Reproduce “Beam Model Technical Report” 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
Reflection Team, Adam Abdulhamid, Osama Abuelsorour, Yasmin Alameddine, Carlo Alfano, Duarte Alves, Brandon Amos, Selvaraj Anandaraj and 222 more
Beam is a sparse mixture-of-experts model trained to balance coding-task success against response length, with a user-set reasoning budget. Its reinforcement learning training rewards successful solutions and penalizes unnecessary tokens. On DeepSWE, pass@1 reached 35% at effort 512 by the 79%-of-run checkpoint and 41% at the final checkpoint; a harder task set was added between them.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.