Research Note
Baka Mohamed
Agadir, Morocco
bakamoohamed@[Link] | +212 652-174360
One challenge I would like to explore at the intersection of Large Language Models (LLMs)
and optimization is bridging the gap between natural language reasoning and constraint satis-
faction guarantees.
LLMs are highly effective at interpreting unstructured requirements and proposing candidate
solutions, but they struggle to ensure that outputs strictly adhere to operational or mathematical
constraints. Conversely, symbolic solvers such as MILP, SAT, or CP guarantee feasibility but
lack flexibility and human interpretability.
My proposal is to study hybrid architectures where LLMs act both as translators and heuristic
guides. In such a framework, LLMs would first convert high-level, human-readable requirements
into formal constraint models. They could then provide informed priors to guide solvers, reducing
the search space and accelerating optimization while still preserving correctness.
I am especially interested in applying this approach to scheduling and planning domains, such
as healthcare staff allocation, energy distribution, or supply chain optimization, where fairness,
robustness, and efficiency are equally critical. Evaluation in these domains would consider not
only accuracy and runtime, but also robustness, transparency, and trustworthiness.
I believe this direction can lead to scalable, constraint-aware AI systems that combine the
strengths of neural and symbolic methods, paving the way for reliable decision-making tools in
real-world operations.