🤖 AI

The Computational Boundary of Inference: Capability Internalization, Training, and the Turing Jump

This paper establishes a formal computability-theoretic limit on recursive self-improvement narratives by proving that finite internal self-modification remains confined within an existing computational layer, while any qualitative ascent to a stronger capability level requires a transition to a higher Turing jump rather than mere repetition within the current regime.

Original authors: Chien-Ping Lu

Published 2026-05-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chien-Ping Lu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Question: Can a System Just "Think Harder" to Become Smarter?

Imagine you have a very smart robot. People often argue that if you let this robot fix its own mistakes over and over again, it will eventually become a super-intelligence that can solve problems we can't even imagine. The idea is: More updates = More power.

This paper says: Not necessarily.

The author, Chien-Ping Lu, uses the strict rules of math (specifically computer science logic) to prove that there is a hard ceiling on how much a system can improve just by "thinking" or "revising" itself internally. To get past that ceiling, something else has to happen.

Here is the breakdown of the three different "modes" of improvement the paper identifies.

1. The "Loop" Mode: Finite Internal Revision

The Analogy: Imagine a chef in a kitchen who is trying to make the perfect soup. They taste the soup, add a pinch of salt, taste it again, add a little pepper, and taste it again. They do this 10 times, then 100 times.

The Paper's Claim:
As long as the chef is working with the same set of ingredients and the same basic cooking skills they started with, no matter how many times they taste and adjust, they cannot suddenly invent a new flavor that was impossible with their original ingredients.

In the paper's math language, this is called finite internal revision. If a system is stuck inside one "layer" of capability, repeating its own updates (like prompting, re-ranking, or self-correction) will never break out of that layer. It can get better at what it already does, but it cannot become qualitatively stronger. It hits a wall.

2. The "Limit" Mode: Stabilized Revision

The Analogy: Now, imagine the chef doesn't just taste the soup once. Instead, they run a simulation where they try to guess the perfect recipe. They make a guess, check it, make a better guess, check it again, and keep going forever. Eventually, their guesses stop changing and settle on one perfect recipe.

The Paper's Claim:
This is called stabilized revision. The paper proves that if a system does this kind of "infinite guessing until it settles," it actually jumps to a higher level of power. It's like the chef suddenly gaining the ability to see the "perfect recipe" that was invisible before.

However, this isn't just "more thinking." In the math world, this specific type of "settling down" is equivalent to a massive leap in power (called a "Turing Jump"). The paper argues that you can't get here just by doing the same thing over and over; you have to be doing a specific kind of "limit" process.

3. The "Upgrade" Mode: Capability Internalization

The Analogy: Imagine the chef realizes they can't make the perfect soup alone. They need a secret spice that only exists outside their kitchen.

Step A: They ask a human (external guidance) for the spice. The human gives it to them, and they make the soup.
Step B: The chef goes back to the market, buys the spice, and bakes it into their own apron. Now, the chef is the spice. They don't need to ask anyone anymore.

The Paper's Claim:
This is the only way to truly "ascend" to a new level of intelligence. The paper calls this capability internalization.

If the guidance (the spice) comes from outside the system (like a human teacher, a reward signal, or a search tool), the system is just borrowing power.
To become a stronger system permanently, that external help must be "baked in" (via training, retraining, or fine-tuning) to create a new, stronger version of the system.

The paper argues that you cannot get a new, stronger system just by the old system "thinking" about the problem. You need a process (like training) that takes that external help and turns it into a permanent part of the system's new brain.

The Three Rules of the Paper

Repetition isn't enough: If you just keep updating a system internally (like a robot fixing its own code), it stays in the same "power tier." It can't jump to a higher tier just by looping.
Stabilization is a jump: If a system's corrections eventually "settle" into a final answer after an infinite process, that answer belongs to a higher power tier. But this is a specific mathematical condition, not just "trying harder."
Training is the bridge: The only way to permanently move from a "weak" system to a "strong" system is to take guidance that was previously outside the system (like human feedback) and internalize it into a new, upgraded system.

What This Means for AI (According to the Paper)

The paper does not say AI can't get smarter. It says we need to be careful with our words.

Don't say: "The AI is improving itself recursively, so it will eventually become a god."
Do say: "The AI is refining its current skills (which is good), but to become a new kind of smarter AI, it needs to be retrained using external guidance to build a stronger foundation."

The paper is a warning against the idea that "more iterations" automatically equals "more power." It draws a clear line: Thinking inside the box is different from building a bigger box. To build the bigger box, you need to bring in new materials from outside and reconstruct the whole thing.

