SEPAL: Separated Expert Pairs with Answer-Level Fusion for Reliable LLM Collaboration
SEPAL creates reliable LLM collaboration by assigning three private Actor-Critic teams to distinct roles: Direct reasoning, Evidence grounding, and Verification. Each team revises its answer four times in isolation using its specific Critic, preventing error propagation. Final answers are fused via a majority vote or a fixed fallback rule, improving accuracy by 1.81 percentage points over single pairs.