Megadose The AI curve, compressed.

Reducing Batch Makespan for LLM Agent Workloads through Inference-Environment Coordination

· HF Daily Papers ·
Coordinated scheduling cut agent-batch completion time by 10.4%–65.0% in live tests, with evaluation scores remaining aligned.

The researchers tested Ling-3.0-Flash and Ling-3.1-Flash across BFCL, SWE-bench Multilingual, WideSearch, and MCP-Atlas. Their approach allocates GPU memory between KV cache and recurrent state, then routes and schedules tasks to reuse caches while balancing worker load. When environment capacity is limited, Chain-Advancing Action Scheduling prioritizes long task chains using observed progress to shorten the wait for the batch’s final tasks. The authors report applying these methods to Ling-3.x training and evaluation workflows.

HF Daily Papers' note

score 5

Categories: Research