mlsanityData to 7 Oct 2026

Transcribe, Translate, and Optimize: Joint Reward Learning for Speech Translation

arXiv 2609.26536 ↗First posted 22 Sep 2026cs.CL eess.AScited by 0

On the map: LLM reasoning → Reinforcement learning for reasoning

Abstract (arXiv)

In LLM-based speech translation, transcription-based chain-of-thought (CoT) suffers from a mismatch between reference transcripts used in supervised fine-tuning (SFT) and model-generated transcripts at inference. To address this, we propose joint recognition and translation fine-tuning via group relative policy optimization (GRPO). We score both transcripts and translations, with translation conditioned on model-generated transcripts, and compare three token advantage strategies. Using Qwen2.5-Omni-3B across four languages, we evaluate CoT against direct speech translation (Direct ST) under SFT and GRPO, training on CoVoST 2 and testing on CoVoST 2 and FLEURS. CoT GRPO outperforms Direct ST GRPO by 1.77 and 0.83 average BLEU points on CoVoST 2 and FLEURS. Compared to CoT SFT, GRPO boosts BLEU by 0.82 and 0.67 points and reduces word error rate (WER) by 8.8% and 7.2% relatively. These results highlight reinforcement fine-tuning as an effective method to mitigate the training-inference mismatch, jointly improving recognition and translation.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Reinforcement learning for reasoning, one of 17 directions in LLM reasoning. Builds on 22 of its 30 references that are in the map. First posted on arXiv 22 Sep 2026.

Builds on

22 of 30 references.

Closest ancestors · 2

Same direction, newest first

  1. 2025Advancing Speech Understanding in Speech-Aware Language Models with GRPO21 Sep
  2. 2025DAPO: An Open-Source LLM Reinforcement Learning System at Scale18 Mar

Other references · 20

  1. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 8,851
  2. 2022Robust Speech Recognition via Large-Scale Weak Supervisioncited by 3,972
  3. 2018A Call for Clarity in Reporting BLEU Scorescited by 1,790
  4. 2025Qwen2.5-Omni Technical Reportcited by 704
  5. 2022FLEURS: Few-shot Learning Evaluation of Universal Representations of Speechcited by 458
  6. 2024SWIFT:A Scalable lightWeight Infrastructure for Fine-Tuningcited by 345
  7. 2017Sequence-to-Sequence Models Can Directly Translate Foreign Speechcited by 121
  8. 2016Listen and Translate: A Proof of Concept for End-to-End Speech-to-Text Translationcited by 113
  9. 2017Minimum Word Error Rate Training for Attention-based Sequence-to-Sequence Modelscited by 68
  10. 2018Tied Multitask Learning for Neural Speech Translationcited by 54
  11. 2023Speech Translation with Large Language Models: An Industrial Practicecited by 28
  12. 2024Making LLMs Better Many-to-Many Speech-to-Text Translators with Curriculum Learningcited by 20
8 more
  1. 2024Chain-of-Thought Prompting for Speech Translationcited by 17
  2. 2022Improving Rare Word Recognition with LM-aware MWER Trainingcited by 12
  3. 2024LLaST: Improved End-to-end Speech Translation System Leveraged by Large Language Modelscited by 12
  4. 2025Group Relative Policy Optimization for Speech Recognitioncited by 8
  5. 2025Speech-to-Text Translation with Phoneme-Augmented CoT: Enhancing Cross-Lingual Transfer in Low-Resource Scenarioscited by 3
  6. 2025Revisiting Direct Speech-to-Text Translation with Speech LLMs: Better Scaling than CoT Prompting?cited by 2
  7. 2025End-to-end Automatic Speech Recognition and Speech Translation: Integration of Speech Foundational Models and LLMscited by 2
  8. 2025Listening or Reading? Evaluating Speech Awareness in Chain-of-Thought Speech-to-Text Translationcited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Reference-Free Reinforcement Learning Fine-Tuning for MT: A Seq2Seq Perspective
  2. 2025Multi-Reward GRPO for Stable and Prosodic Single-Codebook TTS LLMs at Scale
  3. 2025Improving Low-Resource Translation with Dictionary-Guided Fine-Tuning and RL: A Spanish-to-Wayuunaiki Study
  4. 2026Reference-Free Post-Training of Open Large Language Models for Multilingual Machine Translation
  5. 2026Translate-R1: Cost-Aware Translation Tool Use via Reinforcement Learning
  6. 2025Group Causal Policy Optimization for Post-Training Large Language Models
  7. 2025GVPO: Group Variance Policy Optimization for Large Language Model Post-Training
  8. 2025GTPO: Stabilizing Group Relative Policy Optimization via Gradient and Entropy Control