Architectural Decoupling in Multi-Agent Debates Cuts Degradation by 95%

Over 270 crisis simulation runs, a new architecture called KG-CFR kept multi-agent debates from collapsing under shock more than 95% of the time—raising argument quality scores from 0.694 to 0.822.

The Problem: Reactive Multi-Agent Debates Fall Apart Under Pressure

Current multi-agent debate frameworks optimize for final accuracy, not process stability. Throw sustained perturbations at them—long-horizon exchanges with stochastic shocks—and you get logic degradation, argument repetition, role drift. The paper's authors call this "identity loss," and it's a structural weakness baked into architectures that blend thinking and speaking into one reactive loop.

Standard debate setups converge on answers but can't maintain coherence under fire. That failure mode matters for any deployed system that must argue through uncertainty, like resource allocation during a crisis.

The Fix: Separate Planning from Execution with Counterfactual Grounding

KG-CFR enforces a strict separation of concerns. A private, retrieval-augmented planning buffer holds the agent's grounded knowledge and counterfactual reasoning. A separate public execution layer handles what gets said. Thought and speech decoupled.

The key mechanism: counterfactual reasoning grounded in retrieved knowledge before every public utterance. That pre-commitment to a plan prevents the agent from drifting when shocked. The authors evaluate in Dynamic Resource Allocation under Uncertainty (DRAU), a 1v1v1 environment designed to create diversity beyond standard debate settings.

What the Numbers Say: 95% Degradation Prevention and Beyond

Over 270 completely factorial crisis simulation trajectories with stochastic environmental shocks, KG-CFR prevented judge-detected critical post-shock degradation (defined as a quality shift $\Delta \le -0.20$) in more than 95% of perturbed runs. Overall argument quality jumped from 0.694 to 0.822.

Ablation experiments reveal something subtler: proper doctrinal grounding matters as much as prospective planning. The architecture also reduced semantic looping and preserved agent consistency with the original plan. Custom vector metrics for discourse divergence and plan-execution alignment provide directionally consistent evidence of operational stability.

KG-CFR suggests that debate frameworks should treat identity preservation as a first-class design constraint, not an emergent property.

Source: Decoupling Thought from Speech: Knowledge-Grounded Counterfactual Reasoning for Resilient Multi-Agent Argumentation
Domain: arxiv.org