AutoIndexer is a generative model that introduces a novel training and decoding scheme simulating a Chain of Edits, allowing the model to flexibly modify its previously generated output via token insertions, substitutions, and deletions. During decoding the model emits special marker tokens that open and close an edit region and moves a cursor to the target position, much like a text editor.
For training, ground‑truth sequences are perturbed with inverse edit operations; the model learns to undo these inverses, thereby becoming robust to erroneous context and capable of restoring correct content. This approach yields strong performance on editing tasks such as code correction while preserving benchmark results on standard LLM evaluations. Unlike conventional autoregressive models that generate once and diffusion models that can only insert a limited number of tokens into a fixed context, AutoIndexer provides genuine post‑generation context editing. The code is released at https://github.com/AutodeskAILab/autoindexer.
Review