---
title: "RISED: Rubrics for Agentic Multi-Environment Selection and Self-Distillation"
description: "Training a single LLM agent jointly across diverse interactive environments is challenged by scalar rewards that provide limited information about cross-environment relationships and no within-group reward contrast when rewards are identical. RISED addresses this by using rubrics—textual descriptions of rollout behaviors—to guide both online data selection and policy supervision. An LLM judge tags each rollout with a predefined rubric vocabulary shared across environments, and these profiles select data aligned with the batch’s behavioral composition while limiting overlap with already-selected data. Positive rubrics supply token-level supervision via an on-policy self-distillation teacher, while negative rubrics steer subsequent rollouts away from recurring failure modes. Across model backbones, RISED achieves the highest mean pass rate across environments and ranks first or second in every individual environment."
---

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

[Apple](https://yomu.fyi/company/apple) · Oct 6, 2026

**Type:** Problem & solution

## Summary

Training a single LLM agent jointly across diverse interactive environments is challenged by scalar rewards that provide limited information about cross-environment relationships and no within-group reward contrast when rewards are identical. RISED addresses this by using rubrics—textual descriptions of rollout behaviors—to guide both online data selection and policy supervision. An LLM judge tags each rollout with a predefined rubric vocabulary shared across environments, and these profiles select data aligned with the batch’s behavioral composition while limiting overlap with already-selected data. Positive rubrics supply token-level supervision via an on-policy self-distillation teacher, while negative rubrics steer subsequent rollouts away from recurring failure modes. Across model backbones, RISED achieves the highest mean pass rate across environments and ranks first or second in every individual environment.

## Context

RISED is a method for training LLM agents across diverse interactive environments. It uses rubrics—textual descriptions of rollout behaviors—to guide data selection and policy supervision, addressing limitations of scalar rewards.

## Approach / What changed

RISED employs an LLM judge to tag rollouts with predefined rubric vocabulary. Rubric profiles guide data selection to align with batch behavioral composition and limit overlap. Positive rubrics provide token-level supervision via on-policy self-distillation teacher; negative rubrics steer generation away from failures.

## Takeaways

- Scalar rewards in multi-environment RL provide limited information about cross-environment relationships and no within-group reward contrast when rewards are identical, motivating richer textual feedback such as rubrics.
- RISED uses an LLM judge to tag each rollout with a predefined rubric vocabulary shared across environments, and these profiles guide data selection that aligns with the batch’s behavioral composition while limiting overlap with already-selected data.
- Across model backbones, RISED achieves the highest mean pass rate across environments and ranks first or second in every individual environment, demonstrating the effectiveness of rubric-based guidance.

## FAQ

### Why are scalar rewards insufficient for multi-environment RL?

Scalar rewards provide limited information about cross-environment relationships and no within-group reward contrast when rewards are identical, making it hard to guide learning across diverse environments.

### How does RISED use rubrics for data selection?

An LLM judge tags each rollout with predefined rubric vocabulary. These profiles select data that aligns with the batch's behavioral composition while limiting overlap with already-selected data.

### What supervision do positive and negative rubrics provide?

Positive rubrics supply token-level supervision via an on-policy self-distillation teacher, while negative rubrics steer subsequent rollout generation away from recurring failure modes.

**Tags:** [AI Agents](https://yomu.fyi/topic/ai-agents), [LLMs](https://yomu.fyi/topic/llm), [Reinforcement Learning](https://yomu.fyi/topic/reinforcement-learning)

- Source: [Apple](https://machinelearning.apple.com/research/rised-multi-environment-selection)
- Source URL: https://machinelearning.apple.com/research/rised-multi-environment-selection
- Ingested by Yomu: 2026-10-07T03:00:07.639Z

[Read original post](https://machinelearning.apple.com/research/rised-multi-environment-selection)