According to AI at Meta on X, Meta’s new reinforcement learning (RL) training stack delivers smooth, predictable performance scaling, with log-linear improvements in pass@1 and pass@16 as compute increases. As reported by AI at Meta, the approach addresses common large-scale RL instability and demonstrates consistent capability gains under higher compute budgets. According to AI at Meta, these metrics indicate more reliable code or reasoning task success rates, translating into clearer pathways to productionizing RL for model upgrades and cost planning. For AI builders, the business impact includes more forecastable model iteration cycles, better return on GPU spend, and reduced variance in outcomes when scaling RL fine-tuning, as reported by AI at Meta.