Does AGI scaling depend on System-1 and System-2? A new paper

A new perspective paper from Tsinghua University asks whether scaling toward AGI depends on System-1 and System-2 working together, published on arXiv on October 29, 2026. The authors argue that a small, cheap predictor (System-1) should decide which interactions a large, costly executor (System-2) actually runs. Their answer is a "qualified yes", backed by theory. The paper reports no experiments of its own.

It is worth reading because it names a way to scale that is not "more parameters, more data, more tokens". It is also a good example of how much a framing paper can claim before anyone has tested it.

The problem: most compute is spent on interactions that teach nothing

The paper starts from a familiar worry. Human-written text is finite, high-quality data is scarcer still, and some physical experiments can't be sped up. Using a simple power-law model of loss, the authors show that each extra unit of compute buys less than the last when every interaction is run unconditionally.

Their concrete example is reinforcement learning for reasoning models, in the GRPO style with verifiable rewards. For each prompt the model generates several answers. If every answer gets the same reward, all right or all wrong, the prompt produces zero learning signal, yet it still cost the full compute to generate. Robots have the same problem in the physical world: most practice trajectories cover situations they have already mastered, and real rollouts are slow and can't be freely parallelized.

The idea: predict, gate, execute, update

The authors call the mechanism conditional interaction, a four-step loop:

1. System-1 predicts how useful each candidate interaction will be.
2. A gate keeps only candidates whose predicted utility clears an adaptive threshold.
3. System-2, the expensive part (a large model, a simulator, a robot), runs only those.
4. The outcomes are used to update System-1.

"Utility" is left flexible. The running example is expected Bayesian surprise, meaning how much an outcome would change the system's beliefs. Other options are value of information and risk during adaptation.

What the paper shows, in numbers

All the numbers in the paper are theoretical. They are upper bounds under ideal conditions, and the authors say so.

• If System-1 screens perfectly and every candidate costs the same to run, execution compute falls by a factor that approaches 1/ρu, where ρu is the share of interactions that carry useful signal. As an illustration of the formula, if one interaction in ten matters, the ceiling is a tenfold saving. The paper gives no measured ρu.
• Prediction errors, exploration and distribution shift all reduce that saving.
• The regret from picking the wrong candidates is bounded by the predictor's error and the margin in utility between good and bad candidates.
• The authors propose ρu as a measurable "coordinate" of scaling efficiency, and "informative compute" as a better description of scaling than total compute. In their model, gating shifts the loss curve sideways by ln(1/ρu) on a log-compute axis. A second possibility they name is that gating also changes the slope of the curve. Which one happens is left to experiment.

Where it already shows up

The paper surveys existing work as evidence that the pattern is real. In reasoning-model training, DAPO's dynamic sampling throws out prompts whose answers all score the same, but only after generating them. Newer methods such as GRESO and MoPPS predict which prompts are worth sampling before generating anything. The paper mentions "several-fold rollout savings" for one of them, a figure it takes from the authors' earlier work and doesn't reproduce here. For robots, MPTS predicts adaptation risk from optimization history instead of evaluating every task. The paper also extends the loop to test-time compute, deciding when a model should stop thinking, and to agents deciding when a tool call is worth its latency and cost.

Limits the authors admit

• The efficiency gains are "ideal values, not guarantees".
• There is no convergence guarantee for the full co-design loop, where the predictor and executor improve together.
• The scaling claim depends on a hypothesis that ρu keeps shrinking as systems improve. The authors call it empirical and possibly false.
• Gating is not justified when the predictor is poorly calibrated, when exploring has value in itself, or when a missed interaction would be catastrophic.
• Measuring ρu requires an audit pool of ungated interactions, because the gate hides what it skips.
• A stronger System-2 might make its own outcomes harder to predict.
• The test-time and agent claims have a named test but no measured results.

We could not find a code or data release in the paper text we read. The project page also had no code link in the portions we could open, and its paper link was a placeholder.

Why it matters

The useful takeaway is a question to ask of any training or inference system: how much of the compute goes to interactions that change nothing? That question can be measured, and the practical methods that already filter prompts suggest it is a real cost. Whether it amounts to a path to AGI is a much larger claim, and a perspective paper without experiments can't settle it. The next thing to watch is whether anyone measures ρu across model sizes and reports whether the curve shifts or bends. The project page lays out the framing in plainer terms.