Shipeng Cen

Feihu Jin, Shipeng Cen, Ying Tan

Fine-tuning large language models (LLMs) has achieved remarkable success across various NLP tasks, but the substantial memory overhead during backpropagation remains a critical bottleneck, especially as model scales grow. Zeroth-order (ZO) optimization alleviates this issue by estimating gradients through forward passes and Gaussian sampling, avoiding the need for backpropagation. However, conventional ZO methods suffer from high variance in gradient estimation due to their reliance on random perturbations, leading to slow convergence and suboptimal performance. We propose a simple plug-and-play method that incorporates prior-informed perturbations to refine gradient estimation. Our method dynamically computes a guiding vector from Gaussian samples, which directs perturbations toward more informative directions, significantly accelerating convergence compared to standard ZO approaches. We further investigate a greedy perturbation strategy to explore the impact of prior knowledge on gradient estimation. Theoretically, we prove that our gradient estimator achieves stronger alignment with the true gradient direction, enhancing optimization efficiency. Extensive experiments across LLMs of varying scales and architectures demonstrate that our proposed method could seamlessly integrate into existing optimization methods, delivering faster convergence and superior performance. Notably, on the OPT-13B model, our method outperforms traditional ZO optimization across all 11 benchmark tasks and surpasses gradient-based baselines on 9 out of 11 tasks, establishing a robust balance between efficiency and accuracy.

Shipeng Cen, Ying Tan

As optimization problems grow increasingly complex and diverse, advancements in optimization techniques and paradigm innovations hold significant importance. The challenges posed by optimization problems are primarily manifested in their non-convexity, high-dimensionality, black-box nature, and other unfavorable characteristics. Traditional zero-order or first-order methods, which are often characterized by low efficiency, inaccurate gradient information, and insufficient utilization of optimization information, are ill-equipped to address these challenges effectively. In recent years, the rapid development of large language models (LLM) has led to substantial improvements in their language understanding and code generation capabilities. Consequently, the design of optimization algorithms leveraging large language models has garnered increasing attention from researchers. In this study, we choose the fireworks algorithm(FWA) as the basic optimizer and propose a novel approach to assist the design of the FWA by incorporating multi-modal large language model(MLLM). To put it simply, we propose the concept of Critical Part(CP), which extends FWA to complex high-dimensional tasks, and further utilizes the information in the optimization process with the help of the multi-modal characteristics of large language models. We focus on two specific tasks: the \textit{traveling salesman problem }(TSP) and \textit{electronic design automation problem} (EDA). The experimental results show that FWAs generated under our new framework have achieved or surpassed SOTA results on many problem instances.