acceptodds
Withdrawn from ICLR 2027

Where Does Augmentation Repair Compositional Generalization? The Locus Shifts with Compression

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Augmentation is known to repair compositional generalization failures in neural decoders, but what it repairs, and where, is less clear. In a controlled toy regime we establish a full mechanism including a sharp repair-locus transition; in more realistic settings we find a weaker but consistent regularity. We study this in a controlled setting: a compressed (superposed) representation is decoded by a serialized language readout, and augmentation is applied under pre-registered, budget-matched protocols (1000+ trained models, all judgments fixed before execution). Four results emerge. (R1) Failure is a content-addressing mismatch. Probes recover the information (separation ratio ) while slot-level diagnostics locate the collapse in the object slots (0.49–0.74), with the relation slot substantially more robust (0.77–0.96); the pattern holds in both a toy regime and a semi-synthetic bridge. (R2) Repair is unit-local and thresholded. Augmenting a combination rebuilds that combination's readout pathway: 300 examples per unit raise covered units to 0.70–1.00 (median 0.93) while 75 per unit leave accuracy near chance (), and no transfer to unaugmented units occurs. Coverage, not augmentation content, is therefore the driver of repair: mechanism-guided selection is never better than random in the toy regime, and uniform coverage dominates mechanism-guided selection in the bridge (both pre-registered, 5 seeds). (R3) Compression deepens failure (0.000 at versus 0.195 with no compression) and shapes repair efficiency non-monotonically: mid-compression (–) repairs slowest. (R4) In the toy regime the repair locus depends on compression. At the readout pathway alone is sufficient (freezing the encoder still repairs covered units, , 5/5 seeds); at encoder reorganization is required (freezing costs 0.52, 0/5). A seven-point frozen-encoder scan shows the transition is continuous (crossover between and ), and R3's slow-repair plateau disappears when the encoder is frozen: the plateau is a property of encoder reorganization, not of the readout. A first-order account based on Cover's linear-separability capacity predicts the transition band from first principles, with no fitted parameters. On continuous features the locus does not transfer naively (a random frozen code fails by 0.56), while freezing a trained encoder with decoder-only augmentation recovers most of the repair (0.62 vs. 0.75) at lower cost. Under end-to-end training of a small CNN the failure pattern and the weak form both reproduce: unseen accuracy is 0.020 at baseline with object slots at 0.39 against a relation slot at 1.00, and freezing the trained encoder recovers 88.5% of the repair gain (0.539 vs. 0.606). Throughout, geometry changes globally under augmentation while behavior repairs locally: geometric reorganization is incidental at mild compression and is the repair itself at strong compression, a locality that slot-level measurement reveals and geometric analysis cannot.

Primary area:
interpretability and explainable AI
Keywords:
compositional generalization, data augmentation, mechanistic interpretability, neural language models, representation geometry

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_6j6ROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies how augmentation repairs compositional generalization failures and how compression affects whether repair occurs in the encoder or decoder.

Critical strengths

The experiments go beyond showing that augmentation helps. Sections 4.2–4.5 vary coverage, examples per combination, compression, and encoder freezing to examine when and where repair occurs. The robustness checks are informative. The appendix tests input noise, an MLP encoder, and a Transformer decoder, and reports where the original findings fail to reproduce. The frozen-encoder results have practical value. Decoder-only adaptation recovers about 79% of the repair gain on the continuous-feature bridge and 88.5% with the end-to-end CNN.

Critical weaknesses
  1. Proposition 3 misuses Cover’s theorem. Section 4.6 turns the fraction of separable binary labelings into the probability that one particular code is separable from the other eleven. These are different quantities.

  2. The necessity claim is stronger than the freezing results support. At C=4C=4, frozen and trainable encoders reach 0.169 and 0.691 accuracy, respectively. This is a substantial gap, but the actual decoder is a nonlinear LSTM. Linear inseparability in the input code space does not establish that decoder-only repair is impossible; the gap could also reflect optimization or capacity limits.

  3. The headline transition is sensitive to the setup. Appendix C reports that adding noise with σ=0.1\sigma=0.1 removes the crossover, while the Transformer decoder never shows freezing outperforming full training. The continuous-feature compression scan also produces a roughly constant freezing deficit. These results narrow the mechanism considerably beyond what the title and “full mechanism” language suggest.

