Reproduce “Revisable by Design: A Theory of Streaming LLM Agent Execution” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhiyuan Zhai, Ming Li, Xin Wang

Researchers from Fudan University and Guangming Lab developed a "stream paradigm" for large language model (LLM) agents, allowing concurrent execution and user revisions, along with a formal reversibility taxonomy and an adaptive algorithm. The Revision Absorber algorithm achieved comparable task quality to full restarts but reduced wasted computational steps by 14.6 times in handling mid-execution user revisions.

Key insights

  • An agent's ability to absorb revisions is fundamentally limited by the reversibility of actions already executed, termed the Irreversibility Cost Principle.
  • Rolling back to the step immediately preceding the earliest conflicting Compensable or Irreversible action minimizes adaptation cost (Structural Optimality of Earliest-Conflict Rollback).
  • Agent state can be decomposed into an epistemic state (always rollbackable) and a world state (rollbackability depends on action type), informing efficient revision handling.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.