Stream of Revision: LLM Code Generation with On-the-Fly Editing
Researchers propose Stream of Revision, a new framework for LLM-based code generation that allows models to backtrack and edit their own output within a single forward pass, mimicking the iterative nature of human programming. Unlike traditional monotonic token generation, this method uses special action tokens to enable self-correction without external tools or post-hoc agents, reducing latency while leveraging the model's intrinsic reasoning. The approach marks a paradigm shift from linear to dynamic code generation.
Key facts
- Stream of Revision is a framework for LLM code generation.
- It enables backtracking and editing within a single forward pass.
- Uses specific action tokens for self-correction.
- Contrasts with strictly monotonic token generation.
- Aims to reduce latency compared to post-hoc agents.
- Leverages the model's intrinsic semantic reasoning.
- Proposed in arXiv paper 2602.01187.
- Paradigm shift from monotonic stream to dynamic trajectory.
Entities
Institutions
- arXiv