The paper investigates why Transformers struggle more with structural than lexical compositional generalisation. It argues that this disparity stems from low structural type diversity rather than an inherent limitation of Transformers. By creating linguistically diverse variants of the COGS and SLOG datasets, the authors show that type diversity correlates equally with generalisation in both lexical and structural cases, challenging previous explanations of the difficulty.
arXiv:2604. 26157v4 Announce Type: replace-cross Abstract: Structural generalization in semantic parsing requires systems to apply learned compositional rules to novel structural combinations.
The study investigates whether pretrained transformer models encode functional words—such as pronouns and adverbs—in a way that mirrors human usage. By comparing embeddings of nouns with those of their functional counterparts in both isolated and parallel sentences, the authors find that functional words occupy a central yet distinct position in embedding space and that parallel lexicalized and functional sentences reside in different subspaces. Experiments show that only a mixed training set of functional and lexicalized sentences reveals shared syntactic and semantic structure, whereas training on either type alone fails to capture this parallelism.
The paper proposes a new way to evaluate compositional generalization by examining which structural or lexical identifications allow held‑out COGS examples to be considered admissible based on training data. Sentences are modeled as functors from syntactic addresses to lexical tokens, and selective collapses induce Kan extensions that propagate observed associations. Across 21 COGS generalization types, admissibility follows distinct identification profiles, while residual failures highlight unsupported structural templates, providing data‑side diagnoses of what the training corpus licenses without training a predictive model.
The paper surveys over 300 recent NLP studies to address fragmentation in diversity measurement. It proposes an NLP‑specific framework with four perspectives—why, what, where, and how—drawing on unified cross‑domain frameworks from ecology and economics. The authors find that variety and disparity measures dominate, balance measures are scarce, and many current metrics do not fit Stirling’s framework, leading to recommendations for unified terminology and normative debate.
arXiv:2605.27268v2 Announce Type: replace-cross Abstract: Modern Large Language Models (LLMs) are often criticized for producing repetitive and homogeneous text, despite possessing vast latent vocabu...
arXiv:2505. 20161v2 Announce Type: replace-cross Abstract: Effective generalization in language models depends critically on the diversity of their training data.
arXiv:2606.27228v2 Announce Type: replace Abstract: Formal semantics has shown that sentence meanings arise by recursively composing lexical meanings, yet much of the literature on semantic universal...
arXiv:2606.05087v2 Announce Type: replace Abstract: Frequent verbs such as 'have' and 'make' can function either as collocates in light-verb constructions or as full lexical predicates, as in 'make a...
The paper proposes a depth‑aware FrankenMerging technique that blends lower layers of an aligned large language model (LLM) with higher layers of its base counterpart to balance diversity and quality without extra training. Experiments demonstrate that adjusting the layer split controls both lexical and semantic diversity, often surpassing the aligned model in semantic diversity and outperforming temperature sampling in most settings. The method is validated by an ablation that reverses the merge order, supporting the hypothesis about layer‑specific diversity contributions.
arXiv:2610.35102v1 Announce Type: new Abstract: Synthetic supervision can bootstrap semantic parsers in new search domains, but prompt-induced diversity is useful only when generations are faithful t...
The paper introduces a new measure of generative‑process diversity for language models, using Normalised Compression Distance on raw outputs after controlling for permutation effects. Across 38 models, this metric uncovers population structure that semantic similarity misses and predicts lower correlated failures across ten benchmark families, independent of semantic similarity or model capability. The authors argue that higher generative‑process diversity reduces correlated failures in multi‑model systems, offering a practical tool for safety‑relevant applications.
The paper examines how language models learn action nominal constructions (ANCs) across 96 synthetic target languages with varied grammatical features. Using GPT‑2‑small trained from scratch, the study evaluates models on minimal pairs covering 34 grammatical phenomena, revealing that models mirror typological patterns in basic clause configurations but diverge in complement systems and ANC strategies. The findings suggest that natural language distributions may arise from both general learning pressures and functional demands not fully captured by current model experiments.