Reproduce “TAO: Tool-Aware Joint Scheduling for LLM-Based Agentic Serving” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shuhuai Li, Shengdong Ye, Dongdong Ge, Yinyu Ye, Jianghao Lin
TAO schedules agentic inference across GPU nodes by valuing cache reuse, predicted tool-return readiness, and token footprint, then jointly selecting sessions and placements under capacity. On DeepSeek-V4 Flash with SWE-Lite at 64 workers, it reduced mean task latency by 26.8%, lowered TTFT by 36.3%, and raised decode throughput by 29.5% against metric-specific best baselines; gains were not universal across settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.