Leonardo Ranaldi

Leonardo Ranaldi, Giulia Pucci, Andre Freitas

The language ability of Large Language Models (LLMs) is often unbalanced towards English because of the imbalance in the distribution of the pre-training data. This disparity is demanded in further fine-tuning and affecting the cross-lingual abilities of LLMs. In this paper, we propose to empower Instructiontuned LLMs (It-LLMs) in languages other than English by building semantic alignment between them. Hence, we propose CrossAlpaca, an It-LLM with cross-lingual instruction-following and Translation-following demonstrations to improve semantic alignment between languages. We validate our approach on the multilingual Question Answering (QA) benchmarks XQUAD and MLQA and adapted versions of MMLU and BBH. Our models, tested over six different languages, outperform the It-LLMs tuned on monolingual data. The final results show that instruction tuning on non-English data is not enough and that semantic alignment can be further improved by Translation-following demonstrations.

Leonardo Ranaldi, Giulia Pucci

Large Language Models have been demonstrating broadly satisfactory generative abilities for users, which seems to be due to the intensive use of human feedback that refines responses. Nevertheless, suggestibility inherited via human feedback improves the inclination to produce answers corresponding to users' viewpoints. This behaviour is known as sycophancy and depicts the tendency of LLMs to generate misleading responses as long as they align with humans. This phenomenon induces bias and reduces the robustness and, consequently, the reliability of these models. In this paper, we study the suggestibility of Large Language Models (LLMs) to sycophantic behaviour, analysing these tendencies via systematic human-interventions prompts over different tasks. Our investigation demonstrates that LLMs have sycophantic tendencies when answering queries that involve subjective opinions and statements that should elicit a contrary response based on facts. In contrast, when faced with math tasks or queries with an objective answer, they, at various scales, do not follow the users' hints by demonstrating confidence in generating the correct answers.

Leonardo Ranaldi, Giulia Pucci, Federico Ranaldi et al.

Reasoning methods, best exemplified by the well-known Chain-of-Thought (CoT), empower the reasoning abilities of Large Language Models (LLMs) by eliciting them to solve complex tasks in a step-by-step manner. Although they are achieving significant success, the ability to deliver multi-step reasoning remains limited to English because of the imbalance in the distribution of pre-training data, which makes other languages a barrier. In this paper, we propose Cross-lingual Tree-of-Thoughts (Cross-ToT), a method for aligning Cross-lingual CoT reasoning across languages. The proposed method, through a self-consistent cross-lingual prompting mechanism inspired by the Tree-of-Thoughts approach, provides multi-step reasoning paths in different languages that, during the steps, lead to the final solution. Experimental evaluations show that our method significantly outperforms existing prompting methods by reducing the number of interactions and achieving state-of-the-art performance.

Malvina Nissim, Danilo Croce, Viviana Patti et al.

The rapid progress of Large Language Models (LLMs) has transformed natural language processing and broadened its impact across research and society. Yet, systematic evaluation of these models, especially for languages beyond English, remains limited. "Challenging the Abilities of LAnguage Models in ITAlian" (CALAMITA) is a large-scale collaborative benchmarking initiative for Italian, coordinated under the Italian Association for Computational Linguistics. Unlike existing efforts that focus on leaderboards, CALAMITA foregrounds methodology: it federates more than 80 contributors from academia, industry, and the public sector to design, document, and evaluate a diverse collection of tasks, covering linguistic competence, commonsense reasoning, factual consistency, fairness, summarization, translation, and code generation. Through this process, we not only assembled a benchmark of over 20 tasks and almost 100 subtasks, but also established a centralized evaluation pipeline that supports heterogeneous datasets and metrics. We report results for four open-weight LLMs, highlighting systematic strengths and weaknesses across abilities, as well as challenges in task-specific evaluation. Beyond quantitative results, CALAMITA exposes methodological lessons: the necessity of fine-grained, task-representative metrics, the importance of harmonized pipelines, and the benefits and limitations of broad community engagement. CALAMITA is conceived as a rolling benchmark, enabling continuous integration of new tasks and models. This makes it both a resource -- the most comprehensive and diverse benchmark for Italian to date -- and a framework for sustainable, community-driven evaluation. We argue that this combination offers a blueprint for other languages and communities seeking inclusive and rigorous LLM evaluation practices.

