This paper introduces AaLLM, an open-source system that uses Large Language Models (LLMs) to go from a set of user specifications to a working circuit netlist. Unlike prior tools that only pick component sizes or only pick circuit structure, AaLLM combines both steps in one pipeline. It aims to generate novel circuit topologies and then choose component values so the circuit meets the requested performance.

To make design decisions grounded in engineering knowledge, AaLLM builds a searchable knowledge base from textbooks and research papers. A Retrieval-Augmented Generation (RAG) module supplies relevant design facts during both topology selection and sizing. For topology generation the authors fine-tuned a Sequence-to-Sequence (Seq2Seq) model so it can translate specifications into valid circuit netlists and recombine connections to produce new, potentially useful topologies. Candidate topologies are screened for feasibility by the RAG module and then passed to a sizing loop that uses SPICE (Simulation Program with Integrated Circuit Emphasis) to measure actual circuit behavior.

Circuit sizing is handled by a three-agent loop: a Designer proposes component values, a Critic checks those choices, and an Evaluator arbitrates between them to reduce the number of sizing iterations. The sizing process uses simulation feedback and follows a curriculum-style ordering of objectives so the system improves one performance target at a time without undoing previously met goals.

On a test set that includes operational amplifiers and filters, AaLLM-produced novel topologies achieved figures of merit comparable to known designs and in some cases up to three times higher. Compared to state-of-the-art multi-agent LLM pipelines, the authors report 3×–4.5× fewer SPICE calls during inference and about a 40× reduction in wall-clock time. The tri-agent sizing loop met the target specifications in 91.6% of cases on their benchmark.