Reproduce “Forgetting Without Restarting: Execution-State Unlearning for Stateful LLM Agents” 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
Chao Yao, Yangbo Wei, Zhen Huang, Junhong Qian, Chenle Chen, Shaoqiang Lu, Chen Wu, Lei He
The authors present provenance-guided selective replay to make an agent’s later behavior match a counterfactual run that never received a target, restoring a clean state and replaying the affected suffix. Under deterministic decoding, counterfactual observations, and restorable environments, B7 reaches zero any-leak and zero CAD while recomputing 133 versus 1,235 tokens; it uses attribution oracle and does not prove equivalence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.