists. A concept, the semantic whole standing behind a word, is given in several forms — in the notion, the image and the action; the whole is present in each of them, and understanding is therefore expressed in the ability to move between the forms, recognising in one the same thing as in another. If the transition does not take place and a break is observed, the concept is not being held as a whole, even where command of the separate forms is present. The operating mode of the criterion is established: failure of a transition refutes the attribution of understanding, while success does not certify it — and for the purposes of evaluation the former is sufficient. The methods used are: logical and categorial analysis, operationalisation of the construct, historico-cultural analogy and the model case, external validation through adjacent disciplines.