Neural network tuning stops rewriting the whole script: three LLM roles edit only chosen blocks, lifting MNIST accuracy from 80% to 93%

One multi-agent framework can modify only a pre-specified functional block of a neural network instead of rebuilding the network, turning component selection from manual choice into a reproducible automated search. Such key components, including input and output representations, decoding rules, and loss-term balance, were previously selected by experts for each problem with no universal selection rule; existing agent systems either optimize the whole machine learning script or construct a network from a text description. The framework runs on five tasks spanning spiking-network classification and physics-informed neural network solving, and each task is compared with an identity-mapping baseline from the same run.

The framework splits hypothesis generation, code implementation, and experiment documentation into three large language model roles, closing the iteration loop with a sandbox and an external log. The analyst proposes one hypothesis from the task prompt and current experiment log and writes it as a fixed-structure technical specification, the coder translates that specification into Python satisfying a software contract, and the recorder packs specification, code, and evaluation result into a structured JSON record. Generated modules pass a short functional test before evaluation, runtime errors return to the coder for correction within at most 10 attempts, and the metric from a successful run enters the log that feeds the next analyst call.

In MNIST decoder search, the detailed prompt lifts weak analyst models from baseline level toward the strong model, but the coverage map is the component that makes the lift possible for the weakest one. Under the simple prompt, Perplexity Sonar and Qwen3-30B-A3B-Instruct-2507 produced no decoder above the 79.88% baseline across 250 candidates in each of ten runs; under the detailed prompt Sonar exceeded 90% in all ten runs with a maximum of 92.70%, while Qwen reached 89.16% without crossing 90%. Each configuration ran 10 complete searches, paired comparisons used the same trained spiking network instance and identical output spike data, and the two-sided exact Wilcoxon signed-rank test stayed significant for all six tests after Holm correction.

The framework targets tuning scenarios where the architecture and training procedure are fixed and only designated functional blocks may change; users supply the task prompt, software contract, target metric, and baseline. Validated scope covers spiking-network decoders on MNIST and scikit-learn digits, plus boundary-value problems for the generalized nonlinear Schrödinger equation and the Kuramoto–Sivashinsky equation. The authors propose next steps including more complex physics-informed neural network problems such as two-soliton interaction, joint tuning of several blocks, and models that account for neuromorphic system constraints.

The single-soliton problem has an analytical solution, so the agent can embed the known soliton structure into its representations and reconstruction, and the error reduction there is not equivalent to search ability on problems without a known solution. The most expensive configuration on the soliton with Gaussian perturbation reports a single seed, and the ten retrainings cover the conservation error but not the joint criterion. The Kuramoto–Sivashinsky problem was evaluated only with Claude Opus 4.6, without comparison across analyst models. The search space is limited to a fixed catalog of editable blocks, and the experiment log is a chronological JSON record without a hypothesis dependency graph. Two analyst models are accessed through commercial APIs whose availability and pricing may change.