Data-centric analysis of compositional generalization in multimodal models

Investigate, from a data-centric perspective, how multimodal models can understand and generalize to novel compositions of concepts not encountered during pretraining, thereby characterizing and improving compositional generalization distinct from traditional zero-shot learning.

Background

The paper distinguishes compositional generalization—understanding new combinations of concepts—from conventional zero-shot learning and highlights it as an important yet unresolved area.

They explicitly call out the need to analyze compositional generalization through data-centric methods, suggesting that current approaches and empirical understanding are insufficient.

References

This is distinct from traditional zero-shot learning and presents an intriguing, yet unresolved challenge: analyzing compositional generalization from a data-centric perspective.

— No "Zero-Shot" Without Exponential Data: Pretraining Concept Frequency Determines Multimodal Model Performance  (2404.04125 - Udandarao et al., 2024) in Section: Conclusions and Open Problems

Whether these methods also raise EICS is an open question that our released code makes straightforward to test.

— When VLMs Destroy What They Know: Spatial Representations Fall Below Lexical Priors  (2610.13910 - Garg et al., 9 Oct 2026) in Section 2, Related Work, paragraph “Training-time compositional methods”

The experiment cannot show that omission caused the regression, that the internal rule was erased, or that adding composition examples would necessarily repair it.

— Learning When to Withdraw: Curriculum Coverage and Rule-Specific Failures in Fine-Tuned Language Agents  (2610.21045 - Abouzahra, 15 Oct 2026) in Section 5.2, “Composition-Specific Narrowing Regresses”