acceptodds
Under review as a conference paper at ICLR 2027

Compartmentalization: Learning to Translate Is Not Learning to Transfer

Abstract

Language models learn from data presented in many surface forms; they train on the same facts in English and Swahili, the same functions in Python and Haskell. Ideally, what a model learns from one form should help it with others. We test a variant of this in a controlled setting by rendering the same text in _compartments_ that use disjoint vocabularies, each a token-for-token relabeling of the others. Models up to 42M parameters trained on 131B tokens often learn each compartment about as slowly as if trained on its share of the data alone, a failure we call _compartmentalization_. Models initialized with identical embeddings across compartments show no such cost, so a unified solution exists, but is not found from generic initialization. Paired examples, each showing the same text in two compartments, teach models to convert between compartments almost perfectly (99.7-100% exact match), but this reduces the sample efficiency cost only past a threshold determined by and the learning rate. Learning a mapping between surface forms, in other words, does not guarantee transfer between them.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
LLMs, language models, generalization, multi-task learning, transfer learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_QK67OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3(+1)
Soundness
3
Read review
Paper summary

The paper shows that language models trained on identical content across different pseudo-languages (created by duplicating English data), each language having its own disjoint vocabulary (no shared tokens between languages), fail to learn shared representations of underlying “concepts”. As a result, these models suffer from learning slowdown and model capacity cost compared to an optimal setting in which each concept is learned jointly across all pseudo-languages.

Critical strengths
  • The controlled experimental design allows for easily interpretable results (unlike standard multilingual training, where the underlying data mapping is only partially known).

  • In my opinion, the most interesting findings come from the experiments with paired data, which show that even a perfectly learned mapping from vocabulary A to vocabulary B does not necessarily mean shared representations.

Critical weaknesses
  • The implicit hypotheses are not clearly motivated in the paper, leaving it to the reader to understand why such model behavior should be expected. Given that the paper does not use the full nine-page limit, there would be room for further justification for the underlying hypotheses.

  • The paper often describes experimental findings but does not investigate them further, leaving some findings shallow. Therefore, the paper has somewhat limited novelty. For example, the paired data experiments show interesting behavior but do not study or discuss it further. E.g. may such behavior be caused by models with c < 8 having enough capacity to learn separate representations for each compartment, while increasing the number of compartments may create competition for limited model capacity, pushing the model to learn shared representations.

Questions for authors

None

2 comments

We thank the reviewer for suggestions that have helped materially improve our work, in particular for their interest in making hypotheses clearer.

We will denote new results since their initial review with "★\color{green}\bigstar", and rewritings with "Δ\color{orange}\Delta".

Weakness 1: implicit hypotheses

Δ\color{orange}\Delta We have provided further justification within the body of the paper, filling the original page limit. In particular, we stated hypotheses more explicitly:

  1. In the intro: for sufficiently complex data, a solution under multiple mere relabelings should be almost as simple as one with a single labeling.
  2. In section 5.3 (paired translation experiments): very little paired data should be needed to establish a sufficiently simple mapping.
  3. In section 5.2 (mismatched dataset capacity costs), we add two questions that explicitly motivate our choices of datasets.

Weakness 2: findings not investigated further (capacity explanation)

We view our results as consistent with a capacity-based hypothesis. In particular, weight decay creates capacity pressure, and lowers the transition threshold we observe from c=8c{=}8 to c=6c{=}6 at tr=0.75tr{=}0.75. ★\color{green}\bigstar As an additional result, we analyze alignment of the same token embedding across multiple vocabularies, and observe that it rises at tr=0.75tr{=}0.75 for every c∈{4,6,8}c\in\{4,6,8\}, but stays near 0 for c=2c{=}2. We suspect that there may be insufficient capacity pressure to discourage a redundant solution for c=2c{=}2, although we note that we have been unable to induce sharing at c=2c{=}2 for every variation of weight decay, learning rate, and translation ratio we have tried.


