Answers are grounded in this paper and cite their sources.
A recent study tracks the thirteen released reinforcement-learning-with-verifiable-rewards (RLVR) checkpoints of OLMo-2-0425-1B-Instruct on GSM8K, reports that sampled accuracy rises from step 200 to step 2,600 while linear probes for injected arithmetic errors stay flat, and concludes that RLVR sharpens which solution the model leads with rather than what it represents. Those checkpoints record a second RLVR stage that starts from OLMo-2-0425-1B-RLVR1, itself an RLVR stage trained from the DPO model; neither model is measured. We re-measure with the study's prompt, parser and injection code. On all 1,319 test problems greedy accuracy is 51.8% for DPO, 61.4% for RLVR1 and 62.4% at step 2,600: the measured stage changes it by +1.0 points (95% CI -1.1 to +3.1), and RLVR1's own checkpoints show the +9.6-point gain over DPO accruing gradually across the first stage. Within the measured stage, sampled single-try accuracy on 200 problems rises +4.5 points, coverage in 32 samples does not rise, and samples agree with the model's own greedy answer 8.1 points more often, so the study's sampled rise is sharpening of an already trained model. Probes fit on the DPO model and applied unchanged to later checkpoints, on DPO-generated text, detect injected errors no worse than refit probes by more than 0.02 AUROC (upper paired 95% bound), and at a middle layer judge each checkpoint's own answers as well as refit probes. The sharpening reading of the measured stage stands and the verifier transfer the study predicts holds, but most of the gain comes before its checkpoints. Generations, probe scores, code and a script that recomputes every number are released. Preprint, not peer reviewed. Dataset: RLVR Checkpoint Re-measurement v1.0 (see Related works). Original study: Bangera and Gupta, arXiv:2611.02574.
No takes yet. Share an insight, caveat, or question.
Guerin Lee Green (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: