Learning through Internalization
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "Learning through Internalization".
Title and authors: Tom: So, to recap, this paper, "Learning through Internalization," is investigating how transformers absorb an explicit procedure by internalizing chain of thought tokens and then learning to compute the target task directly in one step after those tokens are gone.
Jane: It’s about moving from an explicit reasoning path to an implicit computation path, where the network's weights change to reflect that direct computation.
Lu: The main finding is that for the sparse parity task, a simplified one-layer transformer can provably learn this by first using explicit chain of thought supervision and then internalizing it as those tokens are progressively removed.
Meng: So, they’re showing that direct learning through internalization is possible for tasks that are normally computationally hard to find without that initial supervision.
Tom: And they also looked at how this works for simulating semiautomata, which involves predicting the next state of a semiautomaton after T transition steps given an initial state.
Jane: They found that internalizing chain of thought in these models sometimes leads to shortcut solutions that actually perform worse when we test them on data they haven't seen before.
Lalam: So, while internalization is powerful for learning the task itself, it comes with a trade-off regarding how well the model generalizes to unseen scenarios.
Tom: The paper suggests that this relates to a positive distribution shift framework, where we can use mixtures of chain-of-thought lengths instead of a strict curriculum.
Jane: That’s an interesting idea because it allows us to sidestep the difficulty of direct learning by training on different distribution samples all at once.
Lu: That mixture approach could be more efficient than having to train through a very long, ordered sequence of token removals.
The paper's summary: Tom: One key improvement they suggest is to use a specific next-token prediction loss that specifically targets the correct intermediate chain of thought tokens during each phase of training.
Jane: That means instead of just looking at the final answer, we guide the attention matrix to focus on those relevant coordinates from the hidden parity support as we train.
Lalam: This mechanism should help ensure that even when you remove those explicit reasoning steps later, the model has already built a strong internal representation for that computation.
Tom: And they also look at making models better at simulating semiautomata by designing architectures specifically for that purpose, suggesting using only theta log T layers and an MLP width proportional to the prime power factors of n.
Lu: That’s an architectural suggestion to make the simulation of modular counter semiautomata more efficient by exploiting parallel computation related to the structure of modular arithmetic instead of needing a huge width.
Meng: If we can design these specific architectures, it might mean we don't need massive models just to handle certain types of complex mathematical simulations efficiently.
Jane: It also notes that increasing the number of attention heads actually improves the success rate for internalizing chain of thought for a model called C2 under the inductive curriculum.
Tom: So, in summary, they are proposing both better training loss strategies and more specialized network architectures to enhance this internalization process.
The paper's improvements: Tom: So, wrapping up the "Learning through Internalization" paper, it suggests that we can achieve hard learning tasks like parity by first using explicit chain of thought supervision and then internalizing that generation as tokens are removed.
Jane: The key implication is that this gives us a provable way to move from slow, supervised learning to faster, implicit computation for certain problems.
Lu: The idea of the positive distribution shift framework allows us to use mixtures of chain-of-thought lengths instead of relying on a strict curriculum for training.
Meng: From an engineering standpoint, focusing on wider architectures and specific loss functions seems like a practical way to make these internalization methods more reliable when we try to deploy them.
Lalam: It’s about understanding the mechanics of how the model builds its internal computational path so it can compute things directly later.
Tom: I think the challenge ahead is figuring out which task-specific factors, like prime order semiautomata, make it hardest to internalize and whether we can overcome that difficulty with better training techniques.
Jane: It’s definitely a path forward for understanding how complex reasoning emerges in these systems when they're trained on specific structured data.
Conclusion: Tom: So we’ve been talking about how transformers can learn things they’re actually bad at directly, by first learning them through explicit chain of thought and then just absorbing that computation into their weights.
Jane: That's the core idea of "Learning through Internalization," and it sounds like they showed a way to make models compute things in one step after they've been trained on a slower process.
Lu: The main proof is for the sparse parity task, where this one-layer transformer learns something computationally hard without that initial supervision by internalizing chain of thought tokens as those tokens are progressively removed during training.
Meng: So it means we can teach the model a really tricky math problem by giving it a roadmap first, and then forcing it to figure out the shortcut all on its own later.
