Best-of-Majority Sets New Standard for Pass@k LLM Scaling

Best‑of‑Majority (BoM) filters LLM outputs by frequency then ranks them with a reward model, achieving lower regret and beating majority voting and Best‑of‑N on math reasoning tasks. Read more: getnews.me/best-of-majority-sets-ne... #llm #passatk

AP2O Improves LLM Code Generation by Correcting Errors Type‑by‑Type

AP2O, a new training method for LLM code generation, improves pass@k scores by up to 3% across models from 0.5 B to 34 B parameters. The research was posted on 1 Oct 2025. getnews.me/ap2o-improves-llm-code-g... #ap2o #llmcode #passatk