Reproduce “Learning on the Job: Test-Time Curricula for Targeted Reinforcement Learning” 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
Jonas Hübotter, Leander Diaz-Bone, Ido Hakimi, Andreas Krause, Moritz Hardt
Researchers from ETH Zürich and the Max Planck Institute for Intelligent Systems introduce Test-Time Curricula for Reinforcement Learning (TTC-RL), enabling large language models to continually refine their reasoning abilities for specific tasks at test-time. This method, which uses self-curated training experiences, demonstrates substantial performance gains (e.g., 1.8x on AIME25, 2.1x on CodeElo) and raises the `pass@k` ceiling, confirming genuine reasoning improvements.
Key insights
- Continual, targeted RL training at test-time enables LLMs to effectively compress learned reasoning strategies into their model weights, alleviating context window limitations.
- Self-curated curricula driven by task informativeness and achievability automate the training data selection process, reducing reliance on human data curation.
- TTC-RL leads to highly specialized models that genuinely improve their underlying reasoning capabilities, as evidenced by a novel 'latent improvement' metric, rather than merely learning output formats.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.