- Rating
- 2
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper trains a zoo of 300 GPT-2 models (mini, small and medium) trained on different data mixtures of 5 distinct corporas. The models are pretrained for 85% of the training and then finetuned on the different data mixtures for the last 15%. They provide an analysis into the diversity of the zoo, finding that models trained on the same mixtures are more similar than the models trained on different mixtures. Lastly they use the generation procedure from DeepWeightFlow to show the generation performance by training from this zoo.
Critical strengths
They provide a model zoo of large language models which to the best of my knowledge is a novel contribution to the scientific community and helps expand the field of weight space learning into large language models.
The authors show that a generative approach trained on this zoo can generate models conditioned on data mixtures not seen during training. With the generation outperforming nearest neighbour retrieval baseline.
Critical weaknesses
The paper is essentially training a model zoo, analysing it and using the generative approach from Deepweightflow to generate some models. And whilst the releasing of such a model zoo is valuable to the community, the analysis of the models merely confirmed that models trained on the same data were similar and models trained on different data mixtures are different, which given the shared training trajectory is not surprising. I do not find this scientific contribution substantial enough for ICLR.
Beyond that there are a number of issues in the generalization claims of the contributions. As stated in the paper the Gram-PCA, the weight encoder, is fitted on the full model zoo. Even the models later used for the evaluations. I understand that the generative model is not trained on the evaluation data however this is still a substantial benefit to the approach.
Other issues:
Shared pretrained trunk. The model zoos share a pretrained backbone, referred to as “trunk”. Effectively decomposing the model zoo into a Model Tree, where the fine-tuned models are just 15% of the total training from this trunk. This calls into question the diversity of the model tree in comparison to using independently trained models.
Generalization setting. The experimental protocol in this setup is restricted to 5 different corporas, which are used both for training and evaluation. So while this can show generalization to different data mixtures it limits the generalization claims of the paper. It's also unclear why this setup was chosen instead of using unseen corporas during evaluation.
C2ST. In section 4.3 the authors refer to Appendix B.8 in regards to identifying the requested mixture. And in Table 28 you have a table with a classifier two-sample test (C2ST) which is showing an accuracy of 98%. However it is explained that 50% accuracy represents that the generated and true distributions are truly indistinguishable. Shouldn't this then be the aim of the generative approach, and thus the numbers presented in Table 28 show that the approach is not close to that? I think the author remarks on that on line 431, however no further explanation is provided on the seemingly poor results?
Evaluation. Gold-continuation log score is used to evaluate the performance of the generated models in Table 5 but is not explained in the main text. Thus it is now clear how to interpret the results and how to compare them. Further regarding the accuracy presented in Table 32 in the Appendix, it seems to be showing barely above chance results for the generated models?
Questions for authors
-
How does the generation performance change when the Gram-PCA is only fitted on the training models?
-
How would the performance change if the model zoo did not share a trunk but was effectively independently trained models?
-
Could you explain the C2ST result more, in reference to the aforementioned weakness?
-
Why are both the real and generated models barely above chance in terms of the accuracy in Table 32?
4 comments
Response to Reviewer x5YX - Part 2
Other issue 2. Five fixed corpora
You have identified an exciting direction for future work, which this project enables. Our experiments establish generalization to unseen mixtures of five fixed corpora, not to unseen corpora themselves. Generalization to held-out datasets has been studied in vision weight generation, while generation across pretraining-corpus mixtures is unexplored. Modern LLMs are pretrained on mixtures, which makes the mixture the natural conditioning variable. We make this limitation of our claims clear in the revised text.
Other issue 3. C2ST
We agree that 0.50 is the target. Rechecking that number, we found that the submitted 0.976 was not informative, and we apologize for the error. That design scored 100 real codes against 4,000 generated ones, and 4000/4100 = 0.976 is exactly the majority-class rate. The classifier was reporting the base rate, not necessarily learning. The revision makes three changes (App. B.12, Table 33):
- a balanced design (100 real vs 100 generated, 50 draws);
- the classifier's cross-validation split grouped by held-out member, because a row-wise split puts rows that share a and pushes even a real-vs-real control to 0.35–0.39;
- a real-vs-real control.
We report two feature sets. The first is basis-free: each model's five-domain log-perplexity profile together with . The second is the leading eight codes in a single all-member basis to fix the measurement frame. No generator is trained in or decoded from that basis alone.
| C2ST accuracy (0.50 = indistinguishable) | Mini | Small | Medium |
|---|---|---|---|
| Function space: flow | 0.650 | 0.681 | 0.703 |
| Function space: rank-50 projection of the real target | 0.654 | 0.714 | 0.786 |
| Function space: real-vs-real control | 0.508 | 0.505 | 0.516 |
| Measurement frame : flow | 0.568 | 0.566 | 0.566 |
| Measurement frame: real-vs-real control | 0.498 | 0.488 | 0.472 |
In function space, generated models remain detectable at 0.65–0.70. In the code frame they are close to indistinguishable. However, the exact rank-50 projection of the real target, with no generator involved, is equally or more detectable (0.65 / 0.71 / 0.79). Most of the detectable signal therefore comes from rank truncation. The revision states this directly.