Lalam: I see how that works; you give the AI a teacher, and then you make the teacher disappear while keeping your test questions tough. That makes sense.
Tom: Exactly. And they also looked at simulating semiautomata, which is like making an AI predict what happens to a machine after many steps by just knowing the input and where it started.
Jane: They found that sometimes this internalization process actually creates shortcuts that make the model perform worse when it sees new data it hasn't been trained on.
Lu: That’s a big caveat; if the shortcut isn't robust, then you haven't really learned the underlying behavior of the semiautomaton.
Meng: From a production standpoint, we worry about those shortcuts because they might work perfectly in our test set but completely fail when we hit real-world traffic that looks different.
Lalam: I’m fairly certain this paper is quite sure about those generalization issues; they show that the mixture of chain of thought lengths can actually help solve that distribution shift problem.
Tom: So, by using a mixture of chain of thought lengths instead of just one fixed path, you can train on a wider variety of problems at once?
Jane: That’s right. It suggests we don't need to follow a strict sequence for training; we can just present the AI with different reasoning paths simultaneously.
Lu: That mixture approach could be more efficient than forcing the model to go through a very long, ordered curriculum of removing those tokens one by one.
Meng: If it’s more efficient, that’s good because it means less time and fewer training cycles to get a good result on complex tasks.
Lalam: It really seems like a way to make the learning process itself more robust against distribution shifts, which is something we need when building reliable AI.
Tom: So, "Learning through Internalization" gives us a provable path for making models compute directly after they learn slowly.
Jane: It’s an interesting direction because it connects explicit reasoning steps to the model's final internal computation structure.
Lu: It opens up a new way to think about how complex procedures are encoded in neural networks, moving beyond just pattern matching.
Meng: We need to keep watching this because if we can make these methods more reliable for real traffic, it changes how we deploy these kinds of reasoning models at scale.
Lalam: I’m confident the evidence they present on the parity task is solid; it shows the mechanism works exactly as described in their analysis.
Tom: Yeah, it’s a lot to digest because you have both the theory on internalization and the practical implications for how we train these systems.
Jane: It really makes you think about what kind of input distribution we need to provide to get these kinds of powerful internal representations.
Nikolaos Tsilivis, Nirmit Joshi, Marko Medvedev, Julia Kempe, Nati Srebro
New York University · Toyota Technological Institute at Chicago · University of Chicago
cs.LG
Submitted: 2026-06-18
Updated: 2026-11-18
Comments: NeurIPS camera ready version
License: http://creativecommons.org/licenses/by/4.0/
The gist: We study internalization processes, by which neural-network-based systems absorb an explicit computational procedure into their own weights, and how they facilitate learning <ref:2606.20937#pg2>.
Key concepts
- Internalization
- This is a process where a neural network initially learns a complex procedure over many steps but later adjusts its internal weights to perform the same task much more efficiently. It means moving from a slow, step-by-step implementation to a fast, direct computation.
- Distributional Internalization
- A related concept suggesting success can be achieved by matching the performance of a slow system against a fast system under some underlying input distribution. This framework helps explain how learning can succeed even when direct learning is difficult.
- Semiautomata Internalization
- This investigates if transformers can learn to simulate semiautomata, which involves predicting the next state of a machine after T steps. The research found that internalizing CoT often leads to shortcuts that generalize poorly outside the training distribution.