Leonardo Ranaldi, Fabio Massimo Zanzotto

Instruction-tuned Large Language Models (It-LLMs) have been exhibiting outstanding abilities to reason around cognitive states, intentions, and reactions of all people involved, letting humans guide and comprehend day-to-day social interactions effectively. In fact, several multiple-choice questions (MCQ) benchmarks have been proposed to construct solid assessments of the models' abilities. However, earlier works are demonstrating the presence of inherent "order bias" in It-LLMs, posing challenges to the appropriate evaluation. In this paper, we investigate It-LLMs' resilience abilities towards a series of probing tests using four MCQ benchmarks. Introducing adversarial examples, we show a significant performance gap, mainly when varying the order of the choices, which reveals a selection bias and brings into discussion reasoning abilities. Following a correlation between first positions and model choices due to positional bias, we hypothesized the presence of structural heuristics in the decision-making process of the It-LLMs, strengthened by including significant examples in few-shot scenarios. Finally, by using the Chain-of-Thought (CoT) technique, we elicit the model to reason and mitigate the bias by obtaining more robust models.

Gabriele Maraia, Marco Valentino, Fabio Massimo Zanzotto et al.

Large Language Models (LLMs) often struggle with deductive judgment in syllogistic reasoning, systematically conflating semantic plausibility with formal validity a phenomenon known as content effect. This bias persists even when models generate step-wise explanations, indicating that intermediate rationales may inherit the same semantic shortcuts that affect answers. Recent approaches propose mitigating this issue by increasing inference-time structural constraints, either by encouraging abstract intermediate representations or by intervening directly in the model's internal computations; however, reliably suppressing semantic interference remains an open challenge. To make formal deduction less sensitive to semantic content, we introduce a framework for abstraction-guided reasoning that explicitly separates structural inference from lexical semantics. We construct paired content-laden and abstract syllogisms and use the model's activations on abstract inputs to define an abstract reasoning space. We then learn lightweight Abstractors that, from content-conditioned residual-stream states, predict representations aligned with this space and integrate these predictions via multi-layer interventions during the forward pass. Using cross-lingual transfer as a test bed, we show that abstraction-aligned steering reduces content-driven errors and improves validity-sensitive performance. Our results position activation-level abstraction as a scalable mechanism for enhancing the robustness of formal reasoning in LLMs against semantic interference.

Leonardo Ranaldi, Giulia Pucci, Fabio Massimo Zanzotto

The cognitive essence of humans is deeply intertwined with the concept of animacy, which plays an essential role in shaping their memory, vision, and multi-layered language understanding. Although animacy appears in language via nuanced constraints on verbs and adjectives, it is also learned and refined through extralinguistic information. Similarly, we assume that the LLMs' limited abilities to understand natural language when processing animacy are motivated by the fact that these models are trained exclusively on text. Hence, the question this paper aims to answer arises: can LLMs, in their digital wisdom, process animacy in a similar way to what humans would do? We then propose a systematic analysis via prompting approaches. In particular, we probe different LLMs by prompting them using animate, inanimate, usual, and stranger contexts. Results reveal that, although LLMs have been trained predominantly on textual data, they exhibit human-like behavior when faced with typical animate and inanimate entities in alignment with earlier studies. Hence, LLMs can adapt to understand unconventional situations by recognizing oddities as animated without needing to interface with unspoken cognitive triggers humans rely on to break down animations.

Leonardo Ranaldi, Andrè Freitas

The alignments of reasoning abilities between smaller and larger Language Models are largely conducted via Supervised Fine-Tuning (SFT) using demonstrations generated from robust Large Language Models (LLMs). Although these approaches deliver more performant models, they do not show sufficiently strong generalization ability as the training only relies on the provided demonstrations. In this paper, we propose the Self-refine Instruction-tuning method that elicits Smaller Language Models to self-refine their abilities. Our approach is based on a two-stage process, where reasoning abilities are first transferred between LLMs and Small Language Models (SLMs) via Instruction-tuning on demonstrations provided by LLMs, and then the instructed models Self-refine their abilities through preference optimization strategies. In particular, the second phase operates refinement heuristics based on the Direct Preference Optimization algorithm, where the SLMs are elicited to deliver a series of reasoning paths by automatically sampling the generated responses and providing rewards using ground truths from the LLMs. Results obtained on commonsense and math reasoning tasks show that this approach significantly outperforms Instruction-tuning in both in-domain and out-domain scenarios, aligning the reasoning abilities of Smaller and Larger Language Models.

