Optimal Representation Size: High-Dimensional Analysis of Pretraining and Linear Probing
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: "Optimal Representation Size".
Jane: Here's the gist of the paper. Simulations were distributed over a SLURM cluster across three main experimental setups: Optimal Representation Size, Autoencoder experiment, and LLM probe experiment.
Tom: For the Optimal Representation Size analysis, simulations were run on a SLURM cluster where each job was allocated 4 CPUs and 30 GB of RAM.
Jane: The full grid configuration involved 6 SNR times 4 lambda 1 times 4 lambda 2 times 30 n b and produced 2,880 SLURM jobs.
Tom: Each job processed 30 values of alpha with 100 independent replications at a matrix dimension of p=500.
Title and authors: Tom: Alright, so we're going through "Optimal Representation Size: High-Dimensional Analysis of Pretraining and Linear Probing." The title itself tells you this isn't just about picking a number; it’s an analysis of the high-dimensional regime when you have pretraining followed by linear probing.
Jane: Right, so before we get into the details, what's the main takeaway for someone who doesn't read math? What is this paper actually saying about how to set up these two stages of learning?
Lu: It’s formalizing the idea that pretraining extracts structure from unlabelled data using something like principal component analysis, and then we learn the downstream task by doing linear regression on a separate labelled dataset. This whole pipeline is what they are analyzing.
Meng: So when you look at the practical side, is this paper suggesting we just use a fixed size, say two hundred dimensions, or does it give us a dynamic rule? Because for me, if I’m deploying something, I need to know what’s stable.
Lalam: It suggests that the optimal size depends on how the pretraining data and downstream labelled data are set up. They are deriving exact expressions for training in this high-dimensional regime >
Tom: Exactly. And they use a framework called spiked covariance models, which is a tool from statistics that helps them understand when those principal components actually align with the real underlying structure of the data, versus just being noise.
Jane: That’s where it gets interesting. The idea is that if the representation dimension is too high or too low relative to how much data you have, the quality of your downstream learning changes drastically.
Lu: They focus on a concept called the BBP phase transition, which tells you whether those principal components are useful or not based on the signal strength versus the noise level in your data >
The paper's summary: Tom: So summarizing what they did, they built an analytical model to characterize exactly how big that optimal retained dimension is going to be, depending on pretraining and downstream data. It’s a closed-form characterization >
Jane: When you unpack that for a listener, it means they are trying to solve the question: "Given I have lots of unlabelled text but very little labelled text for my specific task, how should I set the size of the AI's internal knowledge base to get the best results?"
Lu: They show that this optimal dimension isn't just a guess; it’s derived from analyzing how much signal is in your pretraining data compared to how much signal you have in your downstream labelled data >
Meng: Okay, so if I run my model and see that the required dimension for good performance is, say, five hundred instead of two hundred, what does that imply about the computational cost? Is it worth the extra memory allocation?
Lalam: It implies that using a smaller representation size would be detrimental because it wouldn't capture enough structure to transfer well to the labelled task >
Tom: It’s about capturing enough signal. And they use this analysis to show how representations support computation and transfer, or if they just cause catastrophic interference when things get complicated >
Jane: It connects the idea of representation quality directly to how well a model adapts, which is a big step because we often treat the two stages—pretraining and probing—as separate problems.
Lu: And by using this analytical theory, they can tell us if we are in the regime where high-dimensional structure is helpful or if we should stick to lower dimensional representations >
The paper's improvements: Tom: So the paper suggests a few ways to use this analysis. First, it gives a principled way to tune that optimal representation size based on the data availability—specifically when pretraining data is abundant but downstream data is scarce, in that case, maximally compressed representations are actually better for generalization >
Jane: That sounds counterintuitive at first. If we have lots of pretraining but little labelled data, why would we want to compress? What's the logic there?
Lu: Because if you have too much capacity from the pretraining but not enough specific guidance from the labelled data, a smaller representation prevents overfitting on that tiny labelled set >
Meng: So it’s about managing the trade-off between capturing general knowledge and focusing on task specifics. That sounds like something we deal with when we try to fine-tune models >
Lalam: They also show how you can quantify the trade-off between pretraining and supervision by identifying specific regimes where using unlabelled data actually reduces generalization error more than using labelled samples >
Tom: That’s a huge thing for understanding why large models trained on massive amounts of data are so effective for fine-tuning, because it gives us a mathematical reason for that effectiveness >
Jane: And then they get into adapting the representation size based on task structure itself. It suggests we can determine if low dimensional representations are actually useful or if specific tasks need those detailed high dimensional representations >
Conclusion: Tom: So to wrap up, the main implication of "Optimal Representation Size: High-Dimensional Analysis of Pretraining and Linear Probing" is that we now have a formal, analytical way to determine the best representation dimension by looking at the data you have and the task structure.
