Recent Vision-Language Models (VLMs) like BLIP, LLaVA, and Qwen-VL have achieved impressive results in multimodal tasks but still face limitations in true spatial and temporal reasoning. Many current benchmarks conflate visual reasoning with general knowledge and involve shallow reasoning tasks. Furthermore, these models often struggle with understanding complex spatial relations and dynamic scenes due to suboptimal visual feature usage. To address this, recent approaches such as SpatialRGPT, SpaceVLLM, VPD, and ST-VLM have introduced techniques like 3D scene graph integration, spatio-temporal queries, and kinematic instruction tuning to improve reasoning over space and time. This thesis proposes to build on these advances by developing new instruction-tuned models with improved data representation and architectural innovations. The goal is to enable robust spatio-temporal reasoning for applications in robotics, video analysis, and dynamic environment understanding.

Video Anomaly Detection (VAD) aims to automatically identify unusual events in video that deviate from normal patterns. Existing methods often rely on One-Class or Weakly Supervised learning: the former uses only normal data for training, while the latter leverages video-level labels. Recent advances in Vision-Language Models (VLMs) and Large Language Models (LLMs) have improved both the performance and explainability of VAD systems. Despite progress on public benchmarks, challenges remain. Most methods are limited to a single domain, leading to performance drops when applied to new datasets with different anomaly definitions. Additionally, they assume all training data is available upfront, which is unrealistic for real-world deployment where models must adapt to new data over time. Few approaches explore multimodal adaptation using natural language rules to define normal and abnormal events, offering a more intuitive and flexible way to update VAD systems without needing new video samples.

This PhD research aims to develop adaptable Video Anomaly Detection methods capable of handling new domains or anomaly types using few video examples and/or textual rules.

The main lines of research will be the following:
• Cross-Domain Adaptation in VAD: improving robustness against domain gaps through Few-Shot adaptation;
• Continual Learning in VAD: continually enriching the model to deal with new types of anomalies;
• Multimodal Few-Shot Learning: facilitating the model adaptation process through rules in natural language.

To perform an unknown task, a subject (human or robot) has to consult external information, which involves a cognitive cost. After several similar experiments, it masters the situation and can act automatically. The 1980s and 1990s saw explorations in AI using conceptual graphs and schemas, but their large-scale implementation was limited by the technology available at the time.

Today's neural models, including transformers and LLM/VLMs, learn universal representations through pre-training on huge amounts of data. They can be used with prompts to provide local context. Fine-tuning allows these models to be specialised for specific tasks.

RAG and GraphRAG methods can be used to exploit external knowledge, but their use for inference is resource-intensive. This thesis proposes a cognitivist approach in which the system undergoes continuous learning. It consults external sources during inference and uses this information to refine itself regularly, as it does during sleep. This method aims to improve performance and reduce resource consumption.

In humans, these processes are linked to the spatial organisation of the brain. The thesis will also study network architectures inspired by this organisation, with dedicated but interconnected “zones”, such as the vision-language and language models.

These concepts can be applied to the Astir and Ridder projects, which aim to exploit foundation models for software engineering in robotics and the development of generative AI methods for the safe control of robots.

This PhD project focuses on enhancing Large Multimodal Models (LMMs) through the integration of fine-grained and spatio-temporal information into training datasets. While current LMMs such as CLIP and Flamingo show strong performance, they rely on noisy and coarse-grained image-text pairs and often lack spatial or temporal grounding. The thesis aims to develop automatic pipelines to enrich image datasets with geographic and temporal metadata, refine captions using fine-grained semantic descriptors, and balance dataset diversity and compactness by controlling class-wise sample sizes.

Training strategies will incorporate hierarchical class structures and adapt protocols to improve alignment between caption elements and image regions. The work will also explore joint training regimes that integrate fine-grained, spatial, and temporal dimensions, and propose set-based inference to improve the diversity of generated outputs. The enriched datasets and models will be evaluated using existing or newly developed benchmarks targeting contextual relevance and output diversity. The project also addresses challenges in metadata accuracy, efficient model adaptation, and benchmarking methodologies for multi-dimensional model evaluation.

Applications include improved synthetic data generation for autonomous driving, enhanced annotation of media archives through contextual captioning, and better visual reasoning in industrial simulation scenarios.

Conversational agents are increasingly present in our daily lives thanks to advances in natural language processing and artificial intelligence and are attracting growing interest. However, their ability to understand human communication in all its complexity remains a major challenge. This PhD project aims to model linguistic variation to develop agents capable of socially adaptive interactions, taking into account the socio-demographic profile and emotional state of their interlocutors. It also focuses on evaluating linguistic cues at different levels, leveraging both spoken and written language varieties, and assessing the generalization capacity of models trained on multilingual and multi-situational data, with the goal of improving interaction modeling with conversational agents.

The advent of the foundation models led to increase the state-of-the art performance on a large number of tasks in several fields of AI, in particular computer vision and natural language processing. However, despite the huge amount of data used to train them, these models are still limited in their ability to generalize, in particular for a use case of interest that is in a specific domain, not well represented on the Web. A way to formalize this issue is compositional generalization, i.e. generalising to a new, unseen concept from concepts learned during training. This "generalization" is the ability to learn disentangle concepts and to be able to recombine
them into unseen composition when the model is in production. The proposed thesis will address this issue, aiming at proposing visual representations that enable generic visual language models to generalize compositionally within specific domains. It will investigate strategies to reduce shortcut learning, promoting deeper understanding of compositional structures in multimodal data. It will also address the problem of compositional generalization beyond simple attribute–object pairs, capturing more subtle and complex semantics. The proposed thesis aims at proposing preogress at a quite theoretical level but has many potential practical interest, in the fields of health, administration and services sectors, security and defense, manufacturing and agriculture.