Gradient Magnitude Asymmetry as a Structural Honesty Mechanism in Multi-Agent World Model Communication

We investigate whether strategic deception emerges in multi-agent systems that coordinate through shared world model latent representations. Classical cheap talk theory predicts that diverging interests degrade message fidelity, yet we show this does not occur under standard dual-objective training. We propose gradient magnitude asymmetry (GMA) as the responsible mechanism: dense, per-timestep accuracy gradients systematically overpower sparse policy gradients from strategic reward. In 60 training runs across a 3×3 factorial design (reward divergence × verification) with ablations, totaling 467 GPU-hours, we find that (1) incentive divergence produces no significant change in belief-message divergence (p=0.86), (2) sweeping the accuracy weight β yields a monotonic increase in distortion (r=−0.97), and (3) partial verification independently reduces distortion by 38%. These results suggest that representation integrity in shared world models may be a designable property governed by the accuracy-to-strategic gradient ratio.