Neural Foundations · 2015
Pointer Networks
Repurposed attention as a pointer that selects positions in the input as outputs, so a sequence-to-sequence model can emit outputs whose vocabulary is the input itself, solving convex hull, Delaunay triangulation and travelling-salesman instances of variable size.
Editorial record
Plain-language summary
Standard sequence-to-sequence models choose each output from a fixed vocabulary, which fails when the answer is a set of indices into the input. Pointer networks use the attention distribution directly as that choice: at each step the decoder points at one input element. This small change let neural networks tackle combinatorial geometry problems and later inspired copy mechanisms in summarisation and question answering, where the model copies spans from its input.
Source record
Provenance
- Record ID
- P-672
- Record created
- 2026-09-26
- Last reviewed
- 2026-09-26
- Record version
- 1
- https://arxiv.org/abs/1506.03134
- arXiv:1506.03134
Citation caveat: Citation metadata is approximate and marked unverified in the source dataset.