Raffaella Bernardi

Leonardo Bertolazzi, Katya Tentori, Raffaella Bernardi

Large language models (LLMs) are increasingly deployed as autonomous agents in scientific tasks. Yet whether these systems can effectively engage in forms of inductive reasoning relevant to scientific discovery remains an open question. In this work, we introduce FALSIFYBENCH, an evaluation framework for hypothesis-driven reasoning inspired by the classic Wason 2-4-6 task, in which agents must discover hidden semantic properties by iteratively proposing examples and receiving feedback. This task captures key elements of scientific reasoning: hypothesis generation, evidence gathering, and belief revision in response to both confirming and disconfirming evidence. Our evaluation of 12 LLMs across model families and scales shows that reasoning models are generally stronger scientific reasoners than instruction-tuned models, although no model comes close to optimal performance. The primary driver of success is the capacity for negative testing: models that actively seek to falsify their hypotheses consistently outperform those that primarily seek confirmation. Moreover, a fine-grained turn-level analysis, neglected in previous work, reveals that failure is tied to identifiable patterns in how models navigate the hypothesis space.

Davide Mazzaccara, Alberto Testoni, Raffaella Bernardi

Questions are essential tools for acquiring the necessary information to complete information-seeking tasks. However, large language models (LLMs), especially open-source models, often perform poorly in generating informative questions, as measured by expected information gain (EIG). In this paper, we propose a method to enhance the informativeness of LLM-generated questions in 20-question game dialogues. We sample multiple questions from the same model (LLAMA 2-CHAT 7B) for each game and create pairs of low-EIG and high-EIG questions to apply a Direct Preference Optimization (DPO) algorithm. Our results show that this method produces more effective questions (in terms of EIG), even in domains different from those used to train the DPO model.

Leonardo Bertolazzi, Philipp Mondorf, Barbara Plank et al.

The ability of large language models (LLMs) to validate their output and identify potential errors is crucial for ensuring robustness and reliability. However, current research indicates that LLMs struggle with self-correction, encountering significant challenges in detecting errors. While studies have explored methods to enhance self-correction in LLMs, relatively little attention has been given to understanding the models' internal mechanisms underlying error detection. In this paper, we present a mechanistic analysis of error detection in LLMs, focusing on simple arithmetic problems. Through circuit analysis, we identify the computational subgraphs responsible for detecting arithmetic errors across four smaller-sized LLMs. Our findings reveal that all models heavily rely on $\textit{consistency heads}$--attention heads that assess surface-level alignment of numerical values in arithmetic solutions. Moreover, we observe that the models' internal arithmetic computation primarily occurs in higher layers, whereas validation takes place in middle layers, before the final arithmetic results are fully encoded. This structural dissociation between arithmetic computation and validation seems to explain why smaller-sized LLMs struggle to detect even simple arithmetic errors.

Filippo Momentè, Alessandro Suglia, Mario Giulianelli et al.

We examine three evaluation paradigms: standard benchmarks (e.g., MMLU and BBH), interactive games (e.g., Signalling Games or Taboo), and cognitive tests (e.g., for working memory or theory of mind). First, we investigate which of the former two-benchmarks or games-is most effective at discriminating LLMs of varying quality. Then, inspired by human cognitive assessments, we compile a suite of targeted tests that measure cognitive abilities deemed essential for effective language use, and we investigate their correlation with model performance in benchmarks and games. Our analyses reveal that interactive games are superior to standard benchmarks in discriminating models. Causal and logical reasoning correlate with both static and interactive tests, while differences emerge regarding core executive functions and social/emotional skills, which correlate more with games. We advocate for the development of new interactive benchmarks and targeted cognitive tasks inspired by assessing human abilities but designed specifically for LLMs.

Nicola Horst, Davide Mazzaccara, Antonia Schmidt et al.