Technical Summary: The Computational Boundary of Inference

Problem Statement
The paper addresses a conceptual ambiguity prevalent in discussions of recursive self-improvement in Artificial Intelligence. Informal narratives often conflate three distinct processes: (1) repeated internal revision within a fixed system (e.g., prompting, reflection, tool use), (2) revision processes that stabilize over time (convergent approximation), and (3) genuine qualitative ascent to a strictly stronger computational capability regime. The central problem is that these narratives often assume that sufficient accumulation of internal updates automatically yields a stronger capability level. The paper seeks to provide a formal, theorem-level test of this assumption using classical computability theory to determine whether finite internal iteration can mathematically justify a transition to a strictly stronger relative degree.

Methodology and Modeling Framework
The paper employs classical oracle computability theory (Turing reducibility) as a structural idealization rather than a literal model of modern machine learning pipelines. The core framework relies on the following definitions:

Computational Layer ( $C(A)$ ): For a fixed oracle $A$ , the layer is defined as $C(A) = \{B : B \le_T A\}$ , representing all tasks computable relative to $A$ .
Finite Internal Revision: Modeled as a sequence of indices $e_{s+1} = F(e_s)$ where $F$ is an $A$ -computable function. This represents self-modification strictly within the existing effective regime.
Stabilized Revision: Modeled as an $A$ -computable approximation $g(x, s)$ that converges pointwise to a limit. This represents iterative correction that eventually settles on an output.
Capability Internalization: Defined structurally as the process where a "settled deployed system" (a fixed family of procedures) incorporates "externally essential guidance" (tasks not solvable within the current layer) into a new, stronger successor system.

The methodology distinguishes between inference-time activity (finite composition of existing procedures) and training/internalization (the transformation of external guidance into the parameters of a successor system).

Key Contributions and Results
The paper establishes a series of separation theorems that delineate the boundaries between these regimes:

Local Closure vs. Escape (Theorem 3): The computational layer $C(A)$ is locally closed under oracle computation (composition of $A$ -computable functions remains in $C(A)$ ), but it has a canonical boundary: the Turing jump $A'$ . $A'$ is not solvable within $C(A)$ .
Degree-Boundedness of Finite Revision (Theorem 4 & Proposition 1): Finite internal self-modification (repeated application of an $A$ -computable update rule) remains strictly within $C(A)$ . No finite stage of such a process can compute $A'$ or any set strictly stronger than $A$ . Thus, finite iteration alone does not constitute ascent.
Stabilized Revision Corresponds to the Jump (Theorem 5): If revision is modeled as a convergent $A$ -computable approximation, the stabilized limit is computable in $A'$ (and conversely, every $A'$ -computable set arises from such a process). This establishes that stabilized revision belongs to a different recursion-theoretic level (the jump) than finite internal iteration.
Nested Revision and Higher Jumps (Theorem 6): Two-level nested stable revision corresponds exactly to the second jump $A''$ . This suggests a structural correspondence where the depth of nested stabilization tracks the depth of the Turing jump hierarchy.
Capability Internalization Forces Ascent (Proposition 4 & 5): If a successor system internalizes a task that is "externally essential" (solvable with external guidance but not within the original layer $A$ ), the successor system must operate at a strictly stronger degree ( $A <_T B$ ).

Significance and Claims
The paper claims to provide a "clean formal separation" that constrains the logic of recursive-improvement narratives. Its significance lies in the following points:

Refutation of Simple Accumulation: The paper demonstrates that "more updating" (finite internal iteration) is not mathematically equivalent to "qualitative ascent." A system can undergo infinite internal refinement while remaining within the same computational degree.
Structural Role of Training: The paper clarifies that the significance of machine learning training in this context is not that it is "learning" per se, but that it is the empirical mechanism for capability internalization. Training is the process by which externally essential guidance (which may require a jump-level oracle to solve) is baked into a new deployed system, thereby enabling the system to operate at a higher relative degree.
Clarification of AI Discourse: The results argue that claims about open-ended self-improvement must specify which regime is being invoked. If a narrative relies on finite internal revision to explain the emergence of a stronger capability, it is mathematically insufficient. Genuine ascent requires the internalization of guidance that was previously external to the settled layer.

The paper explicitly states it does not claim that current AI systems literally compute Turing jumps or that training pipelines instantiate oracle machines. Rather, it offers a structural language to distinguish between processes that remain within a fixed effective layer and those that necessitate a transition to a stronger relative level. The contribution is a "structural limitation result" that forces recursive-improvement claims to be framed with greater mathematical precision.