- Rating
- 2(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper asks how small a transformer can be and still memorize N random (subject, relation, object) facts, with every entity a single token. For one-layer GPT-2 blocks, the minimum width grows logarithmically in N, and a lack-of-fit F-test rejects power-law and linear fits but not this law. The authors read this as memorization capacity growing exponentially with width, unlike the linear scaling of earlier empirical work, which they attribute to limited tuning and training. Experiments on depth, structured and Zipfian data, and a second architecture (OLMo-3) extend the core measurement.
Critical strengths
-
The measurements are done carefully. Each dataset size gets its own hyperparameter search over three seeds, and a lack-of-fit F-test, not R² alone, selects the functional form. Apart from the training budget (Q1), I find the widths credible, so my concerns are mostly about their interpretation.
-
The parameter accounting is transparent. Table 4 and the "Where the bits live" paragraph (p.9) separate embedding from non-embedding parameters and report bits per parameter with and without embeddings. That makes the data a reusable capacity measurement for models with one embedding row per entity.
Critical weaknesses
- Counted in total parameters, the GPT-2 results match prior work rather than beat it. Each entity is a single token with its own embedding row, and for large datasets these rows hold nearly all the parameters (Table 4). Over all parameters, the paper's own analysis on p.9 finds close to 2 bits per parameter beyond the five smallest datasets, in line with Allen-Zhu & Li (2025), who also count embeddings. Yet the introduction claims substantially better scaling than that work, and this mismatch is the main reason for my rating. Per non-embedding parameter, the largest dataset would need several times what a 32-bit weight can hold, so the facts must live mainly in the embedding rows. A roughly constant rate per embedding coordinate then makes width grow logarithmically with the number of facts. The authors keep the embedding table uncapped to avoid an artificial bottleneck, but their vocabulary test (Table 8) still gives each entity its own token and, by my count, yields a similar rate. Lu et al. (2024) and Allen-Zhu & Li, like deployed LLMs, instead spell entity names with several tokens from a fixed vocabulary and have no per-entity rows. Since Lu et al. count only non-embedding parameters, the gain over their linear scaling shows up precisely in the unit that leaves these rows out. By my count, OLMo-3 stores more bits per total parameter than GPT-2 at every dataset size both were trained on, tied embeddings or not, but it has the same per-entity rows. Neither result therefore supports the suggestion that LLMs may hold more knowledge than estimated (p.3 l.108-110), and settling this needs new experiments that could change the main claim. Concretely, I would like to see the width search repeated at a few sizes with multi-token entity names over a fixed vocabulary, or with frozen random embeddings, reporting bits per trainable parameter.
- The case against a power law is weaker than it looks. Table 1's decisive rejection of the offset power law depends on a constraint that keeps the fitted width positive for every dataset size, however small, which pins the offset at zero. The paper exempts the pure logarithm from this constraint, and a free-offset power law turns negative only below the smallest dataset measured. My refit without the constraint, using per-seed widths reconstructed from Table 4 and pure error inferred from Table 1, rejects the free-offset form only narrowly, just under the 0.05 threshold. The pure error also comes from the seeds that selected each configuration (App. B, Eq. 3), which can bias it in either direction. The paper's claim that only the logarithm survives the test thus rests on a constrained comparison and a pure-error estimate whose calibration is unclear. I would ask for the free-offset test on the actual per-seed widths, ideally with pure error from fresh seeds.
Questions for authors
-
In Fig. 4 (cosine), a run narrower than Table 4's minimum width for a similar N converges only after nearly twice the 40,000-iteration cap. What is the minimum width at twice the budget for the largest N? A binding cap would inflate the widths both weaknesses analyze.
-
Are input and output embeddings tied in GPT-2 and OLMo-3? Table 4 counts one embedding table, so its totals assume they are.
-
The theoretical constructions the paper contrasts with count parameters, not bits, and their parameter count grows only with the square root of the number of facts. Storing random facts, however, takes at least as many bits as the facts carry, and that amount grows with the number of facts. Even at 32-bit precision, the largest dataset would need about a hundred times more parameters than the square-root count, so any construction whose size grows more slowly than the number of facts must let the precision of each parameter grow with it. It would help to state the precision these bounds assume and to make the comparison in bits.
-
The structured-data and OLMo-3 results (Fig. 1, Table 6) compare widths at equal N, but symmetric and two-hop data carry less information than uniform data. Bits per total parameter, computed from each dataset's entropy, would show how fully the model exploits structure and how the two architectures compare.
7 comments
Reply to Reviewer izFw (2/4)
Weakness 2 (power-law with a free offset)
We thank the reviewer for the careful recomputation. We conduct additional analyses without the positivity constraint and find that our main claims remain supported. On our per-seed minimum dimensions, the power-law with a free offset is rejected on the 11 dataset sizes that we test in the submission (), as the reviewer found. To provide more statistical evidence for the comparison of the power-law and logarithmic fits, we added the dataset size (8 entities) to Table 1, Figure 1 and the Abstract of the revised version. With this dataset size, the power-law with a free offset is rejected with , and the logarithmic-with-offset fit is the only fit that is not rejected (). The power-law with a free offset is also rejected () when capacity is measured as the maximum memorizable dataset size at each dimension (Appendix E.1).
