Bonjour à tous,

Vous êtes tous et toutes cordialement invité.es à assister à la présentation de projet du prédoc III de Maryam Hashemzadeh, le 16 décembre à 14h (remote).

Titre : Algorithm and Architecture Design Towards Modularity in LLMs

Date: vendredi 12 décembre à 14h

Location: À distance, lien zoom

Jury

Résumé

As Large Language Models (LLMs) continue to scale, their deployment inreal-world scenarios faces critical bottlenecks regarding computationalefficiency, adaptability, and safety. This thesis proposal argues thatmodularity—both in knowledge representation and model architecture—offers arobust pathway to address these challenges. We investigate this paradigmthrough two distinct but complementary frameworks.

First, we address Knowledge Modularity by introducing Sub-goalDistillation, a method to transfer the reasoning capabilities of massiveLLMs into smaller, resource-constrained agents. By decomposing complexlong-horizon tasks into hierarchical sub-goals, we train a 770M-parameteragent that decouples high-level planning from low-level execution. Testedin the ScienceWorld environment, this approach outperforms standardimitation learning by 16.7% while eliminating the need for real-time LLMinference.

Second, we address Architectural Modularity through SafeMoE, aMixture-of-Experts framework designed to resolve the tension between safetyand informativeness. Unlike tra- ditional alignment methods that often leadto blanket refusals, SafeMoE leverages experts explicitly trained on unsafedomain data, controlled by a safety-aware router. This allows the model tonavigate sensitive topics with nuance, achieving over 20% relativeimprovement in safe response rates compared to baselines whilesimultaneously enhancing informativeness.

Finally, we outline future research directions, including the integrationof generative world-model verifiers for robust planning and the use ofexpert divergence regularization to prevent representation collapse inmodular architectures.