Reproduce “Budget Internalization: Fixed-Budget RLVR Trades Test-Time Scaling for In-Budget Accuracy” 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

Ilia Zavidnyi, Alexander Kovrigin, Arina Puchkova, Kirill Gelvan, Sergey Titov

The authors evaluate fixed-cap RLVR-trained models across multiple token budgets. For Qwen3.5-4B trained with a 4k-token rollout cap, AIME accuracy averaged over three sampled answers is 34.8 points above the base model at 4k, but 30.2 points below it at 64k. The trained model’s output length and accuracy largely saturate above its training budget; evidence centers on this comparison.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.