Reproduce “Slm-mux: Orchestrating small language models for reasoning” 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
Chenyu Wang, Zishen Wan, Hao Kang, Emma Chen, Zhiqiang Xie, Tushar Krishna, Vijay Janapa Reddi, Yilun Du
SLM-MUX introduces an architecture for orchestrating Small Language Models (SLMs) for reasoning tasks, bypassing inter-model discussion by using self-consistency to estimate confidence. This approach achieves accuracy gains of up to 13.4% on MATH and allows SLM ensembles to rival or surpass the performance of much larger frontier models.
Key insights
- Discussion-based orchestration methods, effective for frontier LLMs, often fail or degrade SLM performance due to "groupthink" where models reinforce incorrect reasoning.
- Estimating confidence through an SLM's internal consistency across multiple sampled outputs is an effective, non-discursive strategy for selecting accurate answers.
- Combining complementary SLMs via SLM-MUX can collectively achieve reasoning capabilities comparable to or better than significantly larger, monolithic frontier models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.