Monday, 2 November 2026 · 6 min read
Co-designing a lightweight predictor with a costly executor improves interaction efficiency toward AGI
This paper proposes a conditional interaction loop where a cheap System-1 predicts the utility of candidate interactions, gating which costly System-2 executions occur, improving compute efficiency by up to the inverse density of useful interactions, and formalizes this as a scaling coordinate.
Based on the paper Does Scaling Toward AGI Depend on What System-1 Lets System-2 Execute? by Qi Wang, Yun Qu, Lizhou Cai, Heming Zou and 9 others
The challenge of costly interactions in scaling intelligent systems
Progress toward artificial general intelligence (AGI) has mainly focused on scaling a resource-intensive executive, called System-2, which performs reasoning, planning, and decision-making using large models and extensive compute. However, as systems grow, the cost of executing every possible interaction (such as generating tokens or running rollouts) becomes a bottleneck.
Each interaction consumes resources before its outcome is observed and yields utility only afterward. Many interactions provide little useful information or learning signal, leading to diminishing returns and wasted compute.
Determining which interactions are worth executing is difficult because assessing their value must cost less than executing them. This motivates co-designing a lightweight System-1 that predicts interaction utility cheaply and gates System-2’s execution, focusing resources on valuable interactions.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Three key formulas from the paper, explained symbol by symbol
- Two figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short