Ling team cuts agent eval batch times with trajectory scheduling
AI-generated summary The Ling team released a scheduling method that orders agent evaluation and reinforcement learning rollouts by expected length to free capacity without additional hardware spend. On BFCL, SWE-bench Multilingual, WideSearch, and MCP-Atlas, batch completion times dropped by 10.4 percent to 65.0 percent while evaluation scores remained unchanged for the Ling-3.x models.