The authors fixed the open source agent scaffold OpenCode and implemented five compaction strategies inside it: default self-summarization, a "lossless" version that writes the pre-compaction context to disk, removal of reasoning blocks, and summaries produced by a smaller model.
Context compaction spares benchmark scores but silently drops task constraints
In long-horizon agentic tasks, five compaction strategies leave task performance largely intact across frontier models while frequently erasing anti-cheating rules and user instructions.
Research Lab
Jeanne Salle · Hardik Bhatnagar · Matthias Bethge · Sahar Abdelnabi · Maksym Andriushchenko
ELLIS Institute Tübingen · Max Planck Institute for Intelligent Systems · Tübingen AI Center · University of Tübingen
Research Digest··3 min read
The authors tested five context compaction strategies inside a fixed agent scaffold across three long horizon benchmarks with frontier models.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.