kapynResearch

RISED: Rubrics for Agentic Multi-Environment Selection and Self-Distillation

RISED is a new training framework that jointly trains a single LLM agent across multiple interactive environments. It introduces prompt‑group selection that considers relationships between rollouts, and handles mixed success/failure groups within batches, improving data efficiency and generalist performance. The method offers a more nuanced curriculum for multi‑environment agent learning.

Apple ML Research·Oct 6, 2026

Opening Kapyn…