Reproduce “Intra-request branch orchestration for efficient LLM 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
Weifan Jiang, Rana Shahout, Yilun Du, Michael Mitzenmacher, Minlan Yu
The DUCHESS system offers a method for efficiently serving Large Language Model reasoning requests through intra-request branch orchestration and inter-request complexity-aware scheduling. It reduced token usage by 42-63% and mean latency by 57-81% on benchmarks like GSM8K, MMLU, and MATH, while maintaining or improving accuracy.
Key insights
- Intermediate LLM layer activations (e.g., from the last decoded token) can serve as lightweight, effective signals for real-time correctness prediction of reasoning paths with minimal overhead.
- Dynamic orchestration strategies like early termination of successful branches and selective branch-out from promising paths enable substantial reductions in token usage and latency while avoiding accuracy degradation.
- Combining fine-grained, reasoning-aware intra-request controls with inter-request complexity-aware scheduling can further optimize overall system throughput and reduce end-to-end latency in concurrent LLM serving.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.