Reproduce “Principled Top-$k$ Selection for Language Models with Hybrid Gradients” 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
Xuchen Gong, Junfei Sun, Tian Li
The authors introduce selector training for choosing k distinct items from m candidates in retrieval-augmented generation and mixture-of-experts systems. The method samples an unordered set and combines supervised gradients for selected-item weights with a policy-gradient signal about set choice. On Natural Questions, it achieves 33.6% exact-match accuracy versus 32.4% for supervised-only with Qwen-2.5-1.5B, m=32, and k=3; larger-scale performance remains unevaluated.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.