Thèse Vers des Systèmes Multi-Agents Efficaces et Adaptatifs Grâce à la Distillation de Connaissances et à l'Apprentissage par Renforcement H/F - Doctorat.Gouv.Fr

Établissement : Université Paris-Saclay GS Informatique et sciences du numérique École doctorale : Sciences et Technologies de l'Information et de la Communication Laboratoire de recherche : Laboratoire Méthodes Formelles Direction de la thèse : Fatiha ZAIDI Début de la thèse : 2026-10-01 Date limite de candidature : 2026-09-15T23:59:59 Les systèmes multi-agents (MAS) fondés sur des grands modèles de langage (LLM) constituent aujourd'hui une approche prometteuse pour résoudre des tâches complexes nécessitant planification, collaboration et utilisation d'outils. En répartissant le raisonnement entre plusieurs agents spécialisés, ils obtiennent des performances supérieures à celles d'un agent unique. Toutefois, ces architectures restent coûteuses en calcul et en mémoire, reposent sur des modèles généralistes largement surdimensionnés pour les tâches considérées et utilisent des protocoles de communication et de gestion de la mémoire qui ne sont pas appris mais hérités des capacités des LLM sous-jacents. Une fois déployés, ces systèmes demeurent également peu adaptatifs et nécessitent généralement un réentraînement complet pour intégrer de nouvelles connaissances ou améliorer leurs comportements.
Cette thèse propose de concevoir une nouvelle génération de systèmes multi-agents compacts, spécialisés et capables d'apprentissage tout au long de la vie, en combinant la distillation de connaissances et l'apprentissage par renforcement. La distillation vise à transférer uniquement les connaissances et compétences nécessaires depuis un ou plusieurs modèles teacher vers des agents student plus légers, tandis que l'apprentissage par renforcement est utilisé pour optimiser progressivement leurs stratégies de communication, leur gestion de la mémoire et leur coopération au cours des interactions. L'objectif est ainsi de ne plus considérer la communication et la mémoire comme des mécanismes fixes, mais comme des comportements appris et continuellement améliorés.
Les contributions attendues de cette thèse sont triples : (i) développer des méthodes de distillation dédiées aux systèmes multi-agents afin de produire des agents spécialisés et compacts ; (ii) concevoir des approches d'apprentissage par renforcement permettant d'apprendre des protocoles de communication et des mécanismes de mémoire efficaces, adaptés aux contraintes des systèmes multi-agents ; et (iii) proposer un cadre d'apprentissage continu permettant à ces agents d'améliorer leurs performances après leur déploiement sans réentraînement complet ni oubli catastrophique. L'ambition est de contribuer à l'émergence de systèmes multi-agents plus efficaces, plus adaptatifs et plus autonomes, capables d'apprendre de leurs expériences tout au long de leur cycle de vie.
Multi-agent systems based on large language models (LLM-based Multi-Agent Systems, MAS) have established themselves as a promising architecture for solving complex tasks requiring planning, distributed reasoning, tool use, and collaboration among several specialized agents. Frameworks such as AutoGen, LangGraph, CrewAI, Strands Agents, or Microsoft Agent Framework now facilitate the development of these systems, which demonstrate remarkable performance on numerous reasoning and automation benchmarks.
Despite these advances, current architectures present several fundamental limitations. They generally rely on very large general-purpose LLMs, a large part of whose knowledge and capabilities is not necessary for the specific tasks assigned to the agents. This redundancy leads to significant consumption of compute, memory, and token resources. Moreover, communication strategies between agents are rarely optimized: exchanges take place mainly in natural language, without the agents explicitly learning what to communicate, when, or to which interlocutors.
In addition, memory is today one of the main challenges of multi-agent systems. Current approaches rely essentially on retrieval-augmented generation (RAG) mechanisms, extended context windows, or external memories, but they make it difficult for agents to durably acquire new knowledge and improve their behaviors over the course of interactions. Recent work, notably that of Letta on Memory Models, suggests instead that memory should be regarded as a learning capability in its own right, one that can be optimized and consolidated over the course of an agent's life.
Finally, recent benchmarks such as LifelongAgentBench show that current agents are still mainly evaluated on independent tasks and possess limited continual-learning capabilities. Once deployed, they generally require a new training or fine-tuning cycle to integrate new knowledge, which strongly limits their adaptability in evolving environments.
These observations motivate the development of a new generation of multi-agent systems capable of being at once more compact, more efficient, and capable of learning throughout their life cycle.
This thesis proposes to design a new generation of multi-agent systems that are compact, specialized, and capable of lifelong learning, by combining knowledge distillation and reinforcement learning. Distillation aims to transfer only the knowledge and skills that are necessary from one or more teacher models to lighter student agents, while reinforcement learning is used to progressively optimize their communication strategies, memory management, and cooperation over the course of interactions. The goal is thus to no longer treat communication and memory as fixed mechanisms, but as learned behaviors that are continuously improved.
The expected contributions of this thesis are threefold: (i) developing distillation methods dedicated to multi-agent systems in order to produce specialized and compact agents; (ii) designing reinforcement-learning approaches that make it possible to learn efficient communication protocols and memory mechanisms, adapted to the constraints of multi-agent systems; and (iii) proposing a continual-learning framework enabling these agents to improve their performance after deployment without full retraining or catastrophic forgetting. The ambition is to contribute to the emergence of multi-agent systems that are more efficient, more adaptive, and more autonomous, capable of learning from their experience throughout their life cycle.
The central problem of this thesis is therefore the following:
How can MAS based on LLMs be designed so that, through knowledge distillation and reinforcement learning, they become more compact, learn to communicate efficiently, develop a memory adapted to their tasks, and continue to improve after deployment without forgetting the knowledge they have acquired?
Research Sub-Questions
SQ1 - Distillation
How can the knowledge, skills, and reasoning capabilities useful for a given multi-agent task be identified and distilled from one or more teacher models, in order to build specialized, compact agents?
SQ2 - Communication
How can reinforcement learning be used to enable distilled agents to learn effective communication strategies (what to transmit, when to transmit, and to which agents), rather than relying exclusively on protocols inherited from general-purpose LLMs?
SQ3 - Memory
How can memory management adapted to multi-agent systems be learned by combining parametric memory (encoded in model weights), external memory, and reinforcement learning, so as to enable agents to durably accumulate new knowledge?
SQ4 - Continual Learning
How can these distilled multi-agent systems be enabled to keep learning after deployment through reinforcement learning, while preserving their initial competencies and limiting catastrophic forgetting?

Master en informatique avec compétences en IA et Data Science et programmation

• Distillation
• Machine learning