---
title: "Discussion on \"Multi-Reward RL, Part 3: GDPO + CISPO + REPO-R at 27B, and the Advantage Floor That Stopped Learning\" | Hashnode"
description: "Discussion on \"Multi-Reward RL, Part 3: GDPO + CISPO + REPO-R at 27B, and the Advantage Floor That Stopped Learning\". Qwen3.8-27B, 600 steps, 12 reward channels: CISPO, REPO-R and prompt filtering lifted holdout from −0.47 to +1.80, and an advantage floor silently zeroed 96% of the signal."
url: "https://hashnode.com/posts/multi-reward-rl-part-3-gdpo-cispo-repo-r-at-27b-and-the-advantage-floor-that-stopped-learning/6ac929e55952e02b451b58c9"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Multi-Reward RL, Part 3: GDPO + CISPO + REPO-R at 27B, and the Advantage Floor That Stopped Learning](https://gfactor.hashnode.dev/multi-reward-rl-part-3-gdpo-cispo-repo-r-at-27b-and-the-advantage-floor-that-stopped-learning?utm_source=hashnode&utm_medium=feed)

Qwen3.8-27B, 600 steps, 12 reward channels: CISPO, REPO-R and prompt filtering lifted holdout from −0.47 to +1.80, and an advantage floor silently zeroed 96% of the signal.

Oct 9 · 15 min read · gfactor.hashnode.dev  
Updated Oct 11  
Read the article

-   [#reinforcement-learning](https://hashnode.com/tag/reinforcement-learning)
-   [#llm](https://hashnode.com/tag/llm)
-   [#machine-learning](https://hashnode.com/tag/machine-learning)
-   [#artificial-intelligence](https://hashnode.com/tag/artificial-intelligence)
-   [#deep-learning](https://hashnode.com/tag/deep-learning)

## Written by

[Aleksei Romanov](https://hashnode.com/@gfactor)

@gfactor · United States

  
[View profile](https://hashnode.com/@gfactor)

Ex-Apple GenAI engineer, building an LLM post-training platform at g factor technologies

Ex-Apple GenAI engineer, building an LLM post-training platform at g factor technologies

## Responses

Markdown supported · ⌘↵ to post

No responses yet.
