CLIMB brings together Linguistics, Cognitive Science and Artificial Intelligence to investigate how humans acquire and process language. The project will develop Language Models whose training conditions and design more closely reflect human learning, creating new experimental tools for understanding the structures and mechanisms underlying our linguistic abilities.
1. Motivations and goals
Children learn language through limited but rich experience: listening, observing and interacting with others. Today’s Large Language Models achieve remarkable linguistic performance, yet typically learn from vast quantities of written text under conditions very different from those experienced by human learners. This gap limits what their achievements can tell us about human language acquisition.
CLIMB investigates whether bringing Language Models closer to human learning conditions can make them more informative scientific tools. Its approach combines developmentally appropriate text, speech and visual input with model architectures and learning objectives informed by linguistic and cognitive theories.
The project pursues three main goals:
- Develop cognitively grounded Language Models. Create models trained on input that more closely reflects human experience, and investigate how different learning constraints shape their linguistic abilities.
- Evaluate language development across languages and modalities. Build challenging benchmarks that track how linguistic knowledge emerges over time and enable comparisons between artificial and human learners.
- Explain how Language Models represent and process language. Identify the linguistic information encoded within models and establish how it contributes to their behaviour.
Together, these goals will support controlled experiments on language acquisition and processing, including the interaction between syntax and semantics and the mechanisms that enable generalisation from limited input.
2. Target languages
CLIMB will develop and evaluate Language Models for English, French, German, Italian, Norwegian and Mandarin Chinese.
This selection enables comparisons both among closely related languages and across languages with substantially different structures. Differences in morphology, case marking and word order will help reveal which learning patterns are shared and which depend on the properties of individual languages. Norwegian also brings a focus on language modelling in settings with fewer resources.
Across all six languages, CLIMB will examine how Language Models develop over time, how different learning constraints affect acquisition within each language, and how learning trajectories vary across languages.
3. Work package descriptions
WP1 — Multimodal data for CLIMB-LM training
CLIMB will build training datasets that combine written texts, speech and visual information across the six target languages. Sources will include conversational recordings, children’s programmes, films, television series and news broadcasts, with particular attention to interactive content and material intended for different age groups.
The data will be organised into incremental subsets reflecting developmental stages, allowing models to encounter progressively richer and more complex input. Cross-language comparability and alignment between modalities will guide data collection.
Expected output: Developmentally structured text, speech and visual training datasets for six languages.
WP2 — Theory-driven CLIMB-LM design
This work package will develop Language Models that incorporate learning principles drawn from linguistics and cognitive science. These include hierarchical structure, constraints on dependencies between linguistic elements, and parallel processing of syntax and semantics.
The project will explore different neural architectures (Transformers, Recurrent Neural Networks, JEPA, etc.), integrating these principles into both model design and training objectives. Comparing models with different constraints will help establish how particular assumptions influence language learning.
Expected output: A family of CLIMB language models incorporating different theory-driven learning constraints.
WP3 — Multilingual and multimodal evaluation framework
CLIMB will create a benchmark covering linguistic knowledge from morphology and syntax to semantics and pragmatics. Evaluation materials will combine linguistic, audio and visual stimuli across all six languages.
Tests will use carefully controlled contrasts, including minimal pairs, to investigate specific linguistic abilities. They will also be aligned with human evidence, including eye-tracking data, acceptability judgements and developmental comprehension scores, enabling comparisons between model and human learning trajectories.
Expected output: A richly annotated multilingual and multimodal benchmark for evaluating competence and development of Language Models.
WP4 — Mechanistic interpretability and model explainability
This work package will investigate how linguistic knowledge is represented inside CLIMB Language Models and how it influences their responses. Interpretability methods will identify internal features and computational pathways associated with particular linguistic phenomena.
Analyses will track how representations emerge and change during learning, examining both patterns shared across languages and language-specific differences. The aim is to connect internal model mechanisms with explanations grounded in linguistic theory.
Expected output: A comprehensive “linguistic map” of the internal representations of CLIMB Language Models, supporting causal explanations of their behaviour.
4. Expected results and impact
New insights into human language learning. CLIMB aims to clarify how linguistic input, learning constraints and interactions between language domains shape acquisition and processing. Cross-language comparisons will help distinguish general learning mechanisms from language-specific patterns.
A framework for testing linguistic and cognitive theories. By connecting training conditions, developmental trajectories, internal representations and behaviour, the project will provide a more rigorous basis for using language models as scientific experimental tools.
Open resources for the research community. Training datasets, language models, evaluation benchmarks and analysis tools will support further research on multilingual learning, multimodal processing and model interpretability.
Progress towards more efficient and transparent Artificial Intelligence (AI). CLIMB will explore how cognitively informed design can improve learning from limited data and make model behaviour easier to explain. Its focus on smaller, accessible models is intended to broaden participation in language technology research and support more sustainable AI development.