协调推理与环境调度，Ling-3.x 智能体批量任务 makespan 降低 10.4%-65.0%
Reducing Batch Makespan for LLM Agent Workloads through Inference-Environment Coordination
做智能体评测、合成数据或 RL rollout 的团队值得看这篇。它把推理和环境调 preferring 调度放在一起优化，在 Ling-3.x 上实测 makespan 最多降 65%，分数不掉。如果你被批量跑任务太慢困扰，CAAS 那部分思路可以直接借鉴。
一篇 arXiv 论文提出针对 LLM 智能体工作负载的批量执行优化方法，目标是缩短 batch makespan，即完成整批任务所需时间。这类工作负载常见于基准评测、合成数据生成和 RL rollouts，任务在 LLM 推理与环境交互之间交替进行。
技术上包含几项协调优化：对混合架构 LLM，用基于工作负载的资源模型在 KV cache 与循环状态之间分配 GPU 显存，并给出参考样本并发度；通过轨迹感知路由与调度促进缓存复用并均衡 worker 负载；在环境容量受限时，将样本准入与环境执行解耦，引入 Chain-Advancing Action Scheduling（CAAS），优先推进长任务链以缩短批次尾部。评估用受控回放结合真实推理与录制的环境响应做消融实验。
在 Ling-3.0-Flash 与 Ling-3.1-Flash 上的实测覆盖 BFCL、SWE-bench Multilingual、WideSearch 和 MCP-Atlas 四个基准，makespan 降低 10.4% 到 65.0%，评测分数保持一致。这些方法已用于 Ling-3.x 系列的训练与评测流程，加快模型开发迭代。
Reducing Batch Makespan for LLM Agent Workloads through Inference-Environment Coordination
Efficient batch execution of agent tasks is critical to LLM training and development workflows, including benchmark evaluation, synthetic-data generation, and RL rollouts. Since these tasks alternate between LLM inference and environment interaction, we present coordinated optimizations to reduce batch makespan, the time required to complete all tasks in a batch. For hybrid LLMs, a workload-based resource model guides GPU-memory allocation between key-value (KV) cache and recurrent state and provides a reference sample concurrency to exploit parallelism while limiting recomputation. Trajectory-aware routing and scheduling promote cache reuse while balancing worker load. Under constrained environment capacity, we decouple sample admission from environment execution and introduce Chain-Advancing Action Scheduling (CAAS), which prioritizes long chains using observed progress to shorten the batch tail. Controlled replay combines actual inference with recorded environment responses and durations for ablations of individual optimizations. Live evaluations with Ling-3.0-Flash and Ling-3.1-Flash on the BFCL, SWE-bench Multilingual, WideSearch, and MCP-Atlas benchmarks demonstrate makespan reductions of 10.4% to 65.0%, while evaluation scores remain aligned. Furthermore, these methods have been successfully applied to training and evaluation workflows for the Ling-3.x model series, accelerating the model-development cycle.
This report is categorized under AITOP's "Papers" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Ling-3.1-Flash, SWE-bench, RL rollouts, 推理优化.
Source attribution: curated from arXiv cs.AI (@Wei Feng, Meiling Chen, Boxiao Chen, Lu Liu, Bin Zi, Libin Ye, Xingyu Lu, Lei Xu, Jun Zhou).
