Users state runtime constraints; multi-agent LLMs plus a contract knowledge base generate compact custom LLM inference engines that run under those constraints.

Summary. The manuscript proposes MetaInfer, an “LLM-as-Compiler” system in which users specify only runtime constraints for an LLM inference program; an LLM multi-agent collaboration system coupled with a contract knowledge base (CKB) then automatically generates a compact customized inference framework intended to satisfy those constraints. The abstract claims evaluation along three axes: the effect of source-code reference, runtime behavior and performance of engines generated under a zero-reference constraint on CKB-covered targets, and knowledge-base evolution for new models and platforms. It asserts that generation constraints, validation feedback, and knowledge consolidation form a continuous closed loop that yields runnable customized inference solutions from explicit knowledge, with code released publicly.

Significance. If substantiated, the result would address a real and growing maintenance burden in LLM serving stacks—expanding model families, hardware, quantization, parallelization, and kernels—by replacing heavily abstracted general-purpose frameworks with constraint-driven, compact generated engines. Explicit strengths claimed in the abstract include a closed generation–validation–consolidation loop, zero-reference generation on CKB-covered targets, and a public repository. Those strengths would be material if the full methods and results show that generated engines are not only plausible but correct and performant under stated runtime constraints. On the abstract alone, significance remains conditional on evidence that is not present in the provided text.