Publications

Understanding and modeling human intelligence are parallel goals shared by artificial intelligence (AI) and cognitive science. As AI systems grow increasingly capable, in what ways do model responses resemble human responses, and where do they systematically diverge? The sheer breadth and diversity of the tasks humans can perform and think about pose a challenge for scalable and rigorous comparison between humans and models. We introduce CogGym, a scalable, unified framework grounded in cognitive science for systematically comparing model and human behavior on matched experimental trials. CogGym uses a semi-automated, human-in-the-loop pipeline to standardize diverse experimental paradigms into a task-agnostic Experiment Markup Language (EML), enabling reproducible and faithful comparison at scale. For initial release, we curate and standardize 258 cognitive experiments from 100 papers that focuses on human commonsense reasoning, and evaluate 50 large language models against human responses. We find a clear scaling trend where larger and more recent AI models better reproduce human judgments. Yet AI models' improvement on such common reasoning tasks is considerably slower than the gains observed on formal-reasoning benchmarks like math and coding, and model–human fit remains well below human splithalf reliability (R² = 0.93 on text, 0.95 on image, and 0.92 on video) with the best models achieving R² = 0.59 on text, 0.58 on image, and 0.43 on video experiments. We intend for CogGym to provide a living evaluation framework that continually incorporates new cognitive science experiments to characterize where model behavior resembles human behavior, where it systematically diverges, and how those patterns change as models and experiments evolve.

In many forms of reasoning, including arithmetic reasoning, generalizing across superficial changes in input format is effortless for humans: anyone who can solve 2+5 can also solve 'two plus five'. In contrast, LLMs are more brittle to surface variations of the prompts: for example, they solve numeric arithmetic problems almost perfectly but are substantially less accurate on verbal renditions of the same problems. Here, we ask whether generalization across formats can be predicted from the models' internals. Using attribution patching, we first independently localize the circuit that each model recruits to solve numeric arithmetic problems (2+5) vs. verbal ones, in three languages: English ('two plus five'), Spanish ('dos más cinco'), and Italian ('due più cinque'); then, we test whether overlap with the model's own numeric circuit predicts its generalization to the verbal formats. Indeed, we find support for this idea at three levels: circuit overlap accounts for the relative difficulty of the three verbal formats, for which models generalize best, and for which items are solved correctly, rivaling supervised probes while requiring no labeled data.

For scientific progress, we need benchmarks that test the limits of state-of-the-art models, and evaluation methods that inform us about failure cases. As models get stronger, standard benchmarks for machine translation are approaching saturation. Further, automatic translation metrics are unreliable, opaque, and vulnerable to reward-hacking. Even gold human evaluation is not problem-free, because it often lacks reproducibility, objectivity, and scalability. Overall, this prevents us from tracking progress in the field and identifying pathways for improvement. We introduce the Last Translation Benchmark, a collection of human-authored and peer-reviewed examples (texts, images, audio, videos) that break leading machine translation models. We also present a new evaluation approach: each example comes with handcrafted verification rules describing concrete failure cases on that example, therefore allowing reliable and actionable future evaluation. The Last Translation Benchmark is a live dataset that accepts ongoing contributions. The latest version is LTBv1, containing accepted contributions prior to September 1st 2026, with future releases planned as new data is continuously collected.

Dialogue games represent a challenging setting where complex cognitive skills are required to accomplish tasks while coordinating with other players. Considering that language represents an interface for both understanding the game rules and executing actions, it is reasonable to assume that training on a specific language game will enhance specific capabilities that might be relevant for other tasks as well. Motivated by this rationale, in this paper, we investigate how knowledge transfers across different dialogue games. We study transferability by finetuning LLM models on games from the clembench suite (Chalamalasetti et al., 2023) and performing two analyses: i) we derive a task-transferability graph using a binary integer optimization program from Zamir et al. (2018), using task performance as the main metric; and ii) we compute task vectors (Ilharco et al., 2022) for each game to study similarities across finetuned models and their task transferability. In our first analysis, we find that some games benefit more from transfer than finetuning, and that the visuospatial family (e.g., exploration games) transfers best. With our task vector analysis instead, we find that similarity-based approaches capture game-role relationships but almost no transferability patterns, suggesting that more complex metrics are required.

