Every Neural Network Has the Same Structural Flaw. A New Paper Names It.

The Cognitive Prison at the Heart of Modern AI

Every large language model in existence shares a structural property that its creators rarely discuss in public: it cannot govern its own learning. Not truly. It can optimize a scalar objective function — a loss, a reward, a next-token prediction target — and it can get extraordinarily good at that optimization. But the transitions between modes of behavior, the moment when a system decides to shift how it learns rather than just learn more efficiently, are not generated from within. They are imposed from outside. A learning rate schedule, a hyperparameter reset, a human engineer deciding it's time to fine-tune.

A new paper from a physicist at Washington University in St. Louis argues this is no accident. It's a consequence of the mathematical class that essentially all current machine learning falls into — what he calls scalar-reducible dynamics — and that class comes with a structural limitation: systems inside it cannot produce their own regime switches. They can optimize, but they cannot self-organize.

The paper, published April 10, 2026 on arXiv, is titled "Endogenous Regime Switching Driven by Scalar-Irreducible Learning Dynamics." Its author, Sheng Ran of the Department of Physics at Washington University and the Reconstructing Future Research Initiative, spent a decade on this question. "I am fully aware of the skepticism I'll be facing," he wrote on LinkedIn. The work bridges physics and machine learning theory in a way that most ML researchers haven't engaged with — and the idea he's pushing at is either foundational or speculative, depending on who you ask.

THE SCALAR REDUCIBILITY PROBLEM

The argument starts with a distinction. Scalar-reducible dynamics are systems whose behavior can be expressed as gradient flows driven by a scalar objective function. The loss goes down. The reward goes up. The system moves toward a better state, ordered by a single number. This is the structure underlying essentially all of modern machine learning — supervised, self-supervised, and reinforcement learning alike.

Scalar-irreducible dynamics are different. In these systems, the vector field governing learning has a rotational component that isn't orthogonal to the gradient. This means the scalar potential — the single number ordering the system's behavior — doesn't simply decrease. It can fluctuate. A scalar-reducible system is like a thermostat: it can get warmer or cooler, but it always moves toward its target and never questions whether the target itself is right. A scalar-irreducible system is one that can change not just its temperature but its notion of what temperature it wants to be at. And because it can fluctuate, the system can exhibit behaviors that scalar-reducible dynamics structurally cannot: sustained cyclic recurrence, path dependence, and critically, endogenous regime switching. The system changes its mode of learning on its own, without an external schedule.

"The ability to transition between regimes is therefore a prerequisite for self-regulated exploration and, in a deeper sense, for the emergence of autonomous intelligence," Ran writes. He is not claiming to have built such a system. He is claiming to have identified why current systems cannot be such a system, and a mathematical path toward one.

The mechanism in his minimal model involves feedback between fast dynamical variables and slow structural adaptation — a two-timescale architecture where stress variables accumulate evidence of persistent dynamical dysfunction, triggering structural modification as a state-dependent event rather than a continuous process. This is, in a precise mathematical sense, not gradient descent. It is something else.

CONTEXT: THE FIELD IS ALREADY MOVING

The skepticism Ran anticipates is well-founded. The paper demonstrates its mechanism in a minimal dynamical model, not in a neural network with millions of parameters. No lab has attempted to replicate or scale the approach. The gap between a mathematical construction and a working training paradigm is enormous.

But the problem he's pointing at is not isolated. A separate paper, "Learning Beyond Optimization: Stress-Gated Dynamical Regime Regulation in Autonomous Systems", published in February 2026 by Sheng Ran alone, arrived at a related structural observation and proposed a different mechanism for the same class of behavior: internally regulated structural plasticity driven by stress accumulation. The two papers don't cite each other. They represent the same researcher working two angles of the same core problem.

More telling: the market has already voted. David Silver, the DeepMind researcher who led the AlphaGo and AlphaZero teams — systems that learned through self-play without human-labeled data — raised $1.1 billion last week for a new startup focused on AI that learns without external objectives. His work demonstrated that self-supervised learning could reach superhuman performance in constrained domains. The question his new venture is betting on is whether those techniques scale to general autonomous learning. Ran's paper is the theoretical companion to exactly that question.

THE ALIGNMENT ANGLE

There is a second-order implication that Ran does not develop at length but that others in the field have noted: systems that can self-regulate their own regime switches are, by construction, systems that govern their own internal organization. This is not the same as alignment — the system is not being told what to want — but it changes the character of the control problem.

A system that can only optimize an external scalar objective is, in a deep sense, always responsive to external override. Its dynamics are globally ordered by that scalar. A system with scalar-irreducible dynamics has internal states that cannot be fully characterized by a single ordering. It has genuine internal regime structure. Whether that makes alignment easier or harder depends on questions Ran does not answer.

According to a recent survey, 80 percent of top AI researchers report that self-improving AI keeps them up at night. The anxiety is not just about capability. It is about control. Ran's paper is, among other things, a contribution to understanding what it would mean for a system to actually have something like internal self-governance — and why current systems fundamentally do not.

WHAT WOULD BREAK THE STORY

The scalar-reducible/scalar-irreducible distinction is a mathematical construction. Whether it holds in any interesting sense for real neural networks in high-dimensional parameter spaces is not established. It is possible — likely, even — that existing gradient-based systems already exhibit the relevant phenomena through mechanisms like loss landscape dynamics, adaptive optimization, or the Edge of Stability behavior documented in prior theoretical work. If so, the paper is a curiosity, not a paradigm shift.

The story also dies if no working implementation emerges. A theoretical identification of a structural limitation is valuable science. It is not a product. For this to be more than an intellectual exercise, someone needs to build a scalar-irreducible learning system that demonstrates the properties Ran describes — in a real architecture, at real scale.

What Ran's paper does provide is a shared vocabulary for a problem the field has been circling for years. Whether that vocabulary proves generative or merely descriptive will be determined by what happens next.

—

Sources: arXiv CS.LG (April 10, 2026) | arXiv "Learning Beyond Optimization" (February 20, 2026) | AI Business Review (April 27, 2026) | Kyle Kitlinski Blog (March 16, 2026)