Predictive Maps of Multi-Agent Reasoning: Spectral Diagnostics Predict LLM Communication Topology Failures

arXiv·medium signal

Introduces a successor-representation diagnostic for multi-agent LLM graphs that connects three spectral quantities to three failure modes: spectral radius predicts cumulative error (r_s=-1.0), condition number predicts perturbation robustness (r_s=1.0), spectral gap partially predicts consensus (r_s=0.5). Validated on Qwen2.5-7B with chain/star/mesh topologies across 100 trials.