Large Language Models (LLMs) such as ChatGPT are transforming how scientists conduct and validate research, offering promise as tools to improve scientific reproducibility. However, computational reproducibility and error detection remain expensive and labor-intensive. We experimentally test how collaboration between researchers and LLM assistants influences the reproduction of quantitative social science findings across different levels of AI autonomy. We randomly assigned 288 researchers to 103 teams working under three conditions: human-only, AI-assisted (using ChatGPT as a collaborative tool), or AI-led (ChatGPT operating with minimal human oversight). Teams reproduced published results from leading social science journals, detected coding errors, and proposed robustness checks. Human-only and AI-assisted teams achieved comparable reproduction rates (94% vs. 91%) and performed similarly on most outcomes, except human-only teams identified significantly more major coding errors. Both substantially outperformed AI-led teams, which achieved only a 37% reproduction rate, detected fewer errors across all categories, proposed weaker robustness checks, and required more time. This autonomous approach, however, likely represents only a lower bound of AI capabilities. Despite rapid model advances, expert human judgment currently remains indispensable for reliable empirical verification. While AI assistance did not degrade most outcomes, it provided no measurable advantages and was associated with reduced detection of major errors. However, the 37% autonomous reproduction rate indicates that AI could provide value in settings where scale or cost constraints preclude human review of papers, even though general-purpose LLMs offer no immediate advantages for human-supervised verification.

Iconicity, defined as the potential of linguistic signs to resemble properties or features of their referents, is increasingly recognized as a general property of language. One common approach for quantifying iconicity is to collect iconicity ratings. Although iconicity datasets have been developed for several languages, no comprehensive dataset of iconicity ratings is currently available for Italian. The current study presents IconicITA, the first dataset of Italian iconicity ratings for the 1,121 words of the Italian adaptation of Affective Norms for English Words (ANEW). Ratings were collected from both Italian native speakers (L1) and English native speakers with Italian as a second language (L2). Including L2 participants allowed us to contribute to the debate on whether iconicity ratings genuinely measure form-meaning resemblance, rather than exclusively reflecting semantic properties. We showed that L1 Italian iconicity ratings are positively associated with perceptual strength in the auditory and haptic modalities, and with specificity ratings. Conversely, we found a negative correlation between iconicity and concreteness, age of acquisition, word frequency, and letter frequency. In general, the relationship between Italian iconicity norms and various psycholinguistic variables largely replicated previous findings in the literature on iconicity. Considering L2 data, the ratings provided by L2 speakers correlated more strongly with the Italian L1 data compared to the translation-equivalent English L1 data. This finding suggests that participants' judgments were influenced not only by the semantic information of the words but also by language-specific form-level properties. We take this result as evidence of the validity of iconicity ratings to operationalize the degree of resemblance between words' form and meaning.

Auditory iconic words display a phonological profile that imitates their referents' sounds. Traditionally, those words are thought to constitute a minor portion of the auditory lexicon. In this article, we challenge this assumption by assessing the pervasiveness of onomatopoeia in the English auditory vocabulary through a novel data-driven procedure. We embed spoken words and natural sounds into a shared auditory space through (a) a short-time Fourier transform, (b) a convolutional neural network trained to classify sounds, and (c) a network trained on speech recognition. Then, we employ the obtained vector representations to measure their objective auditory resemblance. These similarity indexes show that imitation is not limited to some circumscribed semantic categories, but instead can be considered as a widespread mechanism underlying the structure of the English auditory vocabulary. We finally empirically validate our similarity indexes as measures of iconicity against human judgments.

Despite being largely spoken and studied by language and cognitive scientists, Italian lacks large resources of language processing data. The Italian Crowdsourcing Project (ICP) is a dataset of word recognition times and accuracy including responses to 130,465 words, which makes it the largest dataset of its kind item-wise. The data were collected in an online word knowledge task in which over 156,000 native speakers of Italian took part. We validated the ICP dataset by (1) showing that ICP reaction times correlate strongly (r = .78) with lexical decision latencies collected in a traditional lab experiment, (2) showing that the effect of major psycholinguistic variables (e.g., frequency, length, etc.) can be replicated in this dataset, and (3) replicating the effect of word prevalence, which we compute here for the first time for Italian. Given the inclusion of many inflectional forms of verbs, adjectives, and nouns, we further showcase the potential of this dataset by exploring two phenomena (inflectional entropy in verb paradigms and the clitic effect in isolated word recognition) that build on the peculiar properties of Italian. In this paper we present the ICP resource and release response times, accuracy, and prevalence estimates for all the words included.

