Paper review · November 3, 2026
Does Scaling Toward AGI Depend on What System-1 Lets System-2 Execute?
The paper formalizes a 'conditional interaction' loop in which a lightweight predictor (System-1) screens execution candidates in advance for a costly executor (System-2), and proposes the hypothesis that the density of useful interactions, ρ_u, is a coordinate that describes scaling efficiency.
Problem
Progress toward AGI has mainly come from increasing System-2 resources such as models, data, compute and reasoning length. However, the marginal gain from each individual interaction is shrinking. For example, prompts where every Rollout in reinforcement learning gets the same reward, reasoning tokens that do not change the answer, and environment steps that repeat already-known information only add cost. A structure is needed that judges, before execution, which interactions are worth executing.
Method
The paper defines a predict → gate → execute → update loop. System-1 cheaply predicts the utility of candidate interactions (prompts, tasks, queries, tool calls). Only candidates above an adaptive threshold are executed by System-2, and the execution results are used to recalibrate System-1. The framework is applied to training (RLVR prompt selection), test-time reasoning, adaptive decision-making and agent tool use, and theoretical bounds on compute savings and selection regret are derived. Unlike static cascades that are trained once and left as is, the authors argue that the two systems must be designed and optimized together.
Results
- With ideal filtering and equal execution costs, execution compute is reduced by a factor of about 1/ρ_u (ρ_u is the fraction of interactions whose utility exceeds the threshold). When costs differ, this becomes a cost-weighted density.
- The analysis finds that prediction error, exploration and distribution shift reduce these savings.
- An upper bound on selection regret is given in terms of predictor error and utility margin.
- On this basis, the authors propose the hypothesis that scaling is better explained by 'informative compute' than by total compute.
Limitations
This is a perspective paper. Its core claims are a hypothesis and theoretical bounds, and large-scale empirical validation could not be found in the provided text. The authors also limit its scope to settings where the utility predictor is reliable, and state that the method does not replace exploration when the predictor is uncertain. ρ_u depends on how utility is defined, so values under different definitions cannot be compared with each other.
Why it matters
It lets 'pre-execution screening' problems that practitioners handled separately, such as Rollout waste in RLVR training, model routing and adaptive reasoning, be viewed through a single framework and metric (ρ_u).
Key terms
- System-1 / System-2
- A distinction borrowed from Kahneman's dual-process theory. Here it refers to a cheap, fast utility predictor (System-1) and a costly, powerful execution model or agent (System-2).
- Conditional interaction
- A loop that executes only candidates whose predicted utility exceeds a threshold and updates the predictor with the execution results.
- ρ_u (useful interaction density)
- The fraction of candidate interactions whose utility exceeds a set threshold. The authors propose it as a coordinate that describes scaling efficiency.
- RLVR
- Reinforcement learning that gives rewards using problems with automatically verifiable answers. Prompts where every Rollout gets the same reward provide no learning signal.
Original
Authors: Qi Wang, Yun Qu, Lizhou Cai, Heming Zou, Haoang Chi, Yiqin Lv, Xiang Li, Haiyan Feng, et al. · Posted 2026-10-29 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.