Internship Opportunity Self-Improving Large Language Models through Multi-Agent Reasoning
Pubblicato il 14-09-2026 - Fondazione Bruno Kessler - Fbk in Italia
--- | --- | | | | | --- | | Internship Opportunity Self-Improving Large Language Models through Multi-Agent Reasoning | |Back| | || --- | --- || | | | --- | | The Bruno Kessler Foundation (FBK) is a leading research institution based in Trento, Italy. With over 450 researchers distributed across 12 Research Centers, FBK promotes excellence in science and technology across both technical and social domains. The internship is offered within the Intelligent Digital Agents (IDA) Research Unit of the Digital Health & Wellbeing Center. The unit focuses on developing flexible, ethically grounded technologies and digital assistants, with particular attention to multi-agent systems, reasoning, human-centered AI, and digital health applications. More information can be found at Planned activities The Research Unit is looking for a Master’s student interested in pursuing an internship experience on Self-Improving Large Language Models through Multi-Agent Reasoning. The internship will explore how a general-purpose Large Language Model, such as Gemma or similar open-weight models,
can improve its reasoning performance without relying on external training data. The work will build on the Pool of Experts (PoE) framework, a multi-agent architecture in which several expert agents analyze a problem from different perspectives and cooperate to produce a final decision. The main goal of the internship is to investigate whether PoE can be used not only as an inference-time reasoning framework, but also as a mechanism for model self-improvement. Starting from a generalist LLM, the intern will study how the multi-agent system can be used to improve model performance without introducing new annotated datasets. Main activities include: * Exploring self-improvement strategies based on internal model outputs, such as self-consistency, self-reflection, critique, debate, and self-distillation; * Designing experiments in which PoE agents generate, evaluate, revise, or select improved re