Reproduce “Where Facts Go Missing: A Layerwise Taxonomy and Per-Layer Attribution of Information Omission in Air-Gapped LLMAgent Pipelines” 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
Santhiya Rajan, Samuel Mugel, Roman Orus
The authors trace facts through a nine-layer agent pipeline to identify where omissions occur. They use canary strings at checkpoints and fault injections across 75,476 trials. The allocation-weighted omission rate is 0.574, with L0–L3 accounting for 73.4% of waterfall loss. In gold-path tests, 32k versus 2k context has 7.43 times higher omission odds, an association limited to tested conditions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.