Reproduce “Does an Agent's History Tell You When Compaction Will Hurt? A Modest, Bounded Effect on the TRACE Paired-Replay Corpus” 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

Egor Pakhomov, Erik Nijkamp

The authors test whether an agent’s pre-compaction history predicts harmful compactions using paired replays of 590 AppWorld boundaries. A two-feature logistic trigger avoided 20.6% of positive-burden boundaries while retaining 83.7% of opportunities, a 4.4-point advantage over random selection. A 17-feature model reached 0.657 AUROC, below a noisy same-boundary replicate; token savings were not measured.

Summary by alphaXiv.

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