一个开发者分享了他对GPT-5.6编排和验证的实践心得，用不同模型做验证能减少奖励破解，很实用。
作者对GPT-5.6在编排与验证任务中的表现感到好奇。他通常使用不同的模型进行验证，并认为这种方式能带来更好的性能并减少奖励破解（reward hacking）导致的异常行为。目前尚缺乏适合这种测试时计算（test-time compute）范式的基准测试，作者通过多个实际项目来评估这些模式的效果。
What I am curious to find out is whether GPT-5.6 is also great at orchestration and verification. Fo...
What I am curious to find out is whether GPT-5.6 is also great at orchestration and verification. For verification, I normally prefer to use a different model in this workflow. It has resulted in better performance and reduced weird behaviors that arise from reward hacking. I really don't have a way to benchmark (much needed in this new test-time compute paradigm) it, but I have a bunch of projects that I apply these patterns to, and that's how I assess it for now. 💬 0 🔄 0 ❤️ 0 👀 753 ⚡