We believe the explicitly stated hypotheses address your first weakness, and the capacity-cost experiment, the learning-rate and weight-decay results, and the embedding-alignment analysis address your second. We welcome your insight on whether these analyses address your question about capacity. If this response doesV address your concerns, we would be grateful if you would consider raising your score.

I want to thank the authors for the response. These changes make the paper easier to follow so I have updated my presentation score accordingly. However, in my opinion the paper is still mostly descriptive, leaving it unclear how the findings will generalize to other settings, what are their implications to downstream behavior, and most importantly, what are the root causes of compartmentalization and how to overcome these limitations. Therefore, I will keep my original rating.

Reviewer_5RZoOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper studies a setting where transformer LMs fail to learn shared representations of underlying latent data when split across disjoint token sets, a problem which they term “comparmentalization”. They show that transformers can learn to share computation by a simple construction when the corresponding embeddings are duplicated at initialization, but that this solution is not often reached in practice. They also measure whether paired data can reduce compartmentalization and find that it helps only in some cases with a larger number of compartments.

Critical strengths

Overall, the experimental setup was easy to understand. The claims are moderate and supported by experimental evidence, and several limitations are acknowledged in the paper. The main compartmentalization experiment is simple and the results are clean.

Critical weaknesses

There are several places where the authors point to the appendix for additional details, but the submission is almost a full page short of the limit, so some of these details could be moved into the main paper to improve readability. Additionally, many of the figures appear fairly far away from the relevant text describing them and their setup, which makes it harder to follow. Some rearranging could improve the gap between figure location and discussion.

While the compartmentalization results seem to hold for the toy setting on small models, it is not clear that this same phenomena is true in pretrained LLMs with much more capacity and data. (This is noted by authors in their limitations section, but is something I think limits the contributions of the paper.) Providing intuition of a potential failure mode for large-scale LMs is helpful, but it's not clear that the particular experiment setting/design choices tested here will apply to large-scale models as there are many other factors of paired data/co-occurence that this paper does not consider.

Questions for authors

Do you have an intuition for why the c=2 case w/ two compartments of tokens did not seem to benefit from paired data? This was surprising to me.

2 comments

We thank the reviewer, in particular for their question about c=2c{=}2, which we address in the final section below. We will denote new results since their initial review with "★\color{green}\bigstar", and rewritings with "Δ\color{orange}\Delta".

Weakness 1: appendix details and figure placement

Δ\color{orange}\Delta We have moved the figure corresponding to the parameter-matched control into the main body. We believe this is a valuable result to highlight because it means that the gap we measure is not merely due to the difficulty of optimizing more parameters, and our paper now fills the original 9-page limit.

Δ\color{orange}\Delta We have reorganized figures so that they are on the same page as their discussion or on the facing page, except for Figure 6, because Figure 5 also discusses the paired data results.

Weakness 2: relevance to large pretrained LMs

★\color{green}\bigstar We strengthen our bilingual results (section 5.4) with a new set of models we trained after submission: 1.4B parameter models on 55B tokens of text, about 1.9x the Chinchilla-optimal count. In particular, we reproduce the same finding at this scale: a shared vocabulary provides no benefit over completely disjoint vocabularies, with a consistent gap to single-language performance. Our evidence provides a new characterization that, at the largest scale we can afford to train on our compute resources, there is not yet any measurable multilingual benefit from a shared vocabulary, i.e., that within our framework, the solution the model learns is fully "compartmentalized".

Δ\color{orange}\Delta Additionally, in our abstract we more explicitly scope our findings: "in our setting," ending by offering our results "as a preliminary methodological caution for future experiments at scale."

