Improving Large Language Model Performance with BRANCH-SOLVE-MERGE

In the realm of artificial intelligence, Large Language Models (LLMs) have become instrumental in various text generation and evaluation tasks. Despite their advanced capabilities, these models often face challenges in maintaining coherence and planning, especially for complex tasks requiring multiple criteria to be met. A recent paper titled "BRANCH-SOLVE-MERGE IMPROVES LARGE LANGUAGE MODEL EVALUATION AND GENERATION" introduces an innovative approach to address these challenges.

Understanding BRANCH-SOLVE-MERGE (BSM)

BRANCH-SOLVE-MERGE (BSM) is a method designed to enhance the performance of LLMs by decomposing complex tasks into manageable sub-tasks.

The process consists of three main components:

• Branch: This module breaks down the main task into several parallel sub-tasks. By simplifying the task, the model can focus on smaller, more manageable pieces of information.
• Solve: Each sub-task is then addressed independently. This step ensures that each aspect of the task receives dedicated attention, improving the overall quality of the solution.
• Merge: Finally, the solutions to the sub-tasks are combined into a cohesive output. This step ensures that the final result is not only accurate but also coherent and well-structured.

Why BSM Matters

Traditional LLMs often struggle with tasks that require adherence to multiple constraints or criteria. This limitation is primarily due to their lack of self-consistency and the ability to effectively plan and decompose problems. BSM addresses these issues by providing a structured framework that enhances the model's ability to handle intricate tasks.

Applications and Benefits

• LLM Response Evaluation: By using BSM, the correctness and consistency of LLM evaluations have improved. The human-LLM agreement increased by up to 26%, and biases were reduced by up to 50%.
• Constrained Text Generation: In tasks like constrained story generation, BSM not only improved the coherence of the stories but also increased constraint satisfaction by 12%.
• These improvements demonstrate that BSM can effectively enhance the performance of various LLMs, including Vicuna, LLaMA-2-chat, and GPT-4, allowing models like LLaMA-2-chat to match or even outperform more advanced models like GPT-4 in certain domains.

Practical Implementation of BSM

To illustrate the practical application of BSM, let's consider a real-world example:

Generating a constrained story based on specific user requirements. The task is to create a story that includes any character named which takes place in any city, and involves interest of point where character discovers something new.

Referred article : https://arxiv.org/abs/2310.15123