Questions for authors
  1. Can Proposition 3 be corrected to address one-versus-rest separation, with explicit assumptions on the code distribution?

  2. Does the C=4C=4 freezing gap persist with longer training, tuned optimization, and a stronger decoder?

Reviewer_N7QkOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The submission seeks to mechanistically understand how data augmentation enables compositional generalization. It primarily studies a synthetic setting in which a scene composed of different geometric objects in different locations needs to be mapped to a linguistic description and uses probing experiments to identify in which parts of the model the failure arises.

Critical strengths

None

Critical weaknesses
  • The paper has severe presentation issues with jargon-heavy language making it difficult to comprehend. The abstract is a clear example of this issue, spanning almost the entire first page and reporting detailed numbers without providing the necessary context that would be required to understand it. The results section is barely readable, with incomplete, "stream-of-thought" like sentences.
  • The experiments are conducted in small, simplistic settings and even within those confines the manuscript itself states that the main claim "The locus shifts with compression" does not consistently arise (L451: "The locus shift itself stays bounded to the toy regime").
Questions for authors
  • Compositional generalization is a well-studied topic with a rich literature that the paper fails to cite and contextualize. How do the findings in this submission relate to established findings of when compositional generalization can succeed and why it sometimes fails?
Reviewer_bEA5OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how data augmentation improves compositional generalization. The authors find that augmentation mainly improves combinations seen during augmentation. They also find that compression affects where the repair happens. Under mild compression, updating the decoder is sufficient. Under strong compression, updating the encoder becomes important. The authors use Cover's theorem to explain these results and test their findings in several settings.

Critical strengths

-The paper investigates where data augmentation repairs compositional generalization failures, focusing on the roles of the encoder and decoder. This is a meaningful research question.

-The authors conduct controlled experiments by varying encoder dimensions, augmentation budgets, and freezing strategies, with multiple random seeds.The authors conduct controlled experiments by varying encoder dimensions, augmentation budgets, and freezing strategies, with multiple random seeds.

-The paper includes experiments with semi-synthetic visual features, an end-to-end CNN, and a Transformer decoder, while also reporting negative results.

Critical weaknesses

-Uselis et al. (ICML 2026, How Can Embedding Models Bind Concepts?) have already investigated recoverable concept information, binding failures, and the role of data coverage in compositional generalization. The additional mechanistic contributions of R1 and R2 relative to such prior work remain unclear.

-The Cover-based analysis concerns random dichotomy separability rather than the one-vs-rest addressability required by the task. Moreover, the measured separable fraction exceeds the claimed theoretical upper bound: at C=4, 1−ρ=0.60 vs. P(4)=0.27; at C=8, 1.00 vs. 0.97. The analysis also overlooks the additive color–shape structure, under which a generic linear projection with C≥5 preserves all 12 convex-hull vertices before ReLU, raising questions about the claimed C=7–8 transition.

-High probe separability may be expected given the simple one-hot inputs and limited attribute vocabulary. Similarly, lower object-slot accuracy may directly follow from holding out object combinations while keeping relation categories observed. These observations do not uniquely establish content-addressing mismatch or readout pathway reconstruction.

-The repair-locus crossover disappears when Gaussian noise is introduced and does not consistently reproduce with continuous visual features or a Transformer decoder. This raises concerns about whether the proposed transition reflects a general mechanism or a property of the specific toy setting.

Questions for authors

Q1: How do the proposed content-addressing mismatch and readout pathway reconstruction differ from the concept-binding and coverage-based explanations in Uselis et al. (ICML 2026)?

Q2: How can Cover's random-dichotomy formula justify one-vs-rest addressability? How do the authors reconcile the empirical measurements with the claimed upper bound and explain the C=7–8 transition given the additive structure of the toy data?

Q3: Can the authors provide evidence distinguishing content-addressing mismatch from conventional OOD generalization failure, including clarification of the probe training protocol and appropriate baselines?

Q4: Under what conditions does the compression-dependent repair-locus transition hold, and how should the central claims be qualified given the negative results beyond the original toy setting?

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

Related papers

Discussion (0)

Sign in to comment.