Interaction between learner and feedback-giver has come into focus recently for post-training of Large Language Models (LLMs), through the use of reward models that judge the appropriateness of a model's response. In this paper, we investigate whether Dialogue Games -- goal-directed and rule-governed activities driven predominantly by verbal actions -- can also serve as a source of feedback signals for learning. We introduce Playpen, an environment for off- and online learning through Dialogue Game self-play, and investigate a representative set of post-training methods: supervised fine-tuning; direct alignment (DPO); and reinforcement learning with GRPO. We experiment with post-training a small LLM (Llama-3.1-8B-Instruct), evaluating performance on unseen instances of training games as well as unseen games, and on standard benchmarks. We find that imitation learning through SFT improves performance on unseen instances, but negatively impacts other skills, while interactive learning with GRPO shows balanced improvements without loss of skills. We release the framework and the baseline training setups to foster research in the promising new direction of learning in (synthetic) interaction.

Sherzod Hakimov, Roland Bernard, Tim Leiber et al.

Negotiation is a fundamental challenge for AI agents, as it requires an ability to reason strategically, model opponents, and balance cooperation with competition. We conduct the first comprehensive study systematically evaluating the effect of (LLM-)reasoning on the negotiation abilities of both commercial and open-weight LLMs, and do this across three languages. Using a self-play setup across three diverse dialogue games, we analyse trade-offs between performance and cost, the language consistency of reasoning processes, and the nature of strategic adaptation exhibited by models. Our findings show that enabling reasoning-that is, scaling test time compute-significantly improves negotiation outcomes by enhancing collaboration and helping models overcome task complexities, but comes at a substantial computational cost: reasoning improves GPT-5's performance by 31.4 % while increasing its cost by nearly 400 %. Most critically, we uncover a significant multilingual reasoning distinction: open-weight models consistently switch to English for their internal reasoning steps, even when negotiating in German or Italian (and thus possibly impacting potential explainability gains through the disclosure of reasoning traces), while leading commercial models maintain language consistency between their reasoning and final output.

Leonardo Bertolazzi, Manuel Vargas Guzmán, Raffaella Bernardi et al.

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.

Davide Testa, Giovanni Bonetta, Raffaella Bernardi et al.

We introduce MAIA (Multimodal AI Assessment), a native-Italian benchmark designed for fine-grained investigation of the reasoning abilities of visual language models on videos. MAIA differs from other available video benchmarks for its design, its reasoning categories, the metric it uses, and the language and culture of the videos. MAIA evaluates Vision Language Models (VLMs) on two aligned tasks: a visual statement verification task, and an open-ended visual question-answering task, both on the same set of video-related questions. It considers twelve reasoning categories that aim to disentangle language and vision relations by highlighting the role of the visual input. Thanks to its carefully taught design, it evaluates VLMs' consistency and visually grounded natural language comprehension and generation simultaneously through an aggregated metric revealing low results that highlight models' fragility. Last but not least, the video collection has been carefully selected to reflect the Italian culture, and the language data are produced by native-speakers.

Anna Bavaresco, Raffaella Bernardi, Leonardo Bertolazzi et al.

There is an increasing trend towards evaluating NLP models with LLMs instead of human judgments, raising questions about the validity of these evaluations, as well as their reproducibility in the case of proprietary models. We provide JUDGE-BENCH, an extensible collection of 20 NLP datasets with human annotations covering a broad range of evaluated properties and types of data, and comprehensively evaluate 11 current LLMs, covering both open-weight and proprietary models, for their ability to replicate the annotations. Our evaluations show substantial variance across models and datasets. Models are reliable evaluators on some tasks, but overall display substantial variability depending on the property being evaluated, the expertise level of the human judges, and whether the language is human or model-generated. We conclude that LLMs should be carefully validated against human judgments before being used as evaluators.

