Russian researchers test multi-agent LLM system for neural-network tuning

Researchers at the Kurchatov Institute, MEPhI and the Moscow Institute of Physics and Technology published a study on 9 October using multi-agent language models to tune selected neural-network components. The framework separates hypothesis generation, coding and experiment logging while leaving the network’s overall architecture fixed. The collaboration also includes a researcher from Harbin Institute of Technology in China.

Tests covered spiking-neural-network classification on MNIST and Digits, plus differential-equation problems using physics-informed neural networks. The paper reports improvements on some measures, but notes that the single-soliton test with an analytical solution allowed the known formula to be supplied and therefore does not measure a complex independent search. The results are experimental and limited to the tasks and configurations studied.