Despite the remarkable progress in neural models, their ability to generalize—a cornerstone for applications like logical reasoning—remains a critical challenge. We delineate two fundamental aspects of this ability: compositionality, the capacity to abstract atomic logical rules underlying complex inferences, and recursiveness, the aptitude to build intricate representations through iterative application of inference rules. In the literature, these two aspects are often confounded together under the umbrella term of generalization. To sharpen this distinction, we investigated the logical generalization capabilities of pre-trained large language models (LLMs) using the syllogistic fragment as a benchmark for natural language reasoning. We extend classical Aristotelian syllogistic forms to build more complex structures, providing a foundational yet expressive subset of formal logic that supports controlled evaluation of essential reasoning abilities. Our findings reflect this non-trivial benchmark: while LLMs demonstrate reasonable proficiency in recursiveness, they struggle with compositionality. This disparity, however, is not uniform, as a more detailed analysis reveals variability in generalization performance across individual syllogistic types, ranging from near-perfect to significantly lower accuracy. To overcome these limitations and establish a reliable logical prover, we propose a hybrid architecture integrating symbolic reasoning with neural computation. This synergistic interaction enables robust and efficient inference—neural components accelerate processing, while symbolic reasoning ensures completeness. Our experiments show that high efficiency is preserved even with relatively small neural components. As part of our proposed methodology, this analysis provides a rationale and highlights the potential of hybrid models to effectively address key generalization barriers in neural reasoning systems.

Abstract We study the isomorphism relation on Borel classes of locally compact Polish metric structures. We prove that isomorphism on such classes is always classifiable by countable structures (equivalently: Borel reducible to graph isomorphism), which implies, in particular, that isometry of locally compact Polish metric spaces is Borel reducible to graph isomorphism. We show that potentially $\boldsymbol {\Pi }^{0}_{\alpha + 1}$ isomorphism relations are Borel reducible to equality on hereditarily countable sets of rank $\alpha $ , $\alpha \geq 2$ . We also study approximations of the Hjorth-isomorphism game, and formulate a condition ruling out classifiability by countable structures.

Abstract We study the complexity of isomorphism of classes of metric structures using methods from infinitary continuous logic. For Borel classes of locally compact structures, we prove that if the equivalence relation of isomorphism is potentially $\mathbf {\Sigma }^0_2$ , then it is essentially countable. We also provide an equivalent model-theoretic condition that is easy to check in practice. This theorem is a common generalization of a result of Hjorth about pseudo-connected metric spaces and a result of Hjorth–Kechris about discrete structures. As a different application, we also give a new proof of Kechris’s theorem that orbit equivalence relations of actions of Polish locally compact groups are essentially countable.

Abstract We define a simple criterion for a homogeneous, complete metric structure X that implies that the automorphism group Aut(X) satisfies all the main consequences of the existence of ample generics: it has the automatic continuity property, the small index property, and uncountable cofinality for nonopen subgroups. Then we verify it for the Urysohn space $$ , the Lebesgue probability measure algebra MALG, and the Hilbert space $\ell _2 $ , thus proving that Iso( $$ ), Aut(MALG), $U\left( {\ell _2 } \right)$ , and $O\left( {\ell _2 } \right)$ share these properties. We also formulate a condition for X which implies that every homomorphism of Aut(X) into a separable group K with a left-invariant, complete metric, is trivial, and we verify it for $$ , and $\ell _2 $ .

Abstract We prove that the set of all Polish groups admitting a compatible complete left-invariant metric (called CLI) is coanalytic non-Borel as a subset of a standard Borel space of all Polish groups. As an application of this result, we show that there does not exist a weakly universal CLI group. This, in particular, answers in the negative a question of H.Becker.

Abstract We construct a non-discrete Polish group none of whose non-discrete subgroups admits a complete left-invariant metric.