One LLM agent framework, unchanged inside, beats baselines on spiking classification and PDE solving across five tasks
Lead
A three-role LLM agent framework that rewrites only a designated functional block while leaving the network structure fixed outperforms baselines on spiking neural network classification and physics-informed neural network solutions of differential equations.
Story
The same multi-agent LLM framework beats the baseline on all five tasks without changing its internal structure. Selecting these key components was previously done manually, with no unified automatic tuning procedure. MNIST accuracy rose from 79.88% to 92.83% and digits from 90.83% to 96.67%; the single-soliton relative error fell about 2491-fold, the conservation error in the Gaussian-perturbed GNLSE case fell about 325-fold, and the KS mass drift fell 14-fold.
The framework splits hypothesis generation, code implementation, and experiment documentation into three LLM roles, closed by a sandbox and an experiment log. Existing systems either optimize the entire ML script or build a network from a textual description, rather than modifying only a designated block of a fixed architecture. The Analyst writes a technical specification, the Coder implements it under a software contract, the sandbox runs a functional test before evaluation, and the Recorder appends a JSON record that serves as external memory.
What to watch
The framework targets settings with a fixed architecture where only a designated functional block may be modified, and its intended readers are researchers and engineers who want to tune specialized neural network modules automatically. The authors propose next steps: increasing PINN problem complexity, jointly tuning several blocks, and building models that account for neuromorphic system constraints.
The search space is restricted to a fixed catalogue of editable blocks, so structural changes outside it cannot be discovered; the experiment log is a flat chronological record without a hypothesis dependency graph; all runs are capped by an iteration budget, so reported values are lower bounds; the most expensive configurations report a single seed; two models depend on commercial APIs whose availability and pricing may change; and no head-to-head comparison with LLM-driven NAS baselines was performed.