Pseudowords such as "knackets" or "spechy"—letter strings that are consistent with the orthotactical rules of a language but do not appear in its lexicon—are traditionally considered to be meaningless, and used as such in empirical studies. However, recent studies that show specific semantic patterns associated with these words as well as semantic effects on human pseudoword processing have cast doubt on this view. While these studies suggest that pseudowords have meanings, they provide only extremely limited insight as to whether humans are able to ascribe explicit and declarative semantic content to unfamiliar word forms. In the present study, we utilized an exploratory-confirmatory study design to examine this question. In a first exploratory study, we started from a pre-existing dataset of words and pseudowords alongside human-generated definitions for these items. Using 18 different language models, we showed that the definitions actually produced for (pseudo)words were closer to their respective (pseudo)words than the definitions for the other items. Based on these initial results, we conducted a second, pre-registered, high-powered confirmatory study collecting a new, controlled set of (pseudo)word interpretations. This second study confirmed the results of the first one. Taken together, these findings support the idea that meaning construction is supported by a flexible form-to-meaning mapping system based on statistical regularities in the language environment that can accommodate novel lexical entries as soon as they are encountered.

We release a database of cloze probability values, predictability ratings, and computational estimates for a sample of 205 English sentences (1726 words), aligned with previously released word-by-word reading time data (both self-paced reading and eye-movement records; Frank et al., Behavior Research Methods, 45(4), 1182–1190. 2013) and EEG responses (Frank et al., Brain and Language, 140, 1–11. 2015). Our analyses show that predictability ratings are the best predictors of the EEG signal (N400, P600, LAN) self-paced reading times, and eye movement patterns, when spillover effects are taken into account. The computational estimates are particularly effective at explaining variance in the eye-tracking data without spillover. Cloze probability estimates have decent overall psychometric accuracy and are the best predictors of early fixation patterns (first fixation duration). Our results indicate that the choice of the best measurement of word predictability in context critically depends on the processing index being considered.

Massively multilingual models such as mBERT and XLM-R are increasingly valued in Natural Language Processing research and applications, due to their ability to tackle the uneven distribution of resources available for different languages. The models' ability to process multiple languages relying on a shared set of parameters raises the question of whether the grammatical knowledge they extracted during pre-training can be considered as a data-driven cross-lingual grammar. The present work studies the inner workings of mBERT and XLM-R in order to test the cross-lingual consistency of the individual neural units that respond to a precise syntactic phenomenon, that is, number agreement, in five languages (English, German, French, Hebrew, Russian). We found that there is a significant overlap in the latent dimensions that encode agreement across the languages we considered. This overlap is larger (a) for long- vis-à-vis short-distance agreement and (b) when considering XLM-R as compared to mBERT, and peaks in the intermediate layers of the network. We further show that a small set of syntax-sensitive neurons can capture agreement violations across languages; however, their contribution is not decisive in agreement processing.

Large language models (LLMs) can reliably distinguish grammatical from ungrammatical sentences, but how grammatical knowledge is represented within the model remains an open question. We investigate whether different syntactic phenomena recruit shared or distinct components in LLMs. Using a functional localization approach inspired by cognitive neuroscience, we identify the LLM units most responsive to 67 English syntactic phenomena in seven open-weight models. These units are consistently recruited across sentence instances and causally support model performance. Critically, different types of syntactic agreement (e.g., subject-verb, anaphor, determiner-noun) recruit overlapping sets of units, suggesting that agreement constitutes a meaningful functional category in LLMs. This pattern holds in Russian and Chinese, and further, across languages: in a cross-lingual analysis of 57 languages, syntactically similar languages share more units for subject-verb agreement. Taken together, these findings reveal that syntactic agreement—a critical marker of syntactic dependencies—constitutes a meaningful category within LLMs' representational spaces.

Recent psycholinguistic theories emphasize the interdependence between linguistic expectations and memory limitations in human language processing. We modify the self-attention mechanism of a transformer model to simulate a lossy context representation, biasing the model's predictions to give additional weight to the local linguistic context. We show that surprisal estimates from our locally-biased model generally provide a better fit to human psychometric data, underscoring the sensitivity of the human parser to local linguistic information.

Massively multilingual models can process text in several languages relying on a shared set of parameters; however, little is known about the encoding of multilingual information in single network units. In this work, we study how two semantic variables, namely valence and arousal, are processed in the latent dimensions of mBERT and XLM-R across 13 languages. We report a significant cross-lingual overlap in the individual neurons processing affective information, which is more pronounced when considering XLM-R vis-à-vis mBERT. Furthermore, we uncover a positive relationship between cross-lingual alignment and performance, where the languages that rely more heavily on a shared cross-lingual neural substrate achieve higher performance scores in semantic encoding.

