# Multi-Reward GRPO for Stable and Prosodic Single-Codebook TTS LLMs at Scale

arXiv: https://arxiv.org/abs/2511.21270 · HTML: https://mlsanity.com/paper/2511.21270/

## Key facts

- Cited by 3 papers in the map (data through 14 Oct 2026).
- Placed in Reinforcement learning for reasoning, one of 17 directions in LLM reasoning.
- Builds on 12 of its 20 references that are in the map.
- First posted on arXiv 26 Nov 2025.

## Placed in

- LLM reasoning / [Reinforcement learning for reasoning](https://mlsanity.com/lr/reinforcement-learning-for-reasoning/)

## Builds on (most cited first, 12 of 12)

- [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://mlsanity.com/paper/2402.03300/) (2024)
- [DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning](https://mlsanity.com/paper/2501.12948/) (2025)
- [Robust Speech Recognition via Large-Scale Weak Supervision](https://mlsanity.com/paper/2212.04356/) (2022)
- [CosyVoice: A Scalable Multilingual Zero-shot Text-to-speech Synthesizer based on Supervised Semantic Tokens](https://mlsanity.com/paper/2407.05407/) (2024)
- [Seed-TTS: A Family of High-Quality Versatile Speech Generation Models](https://mlsanity.com/paper/2406.02430/) (2024)
- [CosyVoice 2: Scalable Streaming Speech Synthesis with Large Language Models](https://mlsanity.com/paper/2412.10117/) (2024)
- [CosyVoice 3: Towards In-the-wild Speech Generation via Scaling-up and Post-training](https://mlsanity.com/paper/2505.17589/) (2025)
- [Spark-TTS: An Efficient LLM-Based Text-to-Speech Model with Single-Stream Decoupled Speech Tokens](https://mlsanity.com/paper/2503.01710/) (2025)
- [Llasa: Scaling Train-Time and Inference-Time Compute for Llama-based Speech Synthesis](https://mlsanity.com/paper/2502.04128/) (2025)
- [FireRedTTS: A Foundation Text-To-Speech Framework for Industry-Level Generative Speech Applications](https://mlsanity.com/paper/2409.03283/) (2024)
- [Enhancing Zero-shot Text-to-Speech Synthesis with Human Feedback](https://mlsanity.com/paper/2406.00654/) (2024)
- [Robust Zero-Shot Text-to-Speech Synthesis with Reverse Inference Optimization](https://mlsanity.com/paper/2407.02243/) (2024)

## Built on by (most cited first, 3 of 3)

- [Multi-Task GRPO: Reliable LLM Reasoning Across Tasks](https://mlsanity.com/paper/2602.05547/) (2026)
- [GLASS: GRPO-Trained LoRA for Acoustic Style Steering in Zero-Shot Text-to-Speech](https://mlsanity.com/paper/2606.05889/) (2026)
- [When Does Predictor-Based RL Align with Human Perception? A Study of Subjective Rewards in Codec-Based Speech Language Models](https://mlsanity.com/paper/2608.31035/) (2026)

## Abstract (arXiv)

Recent advances in Large Language Models (LLMs) have transformed text-to-speech (TTS) synthesis, inspiring autoregressive frameworks that represent speech as sequences of discrete codec tokens. Among them, single-codebook TTS LLMs have emerged as compact and streamable architectures that jointly model semantic and acoustic integration. However, despite their efficiency, these models often exhibit unstable prosody, speaker drift, and degraded naturalness. To address these issues, we propose a multi-reward Group Relative Policy Optimization (GRPO) framework that directly optimizes the token generation policy of single-codebook TTS LLMs. Beyond standard intelligibility and speaker similarity objectives, our design integrates three rule-based rewards: a length penalty for duration consistency, an entropy regularization reward for decoding stability, and an LLM-annotated prosody alignment reward that explicitly supervises rhythm. In this prosody reward, an external reasoning LLM predicts multiple plausible pause structures via in-context learning, providing a human-preference-aligned supervisory signal for GRPO training. To assess universality, we further attach a flow-matching (FM) decoder on top of the GRPO-optimized AR backbone and observe consistent additional gains, indicating that our reinforcement optimization enhances the intrinsic AR policy. We further conduct a scalability analysis across data sizes and model scales, revealing that the proposed method consistently enhances prosodic stability, speaker similarity, and overall speech naturalness in single-codebook TTS LLMs.