ASTOR: Utility-Guided Multi-Task RL for Code LLMs

A new framework called ASTOR (multi-tASk code reinforcement learning via uTility-driven coORdination) addresses limitations in multi-task reinforcement learning for code large language models. Proposed in arXiv:2605.06111, ASTOR introduces task utility—a signal capturing each task's learning potential and cross-task synergy—to guide training. It comprises two modules: a Hierarchical Utility-Routed Data Scheduling module that allocates training budget and prioritizes informative prompts, and an Adaptive Utility-Calibrated Policy Optimization module. This approach aims to overcome the inefficiency of deploying separate task-specific specialists and the limitations of existing multi-task RL methods that treat all coding tasks uniformly with fixed data curricula.