Federico Ranaldi, Elena Sofia Ruzzetti, Dario Onorati et al.

Understanding textual description to generate code seems to be an achieved capability of instruction-following Large Language Models (LLMs) in zero-shot scenario. However, there is a severe possibility that this translation ability may be influenced by having seen target textual descriptions and the related code. This effect is known as Data Contamination. In this study, we investigate the impact of Data Contamination on the performance of GPT-3.5 in the Text-to-SQL code-generating tasks. Hence, we introduce a novel method to detect Data Contamination in GPTs and examine GPT-3.5's Text-to-SQL performances using the known Spider Dataset and our new unfamiliar dataset Termite. Furthermore, we analyze GPT-3.5's efficacy on databases with modified information via an adversarial table disconnection (ATD) approach, complicating Text-to-SQL tasks by removing structural pieces of information from the database. Our results indicate a significant performance drop in GPT-3.5 on the unfamiliar Termite dataset, even with ATD modifications, highlighting the effect of Data Contamination on LLMs in Text-to-SQL translation tasks.

Leonardo Ranaldi, Marco Valentino, Andrè Freitas

Chain-of-Though (CoT) represents a common strategy for reasoning in Large Language Models (LLMs) by decomposing complex tasks into intermediate inference steps. However, explanations generated via CoT are susceptible to content biases that negatively affect their robustness and faithfulness. To mitigate existing limitations, recent work has proposed using logical formalisms coupled with external symbolic solvers. However, fully symbolic approaches possess the bottleneck of requiring a complete translation from natural language to formal languages, a process that affects efficiency and flexibility. To achieve a trade-off, this paper investigates methods to disentangle content from logical reasoning without a complete formalisation. In particular, we present QuaSAR (for Quasi-Symbolic Abstract Reasoning), a variation of CoT that guides LLMs to operate at a higher level of abstraction via quasi-symbolic explanations. Our framework leverages the capability of LLMs to formalise only relevant variables and predicates, enabling the coexistence of symbolic elements with natural language. We show the impact of QuaSAR for in-context learning and for constructing demonstrations to improve the reasoning capabilities of smaller models. Our experiments show that quasi-symbolic abstractions can improve CoT-based methods by up to 8% accuracy, enhancing robustness and consistency on challenging adversarial variations on both natural language (i.e. MMLU-Redux) and symbolic reasoning tasks (i.e., GSM-Symbolic).

Leonardo Ranaldi, Barry Haddow, Alexandra Birch

Retrieval-augmented generation (RAG) has become a cornerstone of contemporary NLP, enhancing large language models (LLMs) by allowing them to access richer factual contexts through in-context retrieval. While effective in monolingual settings, especially in English, its use in multilingual tasks remains unexplored. This paper investigates the effectiveness of RAG across multiple languages by proposing novel approaches for multilingual open-domain question-answering. We evaluate the performance of various multilingual RAG strategies, including question-translation (tRAG), which translates questions into English before retrieval, and Multilingual RAG (MultiRAG), where retrieval occurs directly across multiple languages. Our findings reveal that tRAG, while useful, suffers from limited coverage. In contrast, MultiRAG improves efficiency by enabling multilingual retrieval but introduces inconsistencies due to cross-lingual variations in the retrieved content. To address these issues, we propose Crosslingual RAG (CrossRAG), a method that translates retrieved documents into a common language (e.g., English) before generating the response. Our experiments show that CrossRAG significantly enhances performance on knowledge-intensive tasks, benefiting both high-resource and low-resource languages.

Leonardo Ranaldi, Marco Valentino, Andrè Freitas

Retrieval-augmented generation (RAG) has emerged as a critical mechanism in contemporary NLP to support Large Language Models(LLMs) in systematically accessing richer factual context. However, the integration of RAG mechanisms brings its inherent challenges, as LLMs need to deal with potentially noisy contexts. Recent studies have shown that LLMs still struggle to critically analyse RAG-based in-context information, a limitation that may lead to incorrect inferences and hallucinations. In this paper, we investigate how to elicit critical reasoning in RAG via contrastive explanations. In particular, we propose Contrastive-RAG (C-RAG), a framework that (i) retrieves relevant documents given a query, (ii) selects and exemplifies relevant passages, and (iii) generates explanations that explicitly contrast the relevance of the passages to (iv) support the final answer. We show the impact of C-RAG building contrastive reasoning demonstrations from LLMs to instruct smaller models for retrieval-augmented tasks. Extensive experiments demonstrate that C-RAG improves state-of-the-art RAG models while (a) requiring significantly fewer prompts and demonstrations and (b) being robust to perturbations in the retrieved documents.

