Harness Engineering 本质是即时上下文学习
Original title: Harness Engineering Is Just-in-Time In-Context Learning
The title and summary in the selected language are awaiting translation.
模型能力再强，其"好工作"的先验也未必与你的标准一致，因此始终需要围绕模型整理环境，让它在非功能性要求上偏向你或组织认可的选择。模型变好不会让这种上下文学习（ICL）需求消失。所谓 Harness Engineering，本质是一组技巧，为 ICL 提供即时机会，在不束缚推理模型的前提下对齐模型行为。
While the generally capable models will always become more capable, there’s nothing that requires the models’ priors around what good looks like to align with your own. It follows that it will always be required to curate the environment around the model such that it spikes in the direction of coherent choices for the nonfunctional requirements you or your organization will accept as good work.
No amount of making the model better will obsolete this need for in-context learning (ICL).
All of “harness engineering” is essentially a set of tricks to provide JIT opportunities for ICL to align model behavior with what good looks like for you without unduly restraining these reasoning models.
Source: Ryan Lopopolo · hyperbo.la