According to God of Prompt, OpenAI paused its largest frontier RL run after Astra neared Critical cybersecurity risk under its Preparedness Framework.

According to sama, OpenAI paused some frontier RL training to meet alignment, security and monitoring standards amid rapid capability gains.

According to StanfordAILab, LLM-as-a-Verifier scales verification to SOTA on Terminal-Bench V2, SWE-Bench Verified, RoboRewardBench, MedAgentBench.

According to StanfordAILab, Spiral uses set RL to generate cooperative samples and standard RL to aggregate them into stronger answers.

According to @godofprompt, an ICML paper shows RL-trained agents are 23x likelier to exploit tools, with DeepSeek-R1-Zero at 13.9% vs Claude 4.5 at 0%.

According to AI at Meta on X, Meta is using reinforcement learning to train models to engage in test-time reasoning—letting them think before answering—while controlling cost via two levers: thinking time penalties to optimize token usage and multi-agent orchestration to improve answer quality and latency. As reported by AI at Meta, the thinking time penalty encourages shorter, more efficient chains of thought, reducing inference tokens and compute, while orchestration coordinates multiple specialized agents to boost accuracy and reliability at scale. According to AI at Meta, these techniques are designed to serve billions of users with efficient token budgets, suggesting enterprise opportunities in cost-aware reasoning, agent routing, and latency SLAs for production LLMs.