Stochastic CHAOS: Why Deterministic Inference Is Fatal to Large Language Models, and Distributional Variability Is the Heartbeat of Artificial Cognition

In response to Thinking Machines Lab's argument that deterministic inference ensures reproducibility, an arXiv paper proposes Stochastic CHAOS, treating distributional variability as a measurable and controllable signal. The paper claims that single-sample deterministic evaluation simultaneously underestimates capability and fragility, that the phase-transition-like behavior of emergent capabilities disappears under greedy decoding, and that accuracy declines when multi-path reasoning is compressed onto a deterministic backbone. Deterministic evaluation also masks rare dangerous behaviors that appear only under multi-sample evaluation, thereby underestimating safety risk.