Scaling Closed-Loop Feature Channel Configuration with LLMs

Short summary

This study scales closed-loop LLM-based neural network channel-configuration search to 250 candidates per fine-tuning cycle, analyzing 2,000 generated candidates across 8 cycles with 462 verified CIFAR-100 evaluations. Per-cycle mean accuracy shows a positive linear trend (p=0.043), with best accuracy improving from 0.3144 to 0.3676 while reducing parameters from 166.5M to 11.8M. The larger sample reveals architectural regularities: 41.8% of candidates use non-power-of-two widths, and strong models share structured channel-allocation patterns with moderate early widths and expanded middle or later blocks.

• •Scaled LLM-based channel search to 250 candidates per cycle, confirming positive accuracy trend across 8 cycles
• •Best model achieved 0.3676 accuracy with 11.8M parameters vs early model's 0.3144 with 166.5M parameters
• •Strong models share structured patterns: moderate early widths with expanded middle or later blocks

Generated with AI, which can make mistakes.