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

**Authors:** Qi Wang, Yun Qu, Lizhou Cai, Heming Zou, Haoang Chi, Yiqin Lv, Xiang Li, Haiyan Feng, Mengyu Ge, Yize Geng, Yuhang Jiang, Yixiu Mao, Xiangyang Ji  
**Lab:** Tsinghua University · **Published:** 28 Oct 2026 · **Upvotes:** 56 · **Level:** Intermediate  
**Links:** [arXiv](https://arxiv.org/abs/2610.39497) · [PDF](https://arxiv.org/pdf/2610.39497) · [This page](https://omen-bit.github.io/ai-paper-daily/paper/2610.39497.html)

## Curator's notes

**TL;DR:** Co-design System-1 predictor with System-2 to screen low‑utility actions, reducing execution compute.

- **What's new:** Introduces a predict‑gate‑execute‑update loop and formal bounds linking utility density to compute savings, unlike prior work focusing only on scaling System-2.
- **Key result:** Execution compute falls by a factor approaching 1/ρ_u under ideal screening, where ρ_u is the density of utility‑relevant interactions.
- **Why it matters:** Developers can cut inference costs by letting a cheap model filter tasks before invoking a large model.

## Abstract

Progress toward artificial general intelligence (AGI) has largely relied on scaling a resource-intensive executive, System-2, through larger models, more data and compute, longer reasoning, and more agents. We ask whether the next stage may come instead from co-designing it with System-1, a lightweight, semantically aware predictor that evaluates candidate interactions before execution, so that System-2 acts only on those worth their cost. Our answer is a qualified yes. We study conditional interaction, a predict--gate--execute--update loop: System-1 estimates each candidate's utility, System-2 executes those admitted by an adaptive threshold, and the outcomes update System-1. The loop spans training, test-time reasoning, adaptive decision-making, and agentic tool use, with utility and cost defined by the application. Under ideal screening and equal execution costs, execution compute falls by a factor approaching 1/ρ_u, where ρ_u is the density of utility-relevant interactions (cost-weighted when costs differ); prediction errors, exploration, and distribution shift reduce the saving. We bound selection regret by predictor error and the utility margin. These results motivate our central hypothesis: ρ_u is a measurable coordinate of scaling efficiency and makes informative compute a better description of scaling than total compute.

## What you could build

1. Implement a lightweight utility predictor to pre‑filter queries for a large language model
2. Create an adaptive threshold system that routes high‑value tasks to expensive agents
3. Build a tool‑use agent that updates its predictor based on execution outcomes

## Quiz yourself

What factor reduces execution compute according to the paper’s ideal scenario?

**Answer:** Execution compute is reduced by a factor approaching the inverse of the utility density ρ_u.

---

Notes written by AI from the paper, numbers copied from the abstract. Check the paper before you rely on them. From AI Papers Daily.