Jane: It moves us past just guessing sizes; it gives us a principled way to tune those representations for better generalization when we move from massive pretraining to specific downstream tasks.
Lu: The key is recognizing that there's a phase transition where representation quality changes dramatically based on the signal strength in your data, and this paper explains how to find that point >
Meng: From an engineering standpoint, it helps us understand the cost of running different sizes. If we can predict the optimal size before we train everything, we save time and compute resources >
Lalam: I think it really solidifies the idea that pretraining on large datasets is powerful because it creates a structure that can be optimally refined for any subsequent learning step, as described in this paper >
Valentina Njaradi, Clémentine Dominé
University College London · Institute of Science and Technology Austria · Sainsbury Wellcome Centre
cs.LG
Submitted: 2026-05-19
Updated: 2026-11-09
Comments: Accepted to NeurIPS 2026
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 83/100
The gist: Simulations were distributed over a SLURM cluster across three main experimental setups: Optimal Representation Size, Autoencoder experiment, and LLM probe experiment.
Key concepts
- Optimal Representation Size
- This concept refers to finding the ideal dimensionality ($\alpha^*$) for a model's internal representations. The paper investigates how this optimal size changes depending on factors like the amount of unlabelled versus labelled data, the signal-to-noise ratio (SNR), and task alignment, aiming to minimize generalization error.
- Generalization Error
- This measures how well a model trained on one set of data performs on unseen data. The research seeks to find representation sizes that result in the lowest generalization error. This is crucial because it determines if the learned representations are truly useful for new, related tasks beyond the training set.
- Pretraining vs. Supervision Trade-off
- This examines how much knowledge a model gains from large amounts of unlabelled pretraining data versus specific labelled supervision. The study quantifies regimes where using extensive pretraining helps more than adding labelled samples, explaining why large models are effective for fine-tuning.
Terminology
Summary
Simulations were distributed over a SLURM cluster across three main experimental setups: Optimal Representation Size, Autoencoder experiment, and LLM probe experiment. p. 6 5 Validation in autoencoders and pretrained transformers
In this section, we evaluate the extent to which theoretical predictions are valid beyond the idealised setting employed
in their derivation. We first consider a controlled synthetic experiment in which the representation learned by an
autoencoder approximates PCA.Where the paper says it
For the Optimal Representation Size analysis, simulations were run on a SLURM cluster where each job was allocated 4 CPUs and 30 GB of RAM. The full grid configuration involved 6 SNR times 4 lambda 1 times 4 lambda 2 times 30 n b and produced 2,880 SLURM jobs. p. 6 The full phase transition curve, derived in Corollaries 1 and 2, matches the phase transition observed in the heat map
of Figure 3(a) (shown in white; see also additional heatmaps in Figures 8 and 10 in Appendix B. Increasing the
SNR progressively shrinks the region where compression is beneficial (Figure 3(b)), making full-rank representations
(α = 1) optimal over a wider range of sample sizes. 75
Val accuracy
(g)
Step
2
8
32
128
512
2,000
5,000
20,000
80,000
143,000
0. 0
O
ptim
al
*
100
200
300
400
500
O
ptim
al m
Spiked identity cov.Where the paper says it
Where the paper says it
The Autoencoder experiment utilized a SLURM cluster where each job was allocated 1 GPU, 8 CPUs, and 16 GB of RAM. The full grid configuration involved 20 n base times 20 n task and produced 400 SLURM jobs. p. 4 Importantly, optimising the representation dimensionality reduces generalisation error across a broad region of parameter
space, both relative to low-dimensional representations with a single PC (α ≈ 0, Figure 2(c)) and to full-rank
representations (α = 1, Figure 2(b)), with the optimum shown in Figure 2(a). These gains are particularly pronounced
in regimes with substantial pretraining data and limited downstream samples, where standard PCR or regular regression
fail to achieve comparable performance even in misaligned task settings (see Figure 12 in Appendix C. Finally, spiked
covariances have been used to analyse PCR, showing how the projection onto empirical PCs has a regularising effect
[78]. Our setting differs from these works: the signal subspace is estimated from an independent unlabelled pretraining
sample and then used for a separate labelled task, allowing us to characterise the optimal representation dimension. Overall, our empirical results show that the theoretical predictions derived in the idealised setting extend to trained
linear and nonlinear autoencoders when enough pretraining data is available, and that the qualitative phase structure
persists under general covariance spectra. 2 Linear probing transformers
We next examine whether representations learned by modern architectures exhibit similar dimensionality effects under
linear probing.Where the paper says it
Where the paper says it
Where the paper says it
The LLM probe experiment was also distributed over a SLURM cluster, with jobs allocated 1 GPU, 16 CPUs, and 32 GB of RAM. Two sets of 10 SLURM jobs were submitted: one set for the downstream-PCA condition using a checkpoint of Pythia-70M-deduped, and another for the pretrain-PCA condition. Each job swept over 10 training-set sizes and 3 seeds, fitting a linear probe for up to 10,000 epochs with early stopping across up to 14 PCA dimensionalities (m).