Question 1: why c = 2 does not benefit from paired data

We share QK67's intuition that this is due in part to insufficient capacity pressure to overcome the inefficient solution; we hypothesize that at higher cc, separate copies become too expensive. Supporting this characterization, our weight decay results lower the typical (c,tr)(c,tr) required of a unified solution. We note that at c=2c{=}2, there is no combination (tr,lr)(tr,lr) we have tried that has successfully induced sharing.


We believe this resolves the figure placement and the page budget, and clarifies both the question of scale and the c=2c{=}2 result. Given that you find the setup easy to understand and the results clean, we would be grateful if you would consider raising your score. We look forward to your additional insights, and in particular we'd welcome hearing what, if anything, stands between your position now and a higher one.

Thank you to the authors for the response. I have updated the soundness score of my review to a 3 and will maintain my positive rating. While the scaling of the experiment to a larger model size is nice, the toy nature of the setup is the reason I plan on keeping keep my score at a weak accept, as I think real language data is more messy and complex than the exact pairs data setting you've introduced in this paper, and this is what I'm not sure will transfer. One reason for this is because other research has found that LLMs do learn shared features across languages (e.g., Brinkmann et al., but also others), and others have found that imbalanced data is better for multilingual generalization (Schafer et al., which you cite in your paper as using "similar techniques"), so it may be worth thinking about how messier language data and alternate forms of co-occurrence, other mixtures/ratios might affect your finding of compartmentalization.


Brinkmann et al., Large Language Models Share Representations of Latent Grammatical Concepts Across Typologically Diverse Languages.

Schäfer et al., The Role of Language Imbalance in Cross-lingual Generalisation: Insights from Cloned Language Experiments

Reviewer_VxVwOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies a controlled form of failure to share representations across surface forms. It renders the same text in cc “compartments” that use disjoint vocabularies, each a token-for-token relabeling of the others, by offsetting token IDs by VjVj. Models are GPT-style decoders trained on ~131B tokens of FineWeb at scales from 1.1M to 42M parameters (c=1), with cc up to 8. The authors report that models often learn each compartment about as slowly as if trained on that compartment’s share of the data alone (“compartmentalization”), with a validation-loss gap that grows with cc and persists under continued training. They show that copying embeddings across compartments at initialization, or duplicating a trained c=1 model’s embeddings and LM-head rows post hoc, recovers the c=1 baseline—so a unified solution exists but is not found from generic initialization. Paired “translation” examples teach the model to convert between compartments almost perfectly (99.7–100% exact match), but do not reduce the sample-efficiency cost until a threshold determined by cc, the paired-data ratio, and the learning rate; weight decay accelerates the threshold. A bilingual English–Chinese experiment provides preliminary evidence that a shared tokenizer barely helps relative to disjoint vocabularies.

Critical strengths
  1. The vocabulary-offset trick isolates surface-form disjointness from subword overlap, syntax, and semantics. This is exactly the kind of controlled experiment that is needed to study transfer, and the paper executes it carefully.
  2. The parameter-matched control (same enlarged embedding table, trained like c=1) rules out a simple parameter-count explanation. The initialization-copy and post-hoc-duplication experiments are strong existence proofs: a unified solution is reachable and cheap to recover, but generic training does not find it.
  3. The finding that near-perfect pair-completion does not imply transfer is important for multilingual and multi-format pretraining, and it is stated with appropriate caveats.
  4. Honest limitations. The limitations section acknowledges short context, small scale, exact token correspondence, single-seed results, and the preliminary nature of the bilingual experiment. This is refreshing.
Critical weaknesses
  1. The limitations section acknowledges that the model-scale sweep, the bilingual experiments, and the weight-decay sweep are single-seed. The main cc-sweep in Figure 1 also lacks reported confidence intervals. Given that the paper’s claims are about the presence and persistence of a gap, and that the phase transition is seed-sensitive near the threshold, multi-seed estimates with CIs are needed, especially for the scale sweep that supports “across models from 1.1M to 41.9M.”
  2. The setup is highly synthetic, and the bilingual evidence is preliminary. Exact token-for-token correspondence, disjoint vocabularies, block size 64, and small models are acknowledged limitations, but the title’s broad claim (“learning to translate is not learning to transfer”) is supported only in this controlled setting. The bilingual experiment uses 767M tokens per language, a single seed, and 5k steps; the gap between shared and compartmentalized tokenizers “closes quickly with steps” (Figure 6b), so the conclusion that a shared tokenizer “barely improves” may be transient. The abstract should scope the claim to the controlled setting and present the bilingual result as preliminary.
  3. Compartment embeddings change the setup in Section 4.3. The paired-data experiments add a learned compartment embedding to remove ambiguity about compartment identity. It is not clear whether the tr=0 control also uses these embeddings, and whether the paired-data conclusions would hold without them. Since the compartment embedding gives the model an explicit compartment ID, it may make separate solutions easier and transfer harder (or vice versa); this needs an ablation.
Questions for authors

