Reproduce “AfterDeleteBench: Auditing Post-Deletion Information Persistence in 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
Ehsan Saleh, Roozbeh Nahavandi, Sama Salarian, Raef Bassily, Andrew Perrault
AfterDeleteBench tests whether an agent can still reveal information after its source record is deleted. It pairs agents that saw a synthetic target with controls that never saw it. In the hierarchical-memory setting, Qwen2.5-7B-Instruct had an exact-recovery gap of 0.680 on IID profiles, while controls recovered zero of 200 targets. Flat-memory and exposure-group-deletion tests showed zero observed recovery.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.