ReLLM-C2-3B

Model Card: ReLLM-C2

1. Model Summary

The ReLLM-C2 model is a Large Language Model (LLM) specifically fine-tuned to act as a surrogate model for multi-objective optimization in computationally expensive optimization tasks.

It serves as a core modeling component within the R2SAEA (Reinforced Relation Surrogate-Assisted Evolutionary Algorithm) framework. Unlike general-purpose LLMs, ReLLM-C2 is designed to seamlessly integrate with Evolutionary Algorithms (EAs). By leveraging structured prompt templates containing decision variables and objective data, the model can perform zero-shot relationship reasoning to evaluate and classify candidate solutions in multi-objective optimization scenarios.

2. Intended Use

• Primary Application: Relational-based surrogate modeling in multi-objective Evolutionary Algorithms.
• Out-of-Scope Use: ReLLM-C2 is heavily fine-tuned specifically for numerical optimization and relationship reasoning. It is not intended for general conversational chat, creative writing, or standard code generation tasks.

3. Background & Related Work

This model bridges the gap between Large Language Models (LLMs) and Evolutionary Algorithms (EAs), addressing a critical bottleneck in the field of Surrogate-Assisted Evolutionary Algorithms (SAEAs):

• The Problem with Traditional SAEAs: Conventional machine learning surrogate models (such as Gaussian Processes or Random Forests) require being retrained from scratch at every single generation using new evaluated data, which introduces massive computational overhead.
• Our Methodology: Through the R2SAEA framework, we transform the relationship reasoning problem in optimization tasks into a Reinforcement Learning (RL) problem.
• Training Alignment: ReLLM-C2 is trained using the Group Relative Policy Optimization (GRPO) algorithm. This aligns the LLM's reasoning capabilities directly with multi-objective optimization goals, granting it the ability to perform zero-shot classification across a wide range of unseen tasks. This eliminates the need for generation-by-generation retraining while significantly reducing the computational burden associated with using general-purpose LLMs.

4. GitHub Repository

To utilize ReLLM-C2 effectively, it should be deployed alongside the R2SAEA framework, which handles prompt structuring and the evolutionary loop. The framework provides implementations in both Python (via pymoo) and MATLAB (via PlatEMO).

For deployment instructions, API configuration, and framework integration, please visit our official repository:

• GitHub Repository: [R2SAEA]

5. License

The ReLLM-C2 model and the associated R2SAEA framework are open-sourced under the Apache License 2.0.

6. Citation

If you use this model or the R2SAEA framework in your research, please cite our work:

@misc{r2saeagithub,
  title={R2SAEA: Relation Reasoning with LLMs in Expensive Optimization},
  author={Ye Lu, BingDong Li, Aimin Zhou, Hao Hao},
  year={2026},
}