Terminology
Summary
We study internalization processes, by which neural-network-based systems absorb an explicit computational procedure into their own weights, and how they facilitate learning <ref:2606. p. 0 3University of Chicago. June 23, 2026
Abstract
We study internalization processes, by which neural-network-based systems absorb an explicit computational procedure into their own weights, and how they facilitate learning. We investigate how
transformers internalize the simulation of semiautomata by internalizing chain-of-thought (CoT) tokens,
which classes of semiautomata are harder to internalize, and expose the flip side of internalization, that
is, a progressive degradation of out-of-distribution performance. We then provide the first provable
analysis of successful internalization: for the task of learning parities, we show that a simplified one-layer
transformer provably first learns the target with explicit CoT supervision and then internalizes the
autoregressive generation as CoT tokens are progressively removed, learning to directly compute the parity.Where the paper says it
Where the paper says it
Internalization Definitions
The paper defines internalization as a process by which a neural-network-based system initially implements a procedure over many steps, but later its weights are reconfigured to implement the same procedure in far fewer steps <ref:2606.20937#pg2>. Definition 1 formalizes this by requiring output agreement and faster runtime for the internalized program hθ′ compared to the procedure P(hθ) <ref:2606. p. 2 Output agreement:
hθ
′ (x) = P(hθ)(x) ∀x ∈ X . hθ
′ is faster than P(hθ): For all x ∈ X , it holds:
Timehθ′ (x) < TimeP(hθ)(x), (2)
then we say that hθ
′ internalizes P(hθ). We then describe
the main learning phenomenon of study: instead of learning a distribution directly, we first learn a separate
distribution with a “slow” system and then later internalize the “slow” system within a “faster” one. 1 Definition of internalization
Let Σ be an alphabet, X ⊆ Σ
∗ be an input space, Θ be an encoding (parameter) space, and let Y ⊆ Σ
∗ be an
output space.Where the paper says it
Where the paper says it
Internalizing Semiautomata
The study investigates whether transformers can internalize the simulation of semiautomata, which involves predicting the state of a semiautomaton after T transition steps given inputs and an initial state <ref:2606.20937#pg2>. The research focuses on modular counter semiautomata Cn, where the next state is computed as qt+1 = (qt + σt) mod n <ref:2606. p. 5 In our experiments, we mainly consider modular counter semiautomata Cn, which have inputs and states
from Zn = {0, . , n − 1} and the state at time t + 1 is computed as qt+1 = (qt + σt) mod n. • We expose the flip side of internalization: much like humans, transformers that internalize operations
often tend to perform worse in OOD settings (Figure 1, right). In particular, the internalization process
for semiautomata can lead the model to learn shortcut solutions (Liu et al.Where the paper says it
Where the paper says it
Provable Learning Mechanism
The authors provide the first provable analysis of learning through internalization for the sparse parity task, showing that a stylized one-layer transformer can learn representations computationally hard to find directly by internalizing CoT <ref:2606.20937#pg2>. This is achieved through Algorithm 2, which uses a curriculum that progressively removes CoT tokens, leading to the final model directly computing the target parity in one forward pass <ref:2606. p. 2 , 2023), which can generalize
worse than the representation outputted by CoT training. • We provide the first provable analysis of learning through internalization: for the sparse parity task, we
prove that a stylized one-layer transformer with linear attention and a fixed nonlinear MLP first learns
the target parity of hidden support, using explicit CoT, and then internalizes this computation as CoT
tokens are removed one at a time (Theorem 2), thereby directly computing the parity.Where the paper says it
Positive Distribution Shift Framework
Learning through internalization relates to the Positive Distribution Shift phenomenon introduced by Medvedev et al. (2026), suggesting that success can be achieved by coming up with a distribution that is easier to learn from, which is a generalized f-PDS <ref:2606. p. 4 The perspective of PDS here is that even though learning certain distributions
can be “hard” (e. , difficult to learn directly with transformers and standard gradient descent methods, or
impossible with computationally efficient algorithms), one can succeed in learning by coming up with another
distribution that is not only itself easier to learn from, but, upon learning it, one is certain that is has learned
the original “hard” distribution, too. For example, Deng et al. (2024) showed that such a learning strategy
can sidestep the difficulties of learning certain reasoning tasks with transformers: explicit chain-of-thought
supervision was used to train the slow transformer-based autoregressive system, followed by an internalization
process consisting of curriculum training with progressively amputated chain-of-thought data, eventually
yielding a transformer that can respond correctly in a single forward pass.Where the paper says it
Where the paper says it
Architectural and Task Dependencies
