KG-CFR: New method cuts AI argument quality loss by 95% during crises

Multi-agent debate systems improve LLM performance but suffer from logic degradation, argument repetition, and role drift under sustained perturbations. To address this, researchers Masłowski and Chudziak propose Knowledge-Grounded Counterfactual Reasoning (KG-CFR), which enforces a strict separation between a private retrieval-augmented planning buffer and a public execution layer. This decoupling prevents identity loss and maintains process fidelity during long-horizon exchanges.

In a 1v1v1 Dynamic Resource Allocation under Uncertainty (DRAU) environment with 270 stochastic shock trajectories, KG-CFR prevented judge-detected critical degradation (quality shift ≥ -0.20) in over 95% of perturbed runs, raising overall argument quality from 0.694 to 0.822. Ablation experiments show doctrinal grounding is as important as prospective planning. The system also reduces semantic looping by preserving agent consistency with the original plan, offering a practical blueprint for building stable, multi-agent reasoning pipelines.