Reproduce “Not All Reasoning Generalizes: A Case Study on Learning Generalizable Reasoning Strategies in RLVR” 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
Ziqi Pang, Yu-Xiong Wang
The authors test whether bijective-task training helps recover inputs for string functions under shifts. They compare Qwen2.5-7B-Instruct trained on non-bijective tasks alone or with a 1:1 bijective-task mix. On synthetic Length Change tasks, the mix raises accuracy from 5.3% to 21.7% and derivation share from 48.3% to 93.3%, without establishing derivation as the sole cause or transfer beyond these tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.