Reproduce “Managing Procedural Memory in LLM Agents: Control, Adaptation, and Evaluation” 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

Julia Belikova, Rauf Parchiev, Evgeny Egorov, Grigorii Davydenko, Gleb Gusev, Andrey Savchenko, Maksim Makarenko

The authors introduce AFTER, a benchmark and workflow for evolving reusable procedural skills from execution traces. The harness diagnoses failures, revises skills, and tests them on familiar and held-out tasks, roles, and models. Skills evolved from pooled multi-model traces reached 73.1% reported test accuracy, 13.7 points higher than the best single-model source. AFTER covers technology-sector procedural tool-use tasks.

Key insights

  • LLM-generated skills frequently outperformed handcrafted baselines, and weaker models consistently benefited more from procedural skills, showing gains up to +14.2 points for Gemma 4.
  • Skills evolved from diverse multi-model execution traces significantly outperformed single-model sources, achieving 73.1% test accuracy compared to 36.0–59.4% for single-model sources, highlighting the importance of varied experience.
  • Procedural skills evolved within one professional role tended to over-specialize, demonstrating performance degradation (losses of -4.8 to -7.5 points) when applied to agents in different roles.

Summary by alphaXiv.

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