Matryoshka
Learning to Drive Black-Box LLMs with LLMs

Introduction

Existing research efforts for improving black-box LLM performance can be largely categorized into two main paradigms:

• In-context learning (ICL)-based methods that are designed to guide LLM in exhibiting specific capabilities or adhering to particular directives. These frameworks necessitate meticulously constructing few-shot demonstrations or prompts for LLMs to emulate or follow, rather than fundamentally advancing their intrinsic capabilities.
• Adapter-based methods that exploit the inherent randomness in LLM generation, producing multiple candidate outputs and subsequently selecting those that optimally satisfy domain-predetermined criteria. Nevertheless, these approaches are highly dependent on the intrinsic synthetic capabilities or built-in functionalities of the black-box LLM, potentially resulting in the selection of a suboptimal candidate when all the generated options are less than ideal.

We propose Matryoshka, a modular framework designed to enhance the advanced problem-solving capabilities of black-box LLMs via controllable multi-turn generations.

Experiments

We evaluate Matryoshka on three diverse tasks, including reasoning (GSM8K / GSM-Hard), planning (AlfWorld), and personalization (LaMP) domains.

Main Results

LaMP



AlfWorld



GSM


Matryoshka consistently surpasses existing baselines, achieving state-of-the-art performance across three domains: personalization, planning, and reasoning. Notably, on the LaMP benchmark, we use gpt-4o-mini as the black-box LLM generator, while on the GSM and AlfWorld benchmarks, we use gpt-3.5-turbo as the black-box LLM generator, demonstrating the versatility of our approach.

Plug-and-Play



We demonstrate that Matryoshka effectively applies the optimized white-box controller to other black-box models in a plug-and-play fashion, requiring no additional training. For example, in the LaMP benchmark, our white-box controller functions as a plug-in with black-box models like gpt-3.5-turbo and gemini-1.5-flash. Experimental results indicate that our controller significantly outperforms other baselines, with large performance gains on LaMP-3 and LaMP-4. Additionally, in solving mathematical problems on GSM8K, a Matryoshka model trained with gpt-3.5-turbo can be seamlessly applied to gpt-4o-mini without extra training costs.

Efficiency

LaMP



We presents the accuracy and ROUGE-L curves for LaMP-2M and LaMP-4, with the x-axis representing the total number of profiles per user. We compare Matryoshka and PAG, utilizing same white-box controller and the black-box model gpt-4o-mini. On LaMP-2M, as the profile count increases, PAG’s performance significantly deteriorates, whereas Matryoshka maintains stable performance and surpasses PAG by an increasing margin. For LaMP-4, both Matryoshka and PAG exhibit similar trends, but Matryoshka consistently outperforms PAG by a substantial and steady margin. These results demonstrate the efficacy of IGO in enhancing the summarization capabilities of the black-box controller, especially when dealing with varying and large amounts of profiles.



AlfWorld



We evaluate the performance of Matryoshka against the vanilla LLaMA3-8B-Instruct and AdaPlanner on ALFWorld, varying the number of closed-loop iterations M during the inference phase. As illustrated, Matryoshka achieves an accuracy exceeding 95% in the open-loop inference setting (M=1), significantly surpassing both AdaPlanner and LLaMA3-8B-Instruct. Furthermore, during an 8-iteration closed-loop inference, Matryoshka maintains the highest accuracy of 97%. These findings indicate that Matryoshka is capable of generating exceptionally high-quality plans, enabling the GPT model serving as the task executor to interact with the environment and complete tasks successfully without requiring closed-loop refinement.

Case Study


Here is a case study of Matryoshka on GSM-Hard, AlfWorld and LaMP-4.
• In GSM-Hard, the white-box controller is responsible for breaking down complex problems into simpler sub-problems, while the black-box generator produces executable code for each sub-problem.
• In Alfworld, the white-box controller decomposes tasks into executable plans, and the black-box generator provides an execution strategy for each plan.
• In LaMP, the white-box controller summarizes each user's history, and the black-box generator creates suitable titles based on this summary.

BibTeX

@article{li2024matryoshka,
        title={Matryoshka: Learning to Drive Black-Box LLMs with LLMs},
        author={Li, Changhao and Zhuang, Yuchen and Qiang, Rushi and Sun, Haotian and Dai, Hanjun and Zhang, Chao and Dai, Bo},
        journal={arXiv preprint arXiv:2410.20749},
        year={2024}
      }