Improving Transformer Analogical Reasoning via Copy Task Training
August 21, 2026
Training transformers with Meta-Learning for Compositionality (MLC) and explicit copy tasks improves analogical reasoning performance. A 3-layer encoder-decoder model using this method achieves parity with frontier models on letter-string analogy tasks.
HOW THIS AFFECTS YOU
●
researcherYou can use copy tasks as an intermediate training step to boost compositional and analogical generalization in smaller architectures.