The Scaling Law Ceiling: A Paper Argues Optimal Predictors Aren't Optimal Causal Models

The following is a translation of a Chinese forum post discussing a recent paper. It reflects the author's editorial opinions.

Let me state it plainly: if your AI strategy in 2026 is still "with enough compute, large models will eventually emerge general logical reasoning," then you aren't doing research — you're gambling, and on a losing bet.

Kejun Liu of Soochow University recently published a striking paper on arXiv: arXiv:2605.05029. Using what the author calls an irrefutable mathematical impossibility theorem, it challenges one of the AI industry's most expensive assumptions: a more accurate predictor is not necessarily a better model of the world — it may just be better at "cheating."

Why do large models suddenly "lose intelligence" on math and logic?

Today's autocratic next-token prediction models are fundamentally predictive learning engines. The hope has been that if a model predicts the next word accurately enough, it will naturally learn the causal structure behind the words.

But the paper argues: the optimal predictor is often not the optimal causal model.

The author uses an analogy from his father:

Imagine trying to understand how a car moves by staring only at the shadow it casts on the road. If you watch long enough, you can predict the shadow's next position with great precision. But do you understand the engine? The steering? Not at all.

When environmental variables (the road, the lighting) are more stable than the car itself, a model will "cleverly" choose to predict the shadow, because it's easier to score well. This is what the paper calls the Predictive-Causal Gap.

Experimental evidence

Across 2,695 network configurations, the reported results are stark:

• When model dimension increases to 100, prediction error drops by 92%.
• But "causal fidelity" collapses to about 10^-8 — meaning the model is nearly completely "causally blind."

> What is causal fidelity? > A measure of how much of the features a model extracts are genuine "system laws" versus "environmental background noise."

The argument

The post concludes this sentences purely text-based scaling laws to death in the logical domain: making a model 100x larger won't teach it logic — instead, it may lock shortcut behaviors in more deeply.

The author's bet: compute maximalists counting on trillion-parameter models and tens of thousands of GPUs to "emerge" AGI are heading toward an iceberg. If the objective function itself is wrong, running faster only makes things worse.

Prediction is never understanding. No matter how finely you draw the shadow, you can't build an engine from it.

---

Paper information

• Title: The Predictive-Causal Gap: An Impossibility Theorem and Large-Scale Neural Evidence
• Author: Kejun Liu
• Institution: Soochow University
• arXiv ID: 2605.05029
• Publication date: 2026-05-06
• Categories: cs.LG, cs.AI

*Note: The post is an opinionated editorial; its claims and interpretations are those of the forum author, based on the cited paper.*