Does Scaling Toward AGI Depend on What System-1 Lets System-2 Execute?

Does Scaling Toward AGI Depend on What System-1 Lets System-2 Execute?. By Qi Wang and twelve others, from Tsinghua University. Published on the twenty-ninth of October.

Progress toward artificial general intelligence has largely come from scaling one resource-intensive executive, which the authors call System-2. Scaling it has meant larger models, more data and compute, longer reasoning, and more agents. The question here is whether the next stage might come instead from co-designing System-2 with a System-1: a lightweight, semantically aware predictor that evaluates candidate interactions before they are executed, so that System-2 acts only on the ones that justify their cost.

The object of study is conditional interaction, a loop of four steps: predict, gate, execute, update. System-1 estimates each candidate's utility. System-2 executes the candidates admitted by an adaptive threshold. The outcomes then update System-1. The loop is meant to span training, test-time reasoning, adaptive decision-making, and agentic tool use, with utility and cost defined by the application it sits in.

The answer the authors give is a qualified yes, resting on two results. The first concerns compute. Under ideal screening and equal execution costs, execution compute falls by a factor approaching one over rho sub u, where rho sub u is the density of utility-relevant interactions. The second is a bound: selection regret is bounded by the predictor's error and by the utility margin.

Read together, the two results say where a saving comes from and what limits it. When only a small share of candidate interactions carry utility, rho sub u is small, and a predictor that screens out the rest leaves System-2 with correspondingly less to execute. The regret bound then ties what is lost in the screening to how accurate the predictor is and how wide the utility margin is. From there the authors reach their central hypothesis: that rho sub u is a measurable coordinate of scaling efficiency, and that informative compute describes scaling better than total compute does.

The headline factor holds only under its stated conditions. Screening must be ideal, and execution costs equal. When costs differ, the density is cost-weighted. And three things reduce the saving: prediction errors, exploration, and distribution shift.

The final step is framed as a hypothesis that the results motivate, not as something they establish. What it proposes is a different axis for talking about scale: not how much compute is spent, but how much of it lands on interactions that repay the cost of executing them.