Reproduce “FunDM: Cheap & Label-Free Decision Models from Frozen Language Models” 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
Ujjawal K. Panchal
FunDM trains a small decision head to score candidate answers using a frozen pretrained model. The model reads the question and all options in one causal prompt; the head is fitted to match its option probabilities, without correct-answer labels. On Qwen3-8B, the 24-layer FunDM+1T head reaches 0.750 mean accuracy across eight benchmarks, retaining 99.2% of full-model accuracy in this experiment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.