You Only Need Minimal RLVR Training: Extrapolating LLMs via Rank-1 Trajectories

Researchers found RLVR updates in LLMs are strongly low-rank, with most of the effect captured by a nearly linear rank-1 trajectory over training.

Using this pattern, they built RELEX, which fits an early training direction with linear regression and extrapolates later checkpoints without a learned predictor.

RELEX matches or outperforms full RLVR training on benchmarks while using only about 15% of the compute.

The result suggests a short training window may recover much of RLVR’s benefit and offers a new view into LLM reasoning dynamics.