Drift No More: Spectral Locking for Long-Horizon Open-Weight Multi-Agent AI

ATLAS mitigates task, role, and state drift in long-horizon multi-agent systems by steering hidden states toward shared high-coherence eigendirections. This training-free intervention uses a positive-semidefinite anchor kernel and low-rank Gram decomposition to align representations before token generation, addressing the observed correlation between subspace collapses in drifted turns.