Apple AI Updates: May 5, 2026

1. Apple Research: PORTool Tackles Step-Level Credit Assignment for Tool-Using Agents

Apple. Apple ML Research published PORTool, a policy-optimization algorithm aimed at one of the harder problems in tool-using LLMs — assigning credit to individual tool calls when only the final outcome is rewarded. PORTool reinforces tool-use competence from outcome-level supervision but distributes reward at the step level, which the authors argue produces more sample-efficient training for multi-tool reasoning than naïve outcome-only RL. Worth a look if you’re training agents that chain search, code execution, and retrieval. Source