Systematic Generalization to Novel Concept Combinations

This direction is merged into another direction: compositional generalization testing. Its last version stays readable and citable here.

Version 1 of 153 papers53 conceptsfrozen 2026-10-04

Humans possess a remarkable "algebraic" ability: once we learn the meaning of the word "jump" and the modifier "twice," we instantly understand what it means to "jump twice," even if we have never seen those two words paired together. This ability to recombine known primitives into an infinite variety of novel, predictable structures is known as compositional generalization.

While modern neural networks are world-class pattern matchers, research across natural language, computer vision, and robotics suggests they struggle with this fundamental building block of intelligence. Instead of learning universal rules, models often rely on "shortcuts"—memorizing superficial correlations found in their training data.

The Algebra of Language and the Negation Wall

The core of the problem lies in the difference between statistical interpolation and systematicity. In Still not systematic after all these years: On the compositional skills of sequence-to-sequence recurrent networks, Brenden M. Lake and Marco Baroni introduced the SCAN dataset to prove that standard recurrent networks fail catastrophically when asked to apply a simple rule to a new primitive. If a model is trained on "jump" in isolation but sees "walk twice" and "run twice," it often fails to output the correct sequence for "jump twice."

This isn't just a quirk of language; it's a failure of symbolic logic. Anna Langedijk and Willem Zuidema describe this as a "negation wall." Even if a model understands an operator like "NOT," it may fail to generalize that operator to unseen combinations, suggesting that standard architectures lack the inductive biases to treat operations as modular, "plug-and-play" components. Interestingly, this ability is not unique to humans; Evidence for compositionality in baboons (Papio papio) through the test case of negation shows that non-human primates can treat visual markers as abstract operators, a feat current AI still finds challenging.

Testing for True Understanding: The Combinatorial Holdout

To distinguish between true reasoning and sophisticated mimicry, researchers use combinatorial out-of-distribution (OOD) evaluation. The intuition is simple: if you want to know if a model understands "red" and "cube," you must ensure it never sees a "red cube" during training, even though it sees "red spheres" and "blue cubes."

Several benchmarks have been developed to enforce this rigor:

The Scaling Mirage

A common counter-argument is that "scaling up" models on more data will eventually solve compositionality. However, recent evidence suggests that scaling may simply be hiding the problem. Chengshuai Zhao argues in Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens that Chain-of-Thought (CoT) reasoning is often a "brittle mirage" that collapses as soon as a query drifts away from the training distribution.

Similarly, in video generation, Yue Yang and Rui Lu 0001 found that while models like Sora produce visually stunning results, they are "case-based" reasoners rather than true world models. As shown in How Far is Video Generation from World Model: A Physical Law Perspective, these models prioritize visual attributes like color over fundamental physical laws, failing to generalize when object counts or spatial arrangements deviate from their training "cases."

Why Models Fail: Representational Incompatibility

If scaling isn't the silver bullet, what is the architectural bottleneck? Róbert Csordás and Kazuki Irie suggest a structural cause in CTL++: Evaluating Generalization on Never-Seen Compositional Patterns of Known Functions, and Compatibility of Neural Representations. They found that neural networks often develop "incompatible" internal representations for the same symbol depending on the context in which it appears.

For a model to generalize compositionally, its representation of "blue" must be the same whether it is modifying a "car," a "bird," or a "concept." When these representations are entangled with their context, the model cannot "unplug" the concept of "blue" and "plug" it into a novel scenario. This is why A. Michael Carrell and Ian Berlot-Attwell argue in Attribute Diversity Determines the Systematicity Gap in VQA that increasing the diversity of attribute combinations is more important than increasing the total volume of data.

Moving Toward Productive Reasoning

To move beyond "stochastic parrots," the field is shifting toward evaluations that are impossible to "cheat" via memorization. Dingli Yu and Anirudh Goyal proposed SKILL-MIX: a Flexible and Expandable Family of Evaluations for AI Models, which requires models to combine $k$ random, disparate skills (like "metaphor" and "modus ponens") into a single response. Because the number of possible skill combinations explodes combinatorially, it is statistically impossible for the model to have seen the specific task during training.

Ultimately, the goal is to move from models that interpolate between seen examples to models that possess a "productive" understanding of the world—where the whole is truly understood as a systematic composition of its parts.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_systematic_generalization_to_novel_conce_ab688014,
  title = {Systematic Generalization to Novel Concept Combinations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/systematic-generalization-to-novel-concept-combinations/rdr_ab6880142e0244048c34bf4165ca6df9},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_systematic_generalization_to_novel_conce_ab688014_v1,
  title = {Systematic Generalization to Novel Concept Combinations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/systematic-generalization-to-novel-concept-combinations/rdr_ab6880142e0244048c34bf4165ca6df9/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}