AutoIndexer preprint demonstrates a model that edits its earlier output

Researchers report that AutoIndexer can insert, replace and delete text it has already generated, addressing a limitation of models that produce text in a fixed sequence. Their preprint says the method demonstrates editing after generation without compromising performance on standard language-model benchmarks. AutoIndexer learns to open and close edits and move a cursor through existing text. Training examples teach it to undo deliberately introduced changes, helping it handle erroneous context and tasks such as code correction. The researchers say code is available; the findings concern the method's demonstrated capabilities.