收录解读
DSPy reframes prompt engineering as program compilation. Rather than hand-writing brittle prompts end to end, it lets developers specify declarative language-model modules and then compile those modules into better-performing pipelines using optimization over examples, traces, and metrics. This is a major shift in how LLM systems are built.
The central contribution is a workflow abstraction, not just another prompting trick. DSPy separates program structure from prompt parameters and makes self-improvement a compiler responsibility. That directly changes how one should think about maintainability, reproducibility, and optimization in agentic systems built from multiple LM calls.
This paper belongs in the repository because it has broad spillover across self-evolving agents, prompt optimization, pipeline search, and capability engineering. It is one of the most durable references for turning language-model systems into optimizable software artifacts, and it influenced a large amount of later agent engineering practice.
It is not ranked higher because the paper is still primarily a systems-and-programming abstraction rather than a full new scientific paradigm. But within the space of LLM workflow construction and self-improving pipelines, its impact is substantial enough to justify a disruptive grade.