Reproduce “Scaling Shapley Energy Attribution to Large LLM Batches” 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

Qi Luo, Ziwen Wang, Kunlin Li, Changxi Liu, Yun Chen

StepSplit attributes energy to requests in batched language-model serving. It estimates each request’s own work and divides each forward pass’s shared cost among requests present; for continuous batching, it replays the scheduler on a CPU to correct schedule changes. Across 529,513 runs, normalized L1 errors were 0.4–4.3% for static batches and 0.5–6.4% for continuous batches against measured Shapley references.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.