Leonardo Bertolazzi, Albert Gatt, Raffaella Bernardi

The reasoning abilities of Large Language Models (LLMs) are becoming a central focus of study in NLP. In this paper, we consider the case of syllogistic reasoning, an area of deductive reasoning studied extensively in logic and cognitive psychology. Previous research has shown that pre-trained LLMs exhibit reasoning biases, such as $\textit{content effects}$, avoid answering that $\textit{no conclusion follows}$, display human-like difficulties, and struggle with multi-step reasoning. We contribute to this research line by systematically investigating the effects of chain-of-thought reasoning, in-context learning (ICL), and supervised fine-tuning (SFT) on syllogistic reasoning, considering syllogisms with conclusions that support or violate world knowledge, as well as ones with multiple premises. Crucially, we go beyond the standard focus on accuracy, with an in-depth analysis of the conclusions generated by the models. Our results suggest that the behavior of pre-trained LLMs can be explained by heuristics studied in cognitive science and that both ICL and SFT improve model performance on valid inferences, although only the latter mitigates most reasoning biases without harming model consistency.

Alberto Testoni, Raffaella Bernardi

Generating goal-oriented questions in Visual Dialogue tasks is a challenging and long-standing problem. State-Of-The-Art systems are shown to generate questions that, although grammatically correct, often lack an effective strategy and sound unnatural to humans. Inspired by the cognitive literature on information search and cross-situational word learning, we design Confirm-it, a model based on a beam search re-ranking algorithm that guides an effective goal-oriented strategy by asking questions that confirm the model's conjecture about the referent. We take the GuessWhat?! game as a case-study. We show that dialogues generated by Confirm-it are more natural and effective than beam search decoding without re-ranking.

Alberto Testoni, Raffaella Bernardi

When training a model on referential dialogue guessing games, the best model is usually chosen based on its task success. We show that in the popular end-to-end approach, this choice prevents the model from learning to generate linguistically richer dialogues, since the acquisition of language proficiency takes longer than learning the guessing task. By comparing models playing different games (GuessWhat, GuessWhich, and Mutual Friends), we show that this discrepancy is model- and task-agnostic. We investigate whether and when better language quality could lead to higher task success. We show that in GuessWhat, models could increase their accuracy if they learn to ground, encode, and decode also words that do not occur frequently in the training set.

Alberto Testoni, Raffaella Bernardi

Despite important progress, conversational systems often generate dialogues that sound unnatural to humans. We conjecture that the reason lies in their different training and testing conditions: agents are trained in a controlled "lab" setting but tested in the "wild". During training, they learn to generate an utterance given the human dialogue history. On the other hand, during testing, they must interact with each other, and hence deal with noisy data. We propose to fill this gap by training the model with mixed batches containing both samples of human and machine-generated dialogues. We assess the validity of the proposed method on GuessWhat?!, a visual referential game.

Claudio Greco, Barbara Plank, Raquel Fernández et al.

We study the issue of catastrophic forgetting in the context of neural multimodal approaches to Visual Question Answering (VQA). Motivated by evidence from psycholinguistics, we devise a set of linguistically-informed VQA tasks, which differ by the types of questions involved (Wh-questions and polar questions). We test what impact task difficulty has on continual learning, and whether the order in which a child acquires question types facilitates computational models. Our results show that dramatic forgetting is at play and that task difficulty and order matter. Two well-known current continual learning methods mitigate the problem only to a limiting degree.

Ravi Shekhar, Ece Takmaz, Raquel Fernández et al.

The multimodal models used in the emerging field at the intersection of computational linguistics and computer vision implement the bottom-up processing of the `Hub and Spoke' architecture proposed in cognitive science to represent how the brain processes and combines multi-sensory inputs. In particular, the Hub is implemented as a neural network encoder. We investigate the effect on this encoder of various vision-and-language tasks proposed in the literature: visual question answering, visual reference resolution, and visually grounded dialogue. To measure the quality of the representations learned by the encoder, we use two kinds of analyses. First, we evaluate the encoder pre-trained on the different vision-and-language tasks on an existing diagnostic task designed to assess multimodal semantic understanding. Second, we carry out a battery of analyses aimed at studying how the encoder merges and exploits the two modalities.

