Reviser language model architecture edits mutable canvas

Reviser shifts the LLM paradigm from linear generation to dynamic editing, addressing a major bottleneck in model reasoning and error correction.

• –Autoregressive over edit actions allows for "mental" backspacing and restructuring while keeping decoding efficiency close to standard transformers.
• –The cursor-relative mechanism provides a more natural and compact interface for models to interact with existing context.
• –Native revision capabilities could significantly improve performance on long-form generation and complex problem-solving tasks.
• –Scaling this architecture to foundation-model sizes remains a key hurdle for demonstrating broad utility.