Learning to Extrapolate to New Tasks: A Relational Approach to Task Extrapolation

2026-05-28

AI Summary

Current learning systems exhibit limited generalization when confronted with out-of-distribution tasks, such as those involving out-of-range parameters, increased compositional depth, or novel recombinations of primitives. This work proposes the Relational Task Extrapolator (RTE), which formulates task extrapolation for the first time as transformational relationships between tasks. Specifically, RTE decomposes a target task into known anchor tasks and the transformations mapping them to the target, then learns a relational operator that maps anchor–transformation pairs to predictions for the target task. This approach transcends the limitations of conventional interpolation-based generalization, enabling robust extrapolation across diverse dimensions—including parameter values, sequence lengths, and compositional structures—and can be seamlessly integrated into standard model fine-tuning pipelines. Experiments demonstrate that RTE significantly outperforms existing methods across multiple task extrapolation settings, substantially enhancing systematic generalization to unseen tasks.

Abstract

Modern learning systems excel at interpolation but struggle to generalize to unseen tasks outside the training distribution's support. This failure occurs even in simple settings, such as handling task parameters beyond the training range, and persists despite advances in foundation models. To this end, we develop the Relational Task Extrapolator (RTE), an algorithm designed to enable systematic extrapolation to novel tasks. The key observation is that extrapolation is inherently relational: extrapolating to unseen tasks requires learning how tasks transform into one another. If a model learns the transformation between tasks A and B during training, it can apply that same transformation to relate known tasks to unseen ones at test time. RTE operationalizes this idea by decomposing each target task into a known anchor task and a transformation linking the anchor and target. It then learns a relational operator, mapping an anchor-transformation pair to predictions for the target task. We instantiate RTE across multiple task extrapolation regimes in function prediction, e.g. where target tasks use out-of-range parameters (parameter extrapolation), have greater compositional depth (length extrapolation), and/or recombine function primitives in unseen ways (compositional extrapolation). We further extend RTE to sequence prediction, integrating it into fine-tuning algorithms for foundation models. Across empirical studies, we find that RTE substantially outperforms existing approaches on extrapolation to novel, unseen tasks.

Problem

• task extrapolation
• out-of-distribution generalization
• relational learning
• foundation models
• systematic generalization

Innovation

• task extrapolation
• relational learning
• foundation models
• compositional generalization
• out-of-distribution generalization