Scholarlink

The University Course Timetabling Problem (UCTP) is a tightly constrained NP-hard optimization task that requires strict hard-constraint feasibility while minimizing soft-constraint penalties. Although meta-heuristics such as Simulated Annealing and Tabu Search remain highly competitive, the emergence of Large Language Models (LLMs) raises the question of whether semantic reasoning can contribute to structured timetabling. This paper presents LLM-CALS, a hybrid framework integrating constructive heuristics, adaptive local search, and multi-mode LLM interaction. We evaluate four architectures: (1) Algorithm-Pure, (2) LLM-Pure (one-shot generation), (3) LLM-First (semantic initialization), and (4) Algo-First (LLM-guided refinement). Experiments on the ITC-2007 benchmark show that Algorithm-Pure and Algo-First achieve the highest feasibility and solution quality, whereas LLM-Pure frequently violates hard constraints and LLM-First offers limited benefits due to structural inconsistencies in LLM-generated schedules. Overall, the findings indicate that current LLMs are most effective as auxiliary semantic refiners rather than replacements for domain-specific optimization procedures.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LLM-CALS: A Hybrid Optimization Framework Integrating Large Language Models for Curriculum-Based Course Timetabling

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Dinh-Hong Vu,
  • Quang-Huy Duong,
  • Thi-My-Thanh Lai

摘要

The University Course Timetabling Problem (UCTP) is a tightly constrained NP-hard optimization task that requires strict hard-constraint feasibility while minimizing soft-constraint penalties. Although meta-heuristics such as Simulated Annealing and Tabu Search remain highly competitive, the emergence of Large Language Models (LLMs) raises the question of whether semantic reasoning can contribute to structured timetabling. This paper presents LLM-CALS, a hybrid framework integrating constructive heuristics, adaptive local search, and multi-mode LLM interaction. We evaluate four architectures: (1) Algorithm-Pure, (2) LLM-Pure (one-shot generation), (3) LLM-First (semantic initialization), and (4) Algo-First (LLM-guided refinement). Experiments on the ITC-2007 benchmark show that Algorithm-Pure and Algo-First achieve the highest feasibility and solution quality, whereas LLM-Pure frequently violates hard constraints and LLM-First offers limited benefits due to structural inconsistencies in LLM-generated schedules. Overall, the findings indicate that current LLMs are most effective as auxiliary semantic refiners rather than replacements for domain-specific optimization procedures.