Ravi Shekhar, Aashish Venkatesh, Tim Baumgärtner et al.

We propose a grounded dialogue state encoder which addresses a foundational issue on how to integrate visual grounding with dialogue system components. As a test-bed, we focus on the GuessWhat?! game, a two-player game where the goal is to identify an object in a complex visual scene by asking a sequence of yes/no questions. Our visually-grounded encoder leverages synergies between guessing and asking questions, as it is trained jointly using multi-task learning. We further enrich our model via a cooperative learning regime. We show that the introduction of both the joint architecture and cooperative learning lead to accuracy improvements over the baseline system. We compare our approach to an alternative system which extends the baseline with reinforcement learning. Our in-depth analysis shows that the linguistic skills of the two models differ dramatically, despite approaching comparable performance levels. This points at the importance of analyzing the linguistic output of competing systems beyond numeric comparison solely based on task success.

Hoa Trong Vu, Claudio Greco, Aliia Erofeeva et al.

Capturing semantic relations between sentences, such as entailment, is a long-standing challenge for computational semantics. Logic-based models analyse entailment in terms of possible worlds (interpretations, or situations) where a premise P entails a hypothesis H iff in all worlds where P is true, H is also true. Statistical models view this relationship probabilistically, addressing it in terms of whether a human would likely infer H from P. In this paper, we wish to bridge these two perspectives, by arguing for a visually-grounded version of the Textual Entailment task. Specifically, we ask whether models can perform better if, in addition to P and H, there is also an image (corresponding to the relevant "world" or "situation"). We use a multimodal version of the SNLI dataset (Bowman et al., 2015) and we compare "blind" and visually-augmented models of textual entailment. We show that visual information is beneficial, but we also conduct an in-depth error analysis that reveals that current multimodal models are not performing "grounding" in an optimal fashion.

Ravi Shekhar, Tim Baumgartner, Aashish Venkatesh et al.

Our goal is to explore how the abilities brought in by a dialogue manager can be included in end-to-end visually grounded conversational agents. We make initial steps towards this general goal by augmenting a task-oriented visual dialogue model with a decision-making component that decides whether to ask a follow-up question to identify a target referent in an image, or to stop the conversation to make a guess. Our analyses show that adding a decision making component produces dialogues that are less repetitive and that include fewer unnecessary questions, thus potentially leading to more efficient and less unnatural interactions.

Sandro Pezzelle, Ionut-Teodor Sorodoc, Raffaella Bernardi

The present work investigates whether different quantification mechanisms (set comparison, vague quantification, and proportional estimation) can be jointly learned from visual scenes by a multi-task computational model. The motivation is that, in humans, these processes underlie the same cognitive, non-symbolic ability, which allows an automatic estimation and comparison of set magnitudes. We show that when information about lower-complexity tasks is available, the higher-level proportional task becomes more accurate than when performed in isolation. Moreover, the multi-task model is able to generalize to unseen combinations of target/non-target objects. Consistently with behavioral evidence showing the interference of absolute number in the proportional task, the multi-task model no longer works when asked to provide the number of target objects in the scene.

Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich et al.

In this paper, we aim to understand whether current language and vision (LaVi) models truly grasp the interaction between the two modalities. To this end, we propose an extension of the MSCOCO dataset, FOIL-COCO, which associates images with both correct and "foil" captions, that is, descriptions of the image that are highly similar to the original ones, but contain one single mistake ("foil word"). We show that current LaVi models fall into the traps of this data and perform badly on three tasks: a) caption classification (correct vs. foil); b) foil word detection; c) foil word correction. Humans, in contrast, have near-perfect performance on those tasks. We demonstrate that merely utilising language cues is not enough to model FOIL-COCO and that it challenges the state-of-the-art by requiring a fine-grained understanding of the relation between text and image.

