Reproduce “Late Context Masking: Accuracy and Fidelity to Context-Aware Decoding” 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
Hanyu Yang
The authors introduce the late fork, which shares lower model layers and reruns the final 16 layers with context attention masked as a cheaper alternative to context-aware decoding (CAD). On MemoTrap, Qwen3-4B-Base reached 64.1% accuracy versus CAD’s 60.5% at the same fixed contrast weight. This advantage is task- and setting-specific; the fork did not closely reproduce CAD’s token choices.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.