The paper presents SoL-Pi, a system that automates the discovery of token-efficient mechanisms for . Using a approach, it scales auto-research loops across hundreds of environments to find harness improvements that transfer to unseen tasks.
Four mechanisms were discovered: Action Fusion, Online Context Compact, ObservationPack, and Evidence-Preserving Reducer. On the 51-task EdgeBench benchmark, SoL-Pi achieves performance comparable to the Pi harness while reducing recorded token traffic by 44.7-49.0% and API cost by about one third.
This is interesting because it shows that harness-level optimization can significantly improve token efficiency without modifying the underlying model, and that automated search can discover reusable improvements that generalize beyond the development setting.