Federico Ranaldi, Andrea Zugarini, Leonardo Ranaldi et al.

We introduce the concept of protoknowledge to formalize and measure how sequences of tokens encoding Knowledge Graphs are internalized during pretraining and utilized at inference time by Large Language Models (LLMs). Indeed, LLMs have demonstrated the ability to memorize vast amounts of token sequences during pretraining, and a central open question is how they leverage this memorization as reusable knowledge through generalization. We then categorize protoknowledge into lexical, hierarchical, and topological forms, varying on the type of knowledge that needs to be activated. We measure protoknowledge through Knowledge Activation Tasks (KATs), analyzing its general properties such as semantic bias. We then investigate the impact of protoknowledge on Text-to-SPARQL performance by varying prompting strategies depending on input conditions. To this end, we adopt a novel analysis framework that assesses whether model predictions align with the successful activation of the relevant protoknowledge for each query. This methodology provides a practical tool to explore Semantic-Level Data Contamination and serves as an effective strategy for Closed-Pretraining models.

Leonardo Ranaldi, Federico Ranaldi, Fabio Massimo Zanzotto et al.

Retrieval-augmented generation (RAG) is key to enhancing large language models (LLMs) to systematically access richer factual knowledge. Yet, using RAG brings intrinsic challenges, as LLMs must deal with potentially conflicting knowledge, especially in multilingual retrieval, where the heterogeneity of knowledge retrieved may deliver different outlooks. To make RAG more analytical, critical and grounded, we introduce Dialectic-RAG (DRAG), a modular approach guided by Argumentative Explanations, i.e., structured reasoning process that systematically evaluates retrieved information by comparing, contrasting, and resolving conflicting perspectives. Given a query and a set of multilingual related documents, DRAG selects and exemplifies relevant knowledge for delivering dialectic explanations that, by critically weighing opposing arguments and filtering extraneous content, clearly determine the final response. Through a series of in-depth experiments, we show the impact of our framework both as an in-context learning strategy and for constructing demonstrations to instruct smaller models. The final results demonstrate that DRAG significantly improves RAG approaches, requiring low-impact computational effort and providing robustness to knowledge perturbations.

Mael Jullien, Marco Valentino, Leonardo Ranaldi et al.

Recent works on large language models (LLMs) have demonstrated the impact of prompting strategies and fine-tuning techniques on their reasoning capabilities. Yet, their effectiveness on clinical natural language inference (NLI) remains underexplored. This study presents the first controlled evaluation of how prompt structure and efficient fine-tuning jointly shape model performance in clinical NLI. We inspect four classes of prompting strategies to elicit reasoning in LLMs at different levels of abstraction, and evaluate their impact on a range of clinically motivated reasoning types. For each prompting strategy, we construct high-quality demonstrations using a frontier model to distil multi-step reasoning capabilities into smaller models (4B parameters) via Low-Rank Adaptation (LoRA). Across different language models fine-tuned on the NLI4CT benchmark, we found that prompt type alone accounts for up to 44% of the variance in macro-F1. Moreover, LoRA fine-tuning yields consistent gains of +8 to 12 F1, raises output alignment above 97%, and narrows the performance gap to GPT-4o-mini to within 7.1%. Additional experiments on reasoning generalisation reveal that LoRA improves performance in 75% of the models on MedNLI and TREC Clinical Trials Track. Overall, these findings demonstrate that (i) prompt structure is a primary driver of clinical reasoning performance, (ii) compact models equipped with strong prompts and LoRA can rival frontier-scale systems, and (iii) reasoning-type-aware evaluation is essential to uncover prompt-induced trade-offs. Our results highlight the promise of combining prompt design and lightweight adaptation for more efficient and trustworthy clinical NLP systems, providing insights on the strengths and limitations of widely adopted prompting and parameter-efficient techniques in highly specialised domains.

Fabio Massimo Zanzotto, Elena Sofia Ruzzetti, Giancarlo A. Xompero et al.

Memorization is a fundamental ability of Transformer-based Large Language Models, achieved through learning. In this paper, we propose a paradigm shift by designing an architecture to memorize text directly, bearing in mind the principle that memorization precedes learning. We introduce MeMo, a novel architecture for language modeling that explicitly memorizes sequences of tokens in layered associative memories. By design, MeMo offers transparency and the possibility of model editing, including forgetting texts. We experimented with the MeMo architecture, showing the memorization power of the one-layer and the multi-layer configurations.

Leonardo Ranaldi, Elena Sofia Ruzzetti, Davide Venditti et al.

Cheap-to-Build Very Large-Language Models (CtB-LLMs) with affordable training are emerging as the next big revolution in natural language processing and understanding. These CtB-LLMs are democratizing access to trainable Very Large-Language Models (VLLMs) and, thus, may represent the building blocks of many NLP systems solving downstream tasks. Hence, a little or a large bias in CtB-LLMs may cause huge harm. In this paper, we performed a large investigation of the bias of three families of CtB-LLMs, and we showed that debiasing techniques are effective and usable. Indeed, according to current tests, the LLaMA and the OPT families have an important bias in gender, race, religion, and profession. In contrast to the analysis for other LLMs, we discovered that bias depends not on the number of parameters but on the perplexity. Finally, the debiasing of OPT using LoRA reduces bias up to 4.12 points in the normalized stereotype score.

Leonardo Ranaldi, Elena Sofia Ruzzetti, Fabio Massimo Zanzotto

Pre-trained Language Models such as BERT are impressive machines with the ability to memorize, possibly generalized learning examples. We present here a small, focused contribution to the analysis of the interplay between memorization and performance of BERT in downstream tasks. We propose PreCog, a measure for evaluating memorization from pre-training, and we analyze its correlation with the BERT's performance. Our experiments show that highly memorized examples are better classified, suggesting memorization is an essential key to success for BERT.

Elena Sofia Ruzzetti, Federico Ranaldi, Felicia Logozzo et al.

The impressive achievements of transformers force NLP researchers to delve into how these models represent the underlying structure of natural language. In this paper, we propose a novel standpoint to investigate the above issue: using typological similarities among languages to observe how their respective monolingual models encode structural information. We aim to layer-wise compare transformers for typologically similar languages to observe whether these similarities emerge for particular layers. For this investigation, we propose to use Centered Kernel Alignment to measure similarity among weight matrices. We found that syntactic typological similarity is consistent with the similarity between the weights in the middle layers, which are the pretrained BERT layers to which syntax encoding is generally attributed. Moreover, we observe that a domain adaptation on semantically equivalent texts enhances this similarity among weight matrices.

Leonardo Ranaldi, Aria Nourbakhsh, Arianna Patrizi et al.

Pre-trained Transformers are challenging human performances in many NLP tasks. The massive datasets used for pre-training seem to be the key to their success on existing tasks. In this paper, we explore how a range of pre-trained Natural Language Understanding models perform on definitely unseen sentences provided by classification tasks over a DarkNet corpus. Surprisingly, results show that syntactic and lexical neural networks perform on par with pre-trained Transformers even after fine-tuning. Only after what we call extreme domain adaptation, that is, retraining with the masked language model task on all the novel corpus, pre-trained Transformers reach their standard high results. This suggests that huge pre-training corpora may give Transformers unexpected help since they are exposed to many of the possible sentences.

Elena Sofia Ruzzetti, Leonardo Ranaldi, Michele Mastromattei et al.

Word embeddings are powerful dictionaries, which may easily capture language variations. However, these dictionaries fail to give sense to rare words, which are surprisingly often covered by traditional dictionaries. In this paper, we propose to use definitions retrieved in traditional dictionaries to produce word embeddings for rare words. For this purpose, we introduce two methods: Definition Neural Network (DefiNNet) and Define BERT (DefBERT). In our experiments, DefiNNet and DefBERT significantly outperform state-of-the-art as well as baseline methods devised for producing embeddings of unknown words. In fact, DefiNNet significantly outperforms FastText, which implements a method for the same task-based on n-grams, and DefBERT significantly outperforms the BERT method for OOV words. Then, definitions in traditional dictionaries are useful to build word embeddings for rare words.