Llama 3.2 3B Modular Swarm Scaling (64 vs 256 vs 1,600 Synthetic Traces on 2xT4)
Measured completion-only SFT cross-entropy loss and Verifier DPO preference margin (15 PEFT LoRA adapters, $0 spend)
Specialist vs Monolithic SFT Loss (Lower = Better)
0.0
0.5
1.0
64 traces
256 traces
1,600 traces
Monolithic: 1.01->0.43
Retriever: 0.77->0.35
Decomposer: 0.43->0.36
Verifier DPO Preference Margin (Chosen vs Tombstone)
+13.64
64 pairs
+20.12
256 pairs
+24.54
1,600 pairs
DPO loss: 0.0832 (64) -> 0.0702 (256) -> 0.0654 (1,600) | 100% batch win rate