To make the role of the positivity constraint clearer for future readers, we added the unconstrained fit to Tables 1 and 9 of the revised version and explained the constraint in Section 4.1 and Appendix B.2. In Section 4.1 (Lines 361–363), we added "We require the fitted to be positive for every (positivity constraint) in every fit where this constraint applies. Without this constraint, the fitted power-law with an offset can become indistinguishable from the logarithmic fit (see Appendix B.2)." In Appendix B.2, we added "The power-law with an offset, , approaches as , so it contains the logarithm as a limiting case. On and on OLMo-3, the unconstrained fit drives the exponent to the lower bound of the search (), and the fitted power-law is then indistinguishable from the logarithm. The positivity constraint excludes these degenerate cases."
In order to make the computation of the pure error clearer, we revised Line 843 in Appendix B to state that the pure error is computed from the seeds of the selected hyperparameter configuration.
Question 1 (training budget)
We thank the reviewer for the question, and we agree that a larger budget can lower the minimum dimension. We could not measure the minimum dimension at twice the budget for due to its computational cost. In order to make this limitation clearer for future readers, we added the following to the Limitations and future work paragraph of the revised version (Lines 527–534). "We measure every minimum dimension and every maximum memorizable dataset size under the same budget of 40,000 training iterations, and a larger budget may change these measurements. At , a model of dimension 13 memorizes the dataset after about 75,000 iterations (Figure 4), while the average minimum dimension at under our budget is 14.33 (Table 4). We did not repeat the experiments with a larger budget due to their computational cost. Each measurement requires two 60-trial hyperparameter searches with three seeds. Under our budget, the two searches take 344 GPU-hours at and 779 GPU-hours at in the width experiment, and 88 GPU-hours at and 270 GPU-hours at in the depth experiment."
Question 2 (tied embeddings)
We thank the reviewer for the question. GPT-2 uses tied embeddings, so Table 4 counts its single embedding table once. In order to clarify this, we revised Line 247 in Section 3.2 with the additional phrase "shared between input and output". OLMo-3 uses untied embeddings, with a separate output layer of the same size. We revised Lines 802–806 in Appendix B.1 to state this difference, and Table C counts both OLMo-3 tables (Question 4).
Reply to Reviewer izFw (3/4)
Question 3 (theoretical bounds in bits)
We thank the reviewer for this point, and we agree that a construction with parameters requires a number of bits per parameter that grows with . The theoretical results we cite study a setting that differs from our setting. In order to make this clearer, we made two changes. First, we removed the comparisons between these results and our measurements. Specifically, we removed the clause "substantially worse than what theoretical analyses predict is achievable" from the Abstract, the clause "and are much better than the linear scaling observed in empirical studies" from the Introduction, and the comparison of with from Section 5. Second, in the Related Work section (Lines 146 and 156–158), we added the precision of the parameters to the assumptions that vary across settings, together with the following text. "The construction of Kajitsuka & Sato (2025) uses bits per parameter, and Madden et al. (2025) treat parameters as real numbers with unbounded precision. The memorization bounds above therefore do not predict the capacity of a trained transformer with a fixed precision, such as our models, whose parameters are stored in 32 bits."
Reply to Reviewer izFw (1/4)
We thank the reviewer for the careful and detailed review, and in particular for recomputing our results from the reported numbers. We checked the reviewer's computations on our per-seed data, and we report the results below together with the changes we made in the revised version. Please let us know if anything falls short or if any concerns remain. We would really appreciate the opportunity to continue to improve our paper and address additional concerns before the end of the discussion period. Changes in the revised PDF are highlighted in blue. Additional minor revisions are not highlighted, for clarity.
Weakness 1 (bits per parameter with one token per entity)
We thank the reviewer for this suggestion. As the reviewer requested, we added the Uniform-NL experiment (Sections 3.6 and 4.5 and Appendix E of the revised version). In this experiment, the entities and relations are real-world phrases that each consist of several tokens, and the vocabulary of tokens is fixed.
Each fact of Uniform-NL is written as a natural-language declarative sentence, for example "Shana Bald works at Tim Hortons", and tokenized with the GPT-2 tokenizer. The minimum transformer dimension needed to memorize a dataset of size N behaves similarly to the Uniform experiment and grows sublinearly with N. A lack-of-fit F-test rejects a linear fit (p=6.0×10−6, Table A). Since the revised version includes this experiment, we removed the sentence in the Limitations and future work paragraph that named natural-language facts as a next step.
Table A. Minimum transformer dimension on Uniform-NL (L=1, m=3, averaged over 3 seeds), with the Uniform result of Table 4 for comparison.
Moreover, we believe that we were not sufficiently clear in the submission version of our analysis of embedding and non-embedding parameters, and we rewrote this part of Section 5 in the revised version (starting with the paragraph "Embedding versus non-embedding parameters"). The rewritten part contains the following text.
"The number of embedding parameters does not have an effect on the transformer's information capacity. Consider the Uniform-NL experiment, where the embedding table has the 50,257 rows of the GPT-2 vocabulary at every dataset size. The datasets in this experiment contain at most 1,764 distinct tokens, so at least 48,493 of these rows correspond to tokens that do not occur in any fact. Therefore, the number of bits divided by the total number of parameters is inappropriate for measuring information capacity. Furthermore, we define the capacity of a transformer as the largest dataset size that it can memorize completely (Section 3.1), and we control for how the facts are embedded, which is a confounding factor."