Neural language models are increasingly valued in computational psycholinguistics, due to their ability to provide conditional probability distributions over the lexicon that are predictive of human processing times. Given the vast array of available models, it is of both theoretical and methodological importance to assess what features of a model influence its psychometric quality. In this work we focus on parameter size, showing that larger Transformer-based language models generate probabilistic estimates that are less predictive of early eye-tracking measurements reflecting lexical access and early semantic integration. However, relatively bigger models show an advantage in capturing late eye-tracking measurements that reflect the full semantic and syntactic integration of a word into the current language context. Our results are supported by eye movement data in ten languages and consider four models, spanning from 564M to 4.5B parameters.

It is well known that the surprisal of an upcoming word, as estimated by language models, is a solid predictor of reading times (Smith and Levy, 2013). However, most of the studies that support this view are based on English and few other Germanic languages, leaving an open question as to the cross-lingual generalizability of such findings. Moreover, they tend to consider only the best-performing eye-tracking measure, which might conflate the effects of predictive and integrative processing. Furthermore, it is not clear whether prediction plays a role in non-native language processing in bilingual individuals (Grüter et al., 2014). We approach these problems at large scale, extracting surprisal estimates from mBERT, and assessing their psychometric predictive power on the MECO corpus, a cross-linguistic dataset of eye movement behavior in reading (Siegelman et al., 2022; Kuperman et al., 2020). We show that surprisal is a strong predictor of reading times across languages and fixation measurements, and that its effects in L2 are weaker with respect to L1.

The present work constitutes an attempt to investigate the relational structures learnt by mBERT, a multilingual transformer-based network, with respect to different cross-linguistic regularities proposed in the fields of theoretical and quantitative linguistics. We pursued this objective by relying on a zero-shot transfer experiment, evaluating the model's ability to generalize its native task to artificial languages that could either respect or violate some proposed language universal, and comparing its performance to the output of BERT, a monolingual model with an identical configuration. We created four artificial corpora through a Probabilistic Context-Free Grammar by manipulating the distribution of tokens and the structure of their dependency relations. We showed that while both models were favoured by a Zipfian distribution of the tokens and by the presence of head-dependency type structures, the multilingual transformer network exhibited a stronger reliance on hierarchical cues compared to its monolingual counterpart.

The present paper addresses the study of cross-linguistic and cross-modal iconicity within a deep learning framework. An LSTM-based Recurrent Neural Network is trained to associate the phonetic representation of a concrete word, encoded as a sequence of feature vectors, to the visual representation of its referent, expressed as an HCNN-transformed image. The processing network is then tested, without further training, in a language that does not appear in the training set and belongs to a different language family. The performance of the model is evaluated through a comparison with a randomized baseline; we show that such an imaginative network is capable of extracting language-independent generalizations in the mapping from linguistic sounds to visual features, providing empirical support for the hypothesis of a universal sound-symbolic substrate underlying all languages.

The present paper addresses the study of cross-linguistic phonosemantic correspondences within a deep learning framework. An LSTM-based Recurrent Neural Network is trained to associate the phonetic representation of a word, encoded as a sequence of feature vectors, to its corresponding semantic representation in a multilingual and cross-family vector space. The processing network is then tested, without further training, in a language that does not appear in the training set and belongs to a different language family. The performance of the model is evaluated through a comparison with a monolingual and mono-family upper bound and a randomized baseline. After the assessment of the network's performance, the distribution of phonosemantic properties in the lexicon is inspected in relation to different (psycho)linguistic variables, showing a link between lexical non-arbitrariness and semantic, syntactic, pragmatic, and developmental factors.

The present paper aims to investigate the nature and the extent of cross-linguistic phonosemantic correspondences within a computational framework. An LSTM-based Recurrent Neural Network is trained to associate the phonetic representation of a word, encoded as a sequence of feature vectors, to its corresponding semantic representation in a multilingual vector space. The processing network is tested, without further training, in a language that does not appear in the training set. The performance of the multilingual model is compared with a monolingual upper bound and a randomized baseline. After the quantitative evaluation of its performance, a qualitative analysis is carried out on the network's most effective predictions, showing an inhomogeneous distribution of phonosemantic information in the lexicon, influenced by semantic, syntactic, and pragmatic factors.