Reproduce “Context Is a Pipeline, Not a Window: Information Attrition in LLM Agent Harnesses” 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
Yuvraj Sehgal, Mahsa Panahandeh, Naser Ezzati-Jivan
The paper measures how a harness delivers Linux-kernel-trace evidence to an agent diagnosing injected faults. Across 720 runs, the largest prompt used 5.75% of a 400,000-token window, yet the harness cut 3,827 of 11,189 tool results, making every cut result broken JSON. In controlled spawned-fault runs, cutting increased work but did not establish an accuracy effect.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.