This work investigates whether small-scale LMs can benefit from instruction tuning (IT). We compare conversational and question–answering IT datasets, applied either in a merged or sequential curriculum, using decoder-only models with 100M and 140M parameters. Evaluation spans both fine-tuning (SuperGLUE) and zero-shot (BLiMP, EWoK, WUGs, entity tracking, and psycholinguistic correlation) settings. Results show that IT yields small but consistent gains in fine-tuning scenarios, with sequential curricula outperforming merged data; however, improvements do not consistently transfer to zero-shot tasks, suggesting a trade-off between interaction-focused adaptation and broad linguistic generalization. These results highlight both the potential and the constraints of adapting human-inspired learning strategies to low-resource LMs, and point toward hybrid, curriculum-based approaches for enhancing generalization under ecological training limits.

This paper investigates how decoder-only instruction-tuned LLMs handle lexical ambiguity. Two distinct methodologies are employed: Eliciting rating scores from the model via prompting and analysing the cosine similarity between pairs of polysemous words in context. Ratings and embeddings are obtained by providing pairs of sentences from Haber and Poesio (2021) to the model. These ratings and cosine similarity scores are compared with each other and with the human similarity judgments in the dataset.Surprisingly, the model scores show only a moderate correlation with the subjects’ similarity judgments and no correlation with the target word embedding similarities. A vector space anisotropy inspection has also been performed, as a potential source of the experimental results. The analysis reveals that the embedding spaces of two out of the three analyzed models exhibit poor anisotropy, while the third model shows relatively moderate anisotropy compared to previous findings for models with similar architecture (Ethayarajh 2019). These findings offer new insights into the relationship between generation quality and vector representations in decoder-only LLMs.

The possibility of comparing the linguistic competence of Language Models (LMs) to that of children has gained growing attention lately, raising the need for effective tools for evaluating both the former and the latter. To this purpose, we developed a resource for the linguistic evaluation of BabyLMs, which are LMs trained on datasets that comparable to the linguistic stimulus received by children. This resource adapts four standardized tests for the evaluation of linguistic skills of Italian-speaking children (BVL, TROG-2, TCGB-2 and Peabody). To verify the effectiveness of our benchmark, we administered it to Minerva, a LLM pretrained from scratch on Italian. Our results indicate that Minerva struggles to master certain linguistic aspects, achieving an age-equivalent score of 4 years, and that the type of task administered affects the model’s performance.

We present a model for the Strict-Small track of the BabyLM Challenge 2024 (Choshen et al. 2024). We introduce a Curriculum Learning approach for training a specialized version of GPT-2 (Radford et al. 2019), that we name ConcreteGPT. We utilize the norms from (Brysbaert et al. 2014) which provide concreteness ratings for 40,000 English lexical items based on human subjects. Using these norms, we assign a concreteness score to each sentence in the training dataset and develop two curriculum strategies that progressively introduce more complex and abstract language patterns in the training data. Compared to the baselines, our best model shows lower performance on zero-shot tasks but demonstrates superior performance in fine-tuning tasks. Notably, our curriculum-trained models exhibit significant improvements over a non-curriculum based training of the same model.