Reproduce “Multiverse: Your Language Models Secretly Decide How to Parallelize and Merge Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xinyu Yang, Yuwei An, Hongyi Liu, Tianqi Chen, Beidi Chen
Multiverse is a framework that enables Large Language Models (LLMs) to perform native parallel generation and merging of reasoning steps, inspired by the MapReduce paradigm. This approach significantly reduces inference latency by up to 2x for long-context generation and scales effectively across various batch sizes, while maintaining or improving reasoning performance.
Key insights
- Despite their sequential nature, autoregressive LLMs implicitly generate reasoning paths containing significant latent parallelism, especially in Chain-of-Thought (CoT) trajectories.
- By explicitly structuring and training LLMs to recognize and leverage this intrinsic parallelism through an internalized MapReduce paradigm, it is possible to bridge the gap between AR-LLM reasoning capability and the efficiency of parallel generation.
- A co-design approach involving automated data curation, modified attention mechanisms, and a dynamic inference engine is essential for seamlessly integrating native parallel generation into existing pre-trained AR-LLMs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.