The success of internalization is affected by method-specific factors, such as architecture or curriculum choices, where wider rather than deeper architectures are consistently more successful for internalizing CoT <ref:2606.20937#pg2>. Furthermore, the difficulty of internalization depends on task-specific factors; notably, modular counter semiautomata with prime order appear to be the most difficult to internalize <ref:2606. p. 2 • We identify task specific factors that control whether a semiautomaton is easy or hard to internalize. In
particular, amongst the class of modular counter semiautomata those with prime order appear to be
the most difficult to internalize. We show IID success rates across various counter semiautomata for two different width configurations (depth
is fixed to L = 4). We observe that increasing the MLP width of the transformer improves the odds of
successful internalization, but to a lesser extent for Cn with n being a prime number.Where the paper says it
Where the paper says it
Experimental Findings
Experiments reveal that end-to-end (E2E) training fails to simulate semiautomata, while CoT training succeeds, although it incurs high inference time <ref:2606.20937#pg2>. Internalization with shallow transformers often produces networks that generalize perfectly in distribution but exhibit significantly lower out-of-distribution performance compared to CoT training runs <ref:2606. p. 5 • Internalization succeeds with shallow transformers. Remarkably, we find that the internalization
process with the left curriculum often produces networks that generalize perfectly in distribution, despite
them not generating any intermediate tokens; this holds even for non-commutative semiautomata. This task is computationally hard to learn from data without CoT supervision. Finally, we discuss how
learning through internalization relates to the Positive Distribution Shift phenomenon recently introduced
by Medvedev et al.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Internalization of chain-of-thought tokens allows transformers to learn representations that are computationally hard to find directly, such as learning parities, by first learning a slower system with explicit CoT supervision and then internalizing that computation into its weights. This enables the model to compute the target parity in a single forward pass after progressively removing CoT tokens during training.
-
Transformer architectures can be designed to simulate counter semiautomata using a depth of only theta log T layers and MLP width proportional to the prime power factors of n, rather than requiring O(n 2) width for generic simulations. This allows the model to efficiently represent complex operations by exploiting parallel computation inherent in the structure of modular arithmetic.
-
Learning through internalization can be generalized as a Positive Distribution Shift framework, allowing models to learn hard distributions by training on mixtures of distributions corresponding to different chain-of-thought lengths instead of relying on an ordered curriculum. This mixture approach can be more efficient than sequential curriculum training and supports multiple inference budgets for direct answering or longer CoT generations.
-
The model can be trained using a specific next-token prediction loss that targets the correct intermediate CoT tokens in each phase, which enables the attention matrix to progressively concentrate on the relevant coordinates from the hidden parity support. This mechanism ensures that even when explicit reasoning steps are removed, the final output accurately computes the target parity.
Abstract
We study internalization processes, by which neural-network-based systems absorb an explicit computational procedure into their own weights, and how they facilitate learning. We investigate how transformers internalize the simulation of semiautomata by internalizing chain-of-thought (CoT) tokens, which classes of semiautomata are harder to internalize, and expose the flip side of internalization, that is, a progressive degradation of out-of-distribution performance. We then provide the first provable analysis of successful internalization: for the task of learning parities, we show that a simplified one-layer transformer provably first learns the target with explicit CoT supervision and then internalizes the autoregressive generation as CoT tokens are progressively removed, learning to directly compute the parity. This task is computationally hard to learn from data without CoT supervision. Finally, we discuss how learning through internalization relates to the Positive Distribution Shift phenomenon recently introduced by.
Sources
- Towards a theory of model distillation
- Skip-Thinking: Chunk-wise Chain-of-Thought Distillation Enable Smaller Language Models to Reason Better and Faster
- Compressed Chain of Thought: Efficient Reasoning Through Dense Representations
- Implicit Chain of Thought Reasoning via Knowledge Distillation
- From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step
- Training Large Language Models to Reason in a Continuous Latent Space
- Distilling the Knowledge in a Neural Network
- Transformers Provably Solve Parity Efficiently with Chain of Thought
- Implicit Reasoning in Transformers is Reasoning through Shortcuts
- Shift is Good: Mismatched Data Mixing Improves Test Performance
- Positive Distribution Shift as a Framework for Understanding Tractable Learning
- Hardness of Learning Fixed Parities with Neural Networks
- Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation
- Distilling System 2 into System 1
Related papers
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis
- (Mis)generalization of Helpful-Only Fine-tuning
- Provably Learning Diffusion Models under the Manifold Hypothesis: Collapse and Refine