Reproduce “Best-of-Depth: Learning Exploratory Dynamics in Looped Transformers” 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
Shota Takashiro, Takeshi Kojima, Masanori Koyama, Kohei Hayashi, Yusuke Iwasawa
The paper introduces Best-of-Depth training for looped Transformers, which repeatedly update one hidden state and produce readouts at multiple depths. For each input and valid target, the method trains from the readout with the lowest prediction loss. On 1,000 multi-solution N-Queens problems, it reaches 91.03% mean coverage through depth 64, versus 45.75% for final-depth training, measuring solutions along a trajectory.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.