Other issue 4. Gold-continuation log score and near-chance accuracy
We have defined the gold-continuation log score in the main text and the Table 5 caption. It is the log-probability the model assigns to the correct HellaSwag ending, length-normalized (mean units per ending token), over 2,000 questions and paired between real and generated models. Unlike accuracy, it is continuous and does not saturate at chance, so it separates models that accuracy cannot.
Accuracy is near chance for real and generated models alike (Table 37, formerly Table 32): 0.252 / 0.257 / 0.273 for real interior members, against 0.25 chance. This is expected at our token budget. At 20 tokens per parameter (Chinchilla-optimal; Hoffmann et al., 2022) the zoo members see about 1.0B / 2.5B / 7.1B tokens. Small models with non-trivial HellaSwag accuracy are trained far beyond this. For example, SmolLM-135M (Allal et al., 2024) is trained on more than 50 times our token budget and reaches about 42%. Accuracy rises with scale across our zoo as expected.
On the log score, generated models score slightly higher than their real targets on the interior (Table 5). We attribute this to shrinkage: generated codes are pulled toward the conditional mean of nearby members, which acts like weight averaging. Averaging same-mixture models improves on both endpoints in 19–20 of 20 pairs (Table 17).
Q2. Performance without a shared trunk
We do not have data to answer this directly, and this gap is precisely why we started from a controlled population. We would expect some degradation but have no basis to quantify it. Our evidence is:
(1) The Mini sweep (Table 14), making 60% of training mixture-specific lowers the between/within-mixture distance ratio from 4.08 to 2.88, but mixture structure clearly survives.
(2) Independently trained networks are thought to be linearly connected only modulo permutation (Entezari et al., 2022; Ainsworth et al., 2023), so a trunk-free zoo would likely need re-basining before Gram-PCA, as in prior weight-generation work (Gupta et al., 2026).
W1: Scientific contribution. This paper is not proposing a new generation approach, but rather a new model zoo which you show, by using DeepWeightFlow (Gupta et al., 2026), can be used to train a generative model. However the authors are correct in that mixture-conditioned generation has not been used in this field before.
W2: Gram-PCA fit on the full zoo. I appreciate this change, however I do not understand the use of the wording inductive bias, shouldn't this change just amount to a proper train-test split?
Other issue 1. Shared trunk / model tree. I understand that this design is a deliberate control, however that weakens the claim of mixture conditionied generation of full pretrained LLMs. As this only applies to a low diversity zoo.
Other issue 3. C2ST I welcome the new C2ST results, which look like what i would expect for the generation in this low diversity regime.
Other issue 4. Gold-continuation log score and near-chance accuracy. If the real model are also at chance, why even use them?
Q2. Performance without a shared trunk I thank the authors for this answer and understand the need to start from a controlled population.
Nevertheless I remain unconvinced and will maintain my score.
Response to Reviewer x5YX
We thank the reviewer for continuing to engage with our work, and we respond to the remaining concerns they identify below.
Novelty:
The reviewer is correct that we do not propose a new generator. We see three contributions. The first is the zoo itself: a controlled population of pretrained general-purpose language models in which the pretraining mixture is the only varied factor. The second is mixture-conditioned generation of complete pretrained LM weights, which is, to our knowledge, the first full-weight generation of pretrained general-purpose language models, at scales up to 355M parameters. The third is the evaluation protocol. Its inductive LOMO design withholds the target model from the Gram-PCA fit as well as from flow training, which departs from Gupta et al. (2026). CRPS skill is a proper score that a point mass cannot optimize. The per-cell retrieval control is something a copying generator can at best tie. We regard the protocol as a durable contribution independent of the generator, since it applies to any future weight generator evaluated on this zoo.
On terminology, the reviewer is right: the inductive basis amounts to a proper train-test split for the latent representation. “Bias” would be a typo, and we have checked that it does not persist. We were simply unsure of what wording would be clearest, and are open to suggestions.
Shared trunk and scope of the claims:
We agree that the shared trunk defines the scope of our generation claims, and our revised manuscript scopes our claims in this light. Section 3.2 states that all members share architecture, optimizer, schedule, token budget, and initialization, with only and a branch seed varying. A.5 discusses the regime of independently trained runs, where canonicalization would become necessary and which our results do not cover. The abstract and conclusion restrict the generalization claim to the interior of the mixture simplex within this population.
