Spectral analysis predicts AI debate convergence in advance

A new paper computes when a multi-agent AI debate will converge before the debate starts, using spectral analysis of interaction traces, per arXiv 2608.05956.

Most multi-agent voting and debate setups pick a round budget and hope it's enough. This one gets a computable deadline that tracked real convergence at a 0.93 correlation across 24 configurations, correctly bounding it 96% of the time.

It treats the whole agent collective as one nonlinear dynamical system and estimates its Koopman transfer operator from the traces. The convergence deadline comes off the sub-dominant eigenvalue. That eigenvalue's eigenvector names the coherent factions inside the debate.

Eight of the system's 32 spectral coordinates preserved the final decision at 99.7% fidelity. A certificate trained on just 15 debates held on all 60 held-out debates it was tested against. The whole estimation runs in minutes on a CPU, cheap enough to sit inside a production voting setup as a live round-budget check.

The pruning is the sharper result. Eight of 32 coordinates carrying the full decision means most of what a multi-agent debate computes is discardable motion. A spectral cutoff could trim rounds without losing the answer, something shown so far only across 24 configurations.