Small Models Beat the Monolith

🔥 What happened Shanka Mondal and Jiayun Wang dropped JHiMA-WebRL: a framework that jointly trains a 7B task decomposer and a 7B actor via GRPO. On WebArena-Lite-v2, the 14B pair hits a 37.2% success rate. 💡 Why it matters A monolithic ScaleCUA-32B agent, RL-tuned on the same tasks under the same protocol, manages only 24%. The ablation is the kicker: freezing the planner during RL tanks performance — the role split itself is the lever, not raw parameter count. ⚡ Our take Scaling was yesterday's answer. Anyone still inflating monoliths in 2027 instead of factoring roles cleanly is burning compute for a third less performance.