Within that scope, we believe the population is structured rather than degenerate. Once the drift inherited from the shared trunk is removed, cosines between branch deltas from different mixtures fall to roughly −0.14 to +0.17 (Table 1). Replicates of the same mixture stay strongly aligned (0.80 to 0.89). Functionally, members trained on different mixtures disagree on next-token predictions about twice as often as replicates do (Table 3). The variance decomposition $\eta^2_{\text{corr}}=0.82 \text{ to 0.92} $ is measured at the simplex vertices and bounds the influence of from above. On the interior, a quadratic surface on explains about 22 to 32% of held-out branch-delta variance (Table 24).
Trunk sharing is also a controlled design parameter – not a hidden confound. The branching fraction sets how much of training is shared. Our sweep (Table 14) shows displacement from the trunk rising from 16.5% to 55.9% as goes from 0.15 to 0.60. Over the same range, the between/within-mixture distance ratio falls from 4.08 to 2.88. We chose as the smallest value that passed our domain-separation checks .
Why report HellaSwag:
We agree that accuracy alone is uninformative at these scales, which is why we report the gold-continuation log score as the primary benchmark measure (Table 5). Accuracy is quantized at 1/2000 and sits near chance for real and generated models alike, so on its own it cannot distinguish parity from a failure to measure (Appendix B.17). The log score is continuous and paired question by question, and it separates models even where the argmax does not move. It also tracks scale cleanly, from about −5.2 at Mini to −4.1 at Medium, which indicates it carries signal.
On this measure, generated models match or slightly exceed the real models at the same mixture in every stratum (Table 5). On the interior, accuracy agrees, with no stratum differing by more than one question (Table 37). The comparison of interest is generated versus real at the same mixture, not absolute capability. Our primary functional measure remains across five held-out domains (Table 4).
Regarding the models’ usefulness more broadly: they are intended as a population for studying weight space, not for deployment. They follow the Chinchilla 20:1 token-to-parameter ratio (Hoffmann et al., 2022). Small models trained for downstream utility, such as SmolLM-135M, use substantially more tokens (discussed around Table 37).
Response to Reviewer x5YX - Part 1
We thank the reviewer for an expert and candid review. The basis concern, shared by the other reviewers, led to the largest change in the revision. The C2ST and evaluation questions led to clearer reporting.
Weakness 1. Scientific contribution
We agree that "same mixture$\rightarrow$similar, different mixture$\rightarrow$dissimilar" is expected, and we do not present it as the finding. Its role is a sanity check that the zoo is valid. We make three interconnected contributions:
Mixture-conditioned generation of full pretrained LM weights. Data or dataset-conditioned weight generation has so far targeted image classifiers or LoRA adapters (e.g., Soro et al., 2025b; Schürholt et al., 2022a; LoRA hypernetworks such as Charakorn et al., 2025 among others discussed in the Introduction). Full-weight generation for language models has targeted a single supervised task (BERTs on the Yelp task, Gupta et al., 2026). We generate complete pretrained LMs conditioned on their pretraining mixture, at up to 355M parameters, and beat every retrieval, interpolation, regression, and GP baseline in like settings.
Quantitative structures A shared trunk predicts high raw cosine similarity, which we find (∼0.5 overall). It does not predict how that similarity is organized:
A controlled LLM zoo with an evaluation protocol (grouped LOMO, proper scoring, replicate and truncation floors, data-copying tests). As you note, no such population existed for weight-space learning.
Weakness 2. Gram-PCA fit on the full zoo
We agree; this was the most important point in the reviews. Every generative result in our revised and reuploaded manuscript uses an inductive basis (Sec. 3.3, App. B.1). For a fold targeting θi at πi, θi is excluded from basis fitting, code normalization, flow training, and decoding of generated latents. For a held-out vertex, all four replicates are excluded. Under this protocol the flow obtains 3.78 / 4.11 / 4.70 %ΔPPL against 5.75 / 5.79 / 6.00 for retrieval. The flow beats every baseline (Tables 4, 23). We also scoped the claim to the interior of Δ4, where it holds in every case we have examined.
Other issue 1. Shared trunk / model tree
We agree that the zoo is a model tree, and that a trunk reduces diversity relative to independent pretraining. We would like to make two clarifications:
(1) The branches are not fine-tuned on a task. They continue the same pretraining objective with the same schedule (no warmup restart), differing only in data mixture over the final 15%. This mirrors how data mixtures are changed late in pretraining in practice (mid-training and annealing; e.g., OLMo et al., 2024; Blakeney et al., 2024).
(2) The design is a deliberate control. It isolates mixture-induced variation from initialization and permutation effects (barrier-free interpolation on all scored pairs, Table 17). It is also a setting where simple baselines such as retrieval and interpolation should be competitive, which makes the flow's margin over them more informative. The conclusion frames the result as holding in a controlled environment.
Branch-Train-Merge (Li et al., 2022) uses a related trunk-then-branch design for expert LMs but does not study weight-space structure or generation.