SCAN

The SCAN dataset is a benchmark that contains compositional generalization tasks used to evaluate the ability of models, particularly Transformers, to generalize to new combinations of learned concepts.

Papers using SCAN (4)

ALTA: Compiler-Based Analysis of Transformers2024 · 1 citesCompositional Generalization for Natural Language Interfaces to Web APIs2021 · 2 citesCompositional Generalization and Decomposition in Neural Program Synthesis2022 · 2 citesCompositional Program Generation for Few-Shot Systematic Generalization2023 · 2 cites