LLM Policy Induction for Heuristic Search Control: A Trace-Driven ALNS Framework

Feedback-Driven Population Self-Evolution Framework for Dispatching Rule Generation in Dynamic Job Shop via Knowledge Distillation

The dynamic job shop scheduling problem (DJSSP) is critical for optimizing production efficiency in intelligent manufacturing systems under dynamic constraints. Traditional approaches, including heuristic dispatching rules (HDRs) and evolutionary hyper-heuristics, often struggle to generalize across dynamic and unseen scheduling scenarios. The advent of large language models (LLMs) offers new opportunities for scheduling optimization, yet their deployment in real-world manufacturing is constrained by data privacy concerns that limit the use of online LLM services and by the restricted inference capability of local small-scale models. To address these challenges, this paper proposes a novel feedback-driven self-evolution framework employing dual-stage knowledge distillation. The framework utilizes a feedback-driven refinement mechanism to distill expert knowledge from advanced LLMs, which is then used to fine-tune the Qwen2.5-Coder-7B model into ModouGPT, a specialized model for generating HDRs. Experimental results indicate that our approach performs better than traditional methods and five state-of-the-art LLM-based evolutionary frameworks, particularly in dynamic and unseen scheduling scenarios. ModouGPT demonstrates competitive performance with online LLMs in complex dynamic environments, while ensuring data privacy. This study provides manufacturing enterprises, particularly small and medium-sized enterprises, with a deployable scheduling system that balances intelligence, privacy, and computational efficiency in dynamic job shop environments. Note to Practitioners—DJSSP involves constant disruptions such as machine breakdowns and dynamic order arrivals. Deploying artificial intelligence (AI)-based scheduling in real manufacturing environments remains difficult: cloud-based LLM services risk exposing sensitive production data, while large-scale local models demand prohibitive hardware investments, placing both options out of reach for most enterprises. This paper presents a practical solution. The proposed framework distills scheduling knowledge from a powerful online teacher model into a compact 7B local model, ModouGPT, deployable on a single consumer-grade graphics processing unit (GPU). All scheduling decisions are made entirely on-premises, with no data leaving the factory floor. Practitioners do not need domain expertise to design HDRs. The system automatically generates HDRs tailored to the specific production environment, covering job processing characteristics and dynamic disturbances. An offline HDR library can be built in advance and updated periodically without interrupting production, and new HDRs can be synthesized in under one minute during live operations. This makes the system immediately actionable for engineers seeking to upgrade scheduling capabilities without relying on external LLM services or specialized algorithm designers.

LLM-Guided Evolutionary Search for Constraint Model Reformulation to Improve Solver Efficiency

Combinatorial problems appear in numerous industrial applications. A common approach is to formulate these problems as declarative constraint models that can subsequently be compiled to and solved by a range of back-end solvers. Recent work shows that Large Language Models (LLMs) can produce correct models from natural language, but even a correct model can be expensive to solve because performance remains sensitive to modelling choices. In this work, we investigate whether LLMs can automate performance-oriented model reformulation. Inspired by Automatic Heuristic Design (AHD), we use an evolutionary framework in which an LLM proposes candidate reformulations that are verified and benchmarked against the user-defined baseline model. We compare AHD-adapted search strategies that control which prior attempts, instructions, and measured feedback enter each prompt. Existing retention strategies prioritize recency or performance, but do not explicitly diversify the context. To cover this gap, we introduce Profile-Diverse Retention (PDR), which applies Maximal Marginal Relevance (MMR) to instance-level runtime vectors to retain behaviourally diverse attempts. We systematically evaluate the strategies on eight CSPLib problems using validation-based final model selection. The results show that: (i) iterative reformulation can produce substantial held-out speedups; (ii) strategies that keep the retained context diverse outperform those that retain only recent or the fastest attempts; and (iii) validation-based selection improves the held-out speedup of every strategy.

