Reproduce “Prior Action Control Causally Alters Language-Model Plasticity” 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
Marc-Andre Schulz
The authors test whether reward–answer contingency during training changes models’ later learning. They compare copies receiving matched rewards, with rewards following one copy’s answers but not the other’s, then measure learning on a new task. In a three-answer, three-step task, contingent copies gained 8.67 percentage points more across updates 1–8; evidence is limited to synthetic tasks and two Qwen3 sizes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.