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

Researchers say RLVR weight updates in LLMs follow highly low-rank, predictable trajectories.

They propose RELEX, which estimates a rank-1 direction from early training steps and extrapolates later checkpoints with little or no extra training.

On multiple models and benchmarks, RELEX matches or even beats full RLVR training.

If validated, the method could sharply reduce the cost and time of improving LLM reasoning models.