LLM-Guided Heuristic Design from Simulation Traces: A Case Study in Dynamic Production and AGV Scheduling

Simulation-based optimization (SBO) evaluates executable policies under stochastic dynamics, but most methods treat the simulator as a black box: aggregate scores rank candidates without revealing why they fail or which policy logic should change. We present an LLM-guided heuristic design framework that uses repeated simulation for selection and event-level traces for diagnosis. Each incumbent is assessed through multiple replications, while replaying its lowest-scoring one produces a queryable trace. A manager agent formulates bottleneck hypotheses from this evidence, and editing agents implement parallel code-level revisions. After execution checks and repeated evaluation, best-so-far selection retains only improvements. LLM revision occurs between evaluation batches, while a fixed policy controls each simulation run. We evaluate the framework in a discrete-event simulation of dynamic production and automated guided vehicle (AGV) scheduling. Across five independent optimization runs with Gemini-3.1-Pro, final mean scores averaged 77.51 on the simulator's 0-100 scale. In the highest-scoring run, trace-based diagnoses motivated proactive charging, distance-aware AGV assignment, and rebalanced dispatch priorities, raising the best-so-far mean score from 62.49 to 78.61. On 100 matched seeds, the best final policy outscored representative rolling-MILP, rule-based, and metaheuristic policies on every seed and retained its advantage under random faults without re-optimization. After separate re-optimization for a longer horizon and variable order interarrival times, the resulting policies again outscored all baselines. Ablations with two LLM backbones showed that removing either parallel candidate generation or trace-database access reduced final mean scores. These results show that simulation traces can guide targeted code-level policy improvement in complex simulation-based scheduling.

Preference-Guided Multi-Policy Optimization for Flexible Job Shop Scheduling

Guiding Large Language Models with Genetic Programming-Evolved Heuristic Knowledge for Dynamic Multi-Mode Project Scheduling

In dynamic multi-mode project scheduling, activities have alternative execution modes and uncertain durations, while precedence relations and limited resources constrain their execution. Heuristic priority rules support fast online decisions, but their design requires substantial domain expertise. Genetic programming (GP) hyper-heuristics can automatically evolve such rules. Large language models (LLMs), meanwhile, provide a flexible interface for interpreting scheduling information and explaining decisions. However, zero-shot LLM decisions may lack domain knowledge, consume many tokens, and vary across repeated queries. GP-evolved rules therefore provide a potential source of scheduling knowledge for guiding LLM decisions. Unlike existing LLM--GP hybrids that use LLMs to support heuristic evolution, we transfer knowledge in the reverse direction, using knowledge extracted from high-quality GP rules to guide an online LLM decision maker. We extract knowledge from high-quality GP rules and inject it through Feature Selection, Feature Hint, Rule Reference, and Rule Follow. These mechanisms are evaluated in terms of scheduling performance, token consumption, decision stability, and the feature focus expressed in generated rationales. GP-derived guidance generally improves the unguided LLM, but its representation matters. Simplifying the decision context or supplying explicit decision logic is more effective than highlighting important features. Feature Selection offers the best token efficiency, whereas Rule Follow achieves strong performance at greater token cost. Guidance also improves decision stability and changes the features expressed in generated rationales.

LLM-MetaAS: A Semantic-Statistical Policy Routing Framework for AutoML Execution Strategy Selection

Automated machine learning (AutoML) aims to reduce manual intervention in machine learning pipeline construction. However, AutoML performance is affected not only by the searched pipeline but also by high-level execution policies, such as validation protocols and budget allocation strategies. Auto-sklearn 2.0 has shown that these policies can be selected automatically, but its lightweight task representation mainly captures dataset scale and may not distinguish tasks with similar sizes but different semantic and structural characteristics. To address this limitation, this article proposes LLM-MetaAS, a semantic-statistical framework for AutoML execution policy selection. Building on the established pairwise policy-selection paradigm, LLM-MetaAS focuses on improving task representation rather than introducing pairwise decomposition itself. It constructs a policy-oriented fingerprint by combining LLM-assisted semantic profiling with lightweight statistical descriptors related to validation reliability and computational demand. Policy quality is evaluated using an explicit regret-based criterion relative to the empirical oracle, and vote margins are used to analyze routing uncertainty. Experiments on 39 benchmark tabular classification datasets show that LLM-MetaAS improves overall AutoML performance and selects policies closer to the oracle than fixed strategies, random selection, and the native Auto-sklearn 2.0 selector. Ablation and robustness analyses further support the utility of the complete semantic-statistical representation within the evaluated framework.