Workshop: 21st Workshop on Workflows in Support of Large-Scale Science (WORKS26)
Authors: Hongwei Jin and Krishnan Raghavan (Argonne National Laboratory), Anirban Mandal (Renaissance Computing Institute), and Ewa Deelman (University of Southern California)
Abstract: In decentralized scheduling of scientific workflows, multiple LLM-based agents reason over heterogeneous resources to bid for jobs. However, LLM reasoning incurs substantial latency and compute cost, limiting scalability in high-throughput environments. We therefore ask whether an LLM is necessary for every decision, and study adaptive LLM gating, where an RL-trained policy learns to choose between an LLM and a fast surrogate from downstream scheduling reward. We propose RLSA (Reinforcement Learning-enhanced Scheduling Agents), trained with proximal policy optimization (PPO) under a reward balancing makespan, resource balance, and inference cost. In a controlled auction, RLSA halves LLM calls per job (4.00 to 2.06) while preserving makespan relative to always-LLM bidding (7686 vs. 7285 seconds). We further validate it with a real LLM bidder and a 17-VM FABRIC deployment where LLM calls grow sub-linearly with the number of bidders, positioning RL-based adaptive gating as a cost-control layer for LLM-based scheduling agents.
Back to 21st Workshop on Workflows in Support of Large-Scale Science (WORKS26) Archive Listing
Back to Full Workshop Archive Listing