Large language models (LLMs) are increasingly evaluated on reasoning tasks, yet their logical abilities remain contested. To address this, we study LLMs’ reasoning in a well-defined fragment of logic: syllogistic reasoning. We cast the problem as premise selection and construct controlled datasets to isolate logical competence. Beyond evaluation, an open challenge is enabling LLMs to acquire abstract inference patterns that generalize to novel structures. We propose to apply few-shot meta-learning to this domain, thereby encouraging models to extract rules across tasks rather than memorize patterns within tasks. Although meta-learning has been little explored in the context of logic learnability, our experiments show that it is effective: small models (1.5B–7B) fine-tuned with meta-learning demonstrate strong gains in generalization, with especially pronounced benefits in low-data regimes. These meta-learned models outperform GPT-4o and o3-mini on our syllogistic reasoning task.

We study the ability of neural and hybrid models to generalize logical reasoning patterns. We created a series of tests for analyzing various aspects of generalization in the context of language and reasoning, focusing on compositionality and recursiveness. We used them to study the syllogistic logic in hybrid models, where the network assists in premise selection. We analyzed feed-forward, recurrent, convolutional, and transformer architectures. Our experiments demonstrate that even though the models can capture elementary aspects of the meaning of logical terms, they learn to generalize logical reasoning only to a limited degree.

The paper presents a tool for automatic marking up of quantifying expressions, their semantic features, and scopes. We explore the idea of using a BERT based neural model for the task (in this case HerBERT, a model trained specifically for Polish, is used). The tool is trained on a recent manually annotated Corpus of Polish Quantificational Expressions (Szymanik and Kieraś, 2022). We discuss how it performs against human annotation and present results of automatic annotation of 300 million sub-corpus of National Corpus of Polish. Our results show that language models can effectively recognise semantic category of quantification as well as identify key semantic properties of quantifiers, like monotonicity. Furthermore, the algorithm we have developed can be used for building semantically annotated quantifier corpora for other languages.

We study the role of linguistic context in predicting quantifiers (‘few’, ‘all’). We collect crowdsourced data from human participants and test various models in a local (single-sentence) and a global context (multi-sentence) condition. Models significantly out-perform humans in the former setting and are only slightly better in the latter. While human performance improves with more linguistic context (especially on proportional quantifiers), model performance suffers. Models are very effective in exploiting lexical and morpho-syntactic patterns; humans are better at genuinely understanding the meaning of the (global) context.

Among the readings available for NL sentences, those where two or more sets of entities are independent of one another are particularly challenging from both a theoretical and an empirical point of view. Those readings are termed here as ‘Independent Set (IS) readings'. Standard examples of such readings are the well-known Collective and Cumulative Readings. (Robaldo, 2011) proposes a logical framework that can properly represent the meaning of IS readings in terms of a set-Skolemization of the witness sets. One of the main assumptions of Robaldo's logical framework, drawn from (Schwarzschild, 1996), is that pragmatics plays a crucial role in the identification of such witness sets. Those are firstly identified on pragmatic grounds, then logical clauses are asserted on them in order to trigger the appropriate inferences. In this paper, we present the results of an experimental analysis that appears to confirm Robaldo's hypotheses concerning the pragmatic identification of the witness sets.