LatentSkill Converts Textual Agent Skills to LoRA Adapters, Cuts Prefill Tokens 64–72%

A hypernetwork converts reusable textual agent skills into plug-and-play LoRA adapters stored in weight space instead of context. On ALFWorld, success rate improves 13.4–21.4 points over in-context skill baselines while using 64.1% fewer prefill tokens; Search-QA exact match improves 3.0 points with 72.2% lower skill-token overhead.