Bing Zhu

LLMs can teach themselves new tricks: a simple self-improvement loop, "Ratchet," lets a frozen LLM agent significantly boost its coding performance by autonomously managing its own library of natural-language skills.

End-to-end prompt optimization is often a waste of time and money, succeeding only when coaxing models into specific output formats they're already capable of.

Expert-written rules for coding agents are often useless or even harmful, with random constraints working just as well and negative constraints outperforming positive directives.