Branch-Solve-Merge Improves Large Language Model Evaluation and Generation

The article proposes a novel method, BRANCH-SOLVE-MERGE (BSM), to enhance the performance of Large Language Models (LLMs) in handling intricate language tasks. BSM comprises three modules – ‘branch’, ‘solve’, and ‘merge’ – that work in sync to plan, independently solve and combine solutions to various sub-tasks. BSM has been tested on multiple LLMs such as Vicuna, LLaMA-2-chat, and GPT-4 and has shown significant improvements in evaluation correctness, consistency, coherence and constraint satisfaction.

Publication date: 23 Oct 2023
Project Page: https://arxiv.org/abs/2310.15123
Paper: https://arxiv.org/pdf/2310.15123