Furthermore, Large Language Models (LLMs) were extensively used throughout the research project to assist with writing, editing, and formatting the manuscript. LLMs were also employed as an aid for understanding and discussing technical concepts, as well as for debugging and drafting portions of the experimental and simulation code. All scientific results, analyses, and conclusions derived from these simulations were rigorously verified and validated by the authors. p. 8 co-supervised the project and contributed to the technical development of the work, including proofs, experimental
design, and interpretation of results. co-supervised the project and contributed to the conceptual framing of the
project, interpretation of the theoretical results, experimental design and overall scientific direction.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
-
Optimal representation size tuning for pretraining pipelines can be dynamically determined based on data availability; specifically, when abundant pretraining data is present but downstream data is scarce, maximally compressed representations are optimal to improve generalisation.
-
The system can achieve improved generalisation by finding the optimal representation dimensionality alpha star which minimizes the generalisation error; this is characterized as a function of unlabelled and labelled sample sizes, task alignment, and signal-to-noise ratio.
-
The AI system can quantify the trade-off between pretraining and supervision by identifying regimes where unlabelled data yields larger reductions in generalisation error than labelled samples, providing a principled explanation for why large models trained on extensive data are effective for fine-tuning and linear probing.
-
For low sample regimes, the system can leverage higher dimensional representations to generalise better; this is achieved when pretraining data is limited by using higher dimensional representations.
-
The system can be designed to adapt its representation size based on task structure; it can determine whether low-dimensional representations are useful for generalisation or if tasks benefit from detailed high-dimensional representations, identifying phase transitions where compression becomes useful.
Abstract
Learning to generalise from limited data is a fundamental challenge for both artificial and biological systems. A common strategy is to extract reusable structure from abundant unlabelled data, enabling efficient adaptation to new tasks from limited labelled data. This two-stage paradigm is now standard in modern training pipelines, where pretraining is followed by fine-tuning or linear probing. We provide an analytical model of this process: structure extraction is formalized as principal component analysis on unlabelled data, and downstream learning as linear regression on a separate labelled dataset. In the high-dimensional regime, we derive exact expressions for training and generalisation error showcasing their dependence on representation dimensionality, unlabelled and labelled sample sizes, and task alignment. Our results show that pretrained representations strongly influence downstream generalisation, and we characterize the optimal representation size as a function of task parameters: with abundant pretraining data but scarce downstream data, maximally compressed representations are optimal, whereas with limited pretraining data, higher-dimensional representations generalise better. Furthermore, we establish an exact trade-off between pretraining and supervision, quantifying how much unlabelled data is required to replace a single labelled sample. Beyond our idealised model, we observe similar phenomenology in autoencoders and pretrained LLMs. Altogether, we highlight that optimising representation size is critical, giving conditions for when compression during pretraining improves generalisation.
Sources
- GPT-4 Technical Report
- The Llama 3 Herd of Models
- Optimal Regularization for Performative Learning
- Generalization error of min-norm interpolators in transfer learning
- A Theoretical Analysis of Contrastive Unsupervised Representation Learning
- Understanding intermediate layers using linear classifier probes
- A mathematical theory for understanding when abstract representations emerge in neural networks
- A Theory of How Pretraining Shapes Inductive Bias in Fine-Tuning
- Datasheet for the Pile
- From Principal Subspaces to Principal Components with Linear Autoencoders
- The Pile: An 800GB Dataset of Diverse Text for Language Modeling
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