Could System-1 Decide What System-2 Executes on the Road to AGI?

Introduction

The dominant route toward more capable AI systems has been to scale the executive machinery: larger models, more data, more compute, longer reasoning traces, and additional agents. This strategy can improve capability, but it also increases the cost of inference. A paper titled Does Scaling Toward AGI Depend on What System-1 Lets System-2 Execute? asks whether the next efficiency gains could come from controlling which computations are allowed to reach the expensive reasoning stage.

A lightweight gate for expensive reasoning

The study uses the familiar distinction between System-1 and System-2 in a computational design sense. System-2 is the resource-intensive executive. It performs the reasoning step, decision, or tool interaction that may produce value. System-1 is a lighter, semantically aware predictor that evaluates candidate interactions before they are executed.

The proposed architecture follows a four-stage loop:

• Predict: estimate the likely utility and cost of each candidate interaction;
• Gate: compare candidates with an adaptive threshold and admit only selected ones;
• Execute: let System-2 perform the interactions that pass the gate;
• Update: use the observed outcomes to improve future predictions.

The framework is intentionally broad. It can be applied during training, at test-time reasoning, in adaptive decision-making, or in agentic tool use. The relevant notion of utility is application-dependent. In a tool-using agent, for example, a candidate call may be valuable if it is likely to change the answer or resolve uncertainty. In a reasoning process, an intermediate step may be skipped when its expected contribution is low. The paper does not claim that one fixed utility function works across all of these settings.

From total compute to useful interactions

The paper’s central quantity is the density of utility-relevant interactions, denoted by ρ_u. It describes the share of candidate interactions that actually matter for the task, with a cost-weighted version used when interactions have different execution costs. If screening were ideal and every execution cost were equal, the compute handled by System-2 could approach a reduction factor of 1/ρ_u relative to executing everything.

That result is a theoretical efficiency picture, not an unconditional promise for deployed systems. A predictor can reject valuable actions or admit unproductive ones. The system may also need to explore uncertain options, and the deployment distribution may differ from the data used to train System-1. All three effects reduce the possible savings. The paper also relates selection regret to predictor error and the utility margin, highlighting why the separation between worthwhile and unworthwhile candidates matters.

Why the proposal matters

The study’s main contribution is a way to rethink scaling efficiency. Instead of treating all executed computation as equally informative, it suggests measuring how much of that computation is actually relevant to the objective. In this view, ρ_u could become a measurable coordinate for comparing systems and identifying opportunities for optimization.

For inference infrastructure and AI agents, the implication is a co-design problem. System-1 must be cheap enough that screening does not consume the savings, yet reliable enough to preserve rare, high-value actions. Practical evaluation would need to examine task-specific utility, the cost of false rejection, exploration policies, and distribution shift. The paper therefore offers a qualified hypothesis: future progress may come not only from scaling System-2, but from making the boundary around its execution more selective.

Source: Hugging Face Daily Papers