see weaknesses

2 comments

We thank the reviewer for three points that have helped shape our work. We will denote new results since their initial review with "★\color{green}\bigstar", and rewritings with "Δ\color{orange}\Delta".

Weakness 1: seeds and confidence intervals

We trained many of our experiments with three seeds, as permitted by our limited compute. ★\color{green}\bigstar We refer the reviewer to Tables 2–6 in Appendix E, where we list seed variance for runs where we perform multiple seeds. Most results vary by about 0.005 nats, except near the phase transition, where seeds differ by up to 0.24 nats. We highlight in the paper that the phase transition is seed-sensitive, with seed agreement away from the transition. We acknowledge that we do not have the compute required to train multiple seeds for every condition, including parts of the model scale sweep, and list the single-seed conditions in Table 6 in the appendix, but note that for most such runs, comparable multi-seed conditions exist with negligible variance, i.e., that most of our experiments are in seed-insensitive regions. An important exception is the weight decay results, which we note in Table 6.

Weakness 2: synthetic setting and preliminary bilingual evidence

Δ\color{orange}\Delta We improved our abstract: we state that the strongest claim holds "in our setting," and add a new ending: "we offer this as a preliminary methodological caution for future experiments at scale." We state the remaining limitations in our limitations section.

★\color{green}\bigstar After our initial submission, we trained 1.4B parameter language models on 55B tokens, or about 1.9x the Chinchilla-optimal number of tokens for the parameter count. These models support our updated multilingual comparison, which provides a similar characterization to our previous one, at greater scale.

Additionally, we note that the quickly closing gap the reviewer refers to is our result precisely; the persistent gap of concern is the one between our multilingual and monolingual conditions. Δ\color{orange}\Delta We have improved clarity about these conditions in our updated draft.

Δ\color{orange}\Delta We label the new, stronger multilingual run a "preliminary extension," and clearly list that it is single-seed due to our compute limitations.

Weakness 3: compartment embeddings and the tr = 0 control

Δ\color{orange}\Delta We have updated section 5.3 to clarify that compartment embeddings are applied to every condition, including those without paired data.

★\color{green}\bigstar In addition, we performed an ablation (Appendix B.2, Figure 8), training a new variant at the 14.7M scale without compartment embeddings for 300k steps, for c∈{2,8}c \in \{2, 8\} and tr∈{0,0.5}tr \in \{0, 0.5\}. The conditions track each other closely except at c=8c{=}8 and tr=0.5tr{=}0.5, where the model converges faster without compartment embeddings. Our paired-data conclusions do not appear to depend on compartment embeddings.


We believe this serves to address your concern about seed robustness, better scopes our claims to the setting we study, and resolves the question of compartment embeddings in the tr=0tr{=}0 control. If this response addresses the weaknesses you raised, we would be grateful if you would consider raising your score. Otherwise, if any concerns remain, we welcome your feedback so we can address it during the discussion phase.

Thank you for the feedback, I'll keep my score unchanged.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.