Skill-naming tokens get almost no RL credit, and splitting the loss into two channels lifts success 40.8% to 53.2%

arXiv 2608.18852·high signalXBlueskyLinkedInCopy link

SkillGate names the failure mode selector credit starvation. In long-horizon agents the tokens that name which skill to read receive a vanishing share of the loss signal, and the credit gets more wrong as trajectories lengthen. The fix partitions token support into two disjoint channels, outcome credit for execution tokens and an action-local advantage that only rewards skill-naming tokens when the single skill read was correct, taking a 9B policy from 40.8% to 53.2% across five agentic benchmarks with 16-candidate skill slates while cutting candidate exposure by two thirds.