Modal reports off-policy GRPO corrections stabilize training on SWE-bench Pro V2

Modal says it stabilized GRPO training on mixed samplers by adding masked importance sampling, score centering, and top-p sampling with mask replay to standard GRPO. Modal reports that training on the mixed samplers became stable as evaluated on SWE-bench Pro V2.

The only source is a post on Modal's X account. The post gives no figures for the stability measurement and does not describe a comparison against standard GRPO results.