Ionut Sorodoc, Sandro Pezzelle, Aurélie Herbelot et al.

Major advances have recently been made in merging language and vision representations. But most tasks considered so far have confined themselves to the processing of objects and lexicalised relations amongst objects (content words). We know, however, that humans (even pre-school children) can abstract over raw data to perform certain types of higher-level reasoning, expressed in natural language by function words. A case in point is given by their ability to learn quantifiers, i.e. expressions like 'few', 'some' and 'all'. From formal semantics and cognitive linguistics, we know that quantifiers are relations over sets which, as a simplification, we can see as proportions. For instance, in 'most fish are red', most encodes the proportion of fish which are red fish. In this paper, we study how well current language and vision strategies model such relations. We show that state-of-the-art attention mechanisms coupled with a traditional linguistic formalisation of quantifiers gives best performance on the task. Additionally, we provide insights on the role of 'gist' representations in quantification. A 'logical' strategy to tackle the task would be to first obtain a numerosity estimation for the two involved sets and then compare their cardinalities. We however argue that precisely identifying the composition of the sets is not only beyond current state-of-the-art models but perhaps even detrimental to a task that is most efficiently performed by refining the approximate numerosity estimator of the system.

Sandro Pezzelle, Marco Marelli, Raffaella Bernardi

People can refer to quantities in a visual scene by using either exact cardinals (e.g. one, two, three) or natural language quantifiers (e.g. few, most, all). In humans, these two processes underlie fairly different cognitive and neural mechanisms. Inspired by this evidence, the present study proposes two models for learning the objective meaning of cardinals and quantifiers from visual scenes containing multiple objects. We show that a model capitalizing on a 'fuzzy' measure of similarity is effective for learning quantifiers, whereas the learning of exact cardinals is better accomplished when information about number is provided.

Denis Paperno, Germán Kruszewski, Angeliki Lazaridou et al.

We introduce LAMBADA, a dataset to evaluate the capabilities of computational models for text understanding by means of a word prediction task. LAMBADA is a collection of narrative passages sharing the characteristic that human subjects are able to guess their last word if they are exposed to the whole passage, but not if they only see the last sentence preceding the target word. To succeed on LAMBADA, computational models cannot simply rely on local context, but must be able to keep track of information in the broader discourse. We show that LAMBADA exemplifies a wide range of linguistic phenomena, and that none of several state-of-the-art language models reaches accuracy above 1% on this novel benchmark. We thus propose LAMBADA as a challenging test set, meant to encourage the development of new models capable of genuine understanding of broad context in natural language text.

Raffaella Bernardi, Ruket Cakici, Desmond Elliott et al.

Automatic description generation from natural images is a challenging problem that has recently received a large amount of interest from the computer vision and natural language processing communities. In this survey, we classify the existing approaches based on how they conceptualize this problem, viz., models that cast description as either generation problem or as a retrieval problem over a visual or multimodal representational space. We provide a detailed review of existing models, highlighting their advantages and disadvantages. Moreover, we give an overview of the benchmark image datasets and the evaluation measures that have been developed to assess the quality of machine-generated image descriptions. Finally we extrapolate future directions in the area of automatic image description generation.

Angeliki Lazaridou, Dat Tien Nguyen, Raffaella Bernardi et al.

We introduce language-driven image generation, the task of generating an image visualizing the semantic contents of a word embedding, e.g., given the word embedding of grasshopper, we generate a natural image of a grasshopper. We implement a simple method based on two mapping functions. The first takes as input a word embedding (as produced, e.g., by the word2vec toolkit) and maps it onto a high-level visual space (e.g., the space defined by one of the top layers of a Convolutional Neural Network). The second function maps this abstract visual representation to pixel space, in order to generate the target image. Several user studies suggest that the current system produces images that capture general visual properties of the concepts encoded in the word embedding, such as color or typical environment, and are sufficient to discriminate between general categories of objects.