# Reducing the Output-Mode Gap in Speech Language Models via Joint-Output On-Policy Distillation

arXiv: https://arxiv.org/abs/2609.15313 · HTML: https://mlsanity.com/paper/2609.15313/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Speech and audio distillation, one of 12 directions in Knowledge distillation.
- Builds on 25 of its 26 references that are in the map.
- First posted on arXiv 14 Sep 2026.

## Placed in

- Knowledge distillation / [Speech and audio distillation](https://mlsanity.com/distill/speech-and-audio-distillation/)

## Builds on (most cited first, 25 of 25)

- [Training Verifiers to Solve Math Word Problems](https://mlsanity.com/paper/2110.14168/) (2021)
- [Tulu 3: Pushing Frontiers in Open Language Model Post-Training](https://mlsanity.com/paper/2411.15124/) (2024)
- [On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes](https://mlsanity.com/paper/2306.13649/) (2023)
- [Qwen-Audio: Advancing Universal Audio Understanding via Unified Large-Scale Audio-Language Models](https://mlsanity.com/paper/2311.07919/) (2023)
- [Moshi: a speech-text foundation model for real-time dialogue](https://mlsanity.com/paper/2410.00037/) (2024)
- [SpeechGPT: Empowering Large Language Models with Intrinsic Cross-Modal Conversational Abilities](https://mlsanity.com/paper/2305.11000/) (2023)
- [SALMONN: Towards Generic Hearing Abilities for Large Language Models](https://mlsanity.com/paper/2310.13289/) (2023)
- [Qwen3-Omni Technical Report](https://mlsanity.com/paper/2509.17765/) (2025)
- [GLM-4-Voice: Towards Intelligent and Human-Like End-to-End Spoken Chatbot](https://mlsanity.com/paper/2412.02612/) (2024)
- [VoiceBench: Benchmarking LLM-Based Voice Assistants](https://mlsanity.com/paper/2410.17196/) (2024)
- [Spirit LM: Interleaved Spoken and Written Language Model](https://mlsanity.com/paper/2402.05755/) (2024)
- [Step-Audio 2 Technical Report](https://mlsanity.com/paper/2507.16632/) (2025)
- [NaturalReasoning: Reasoning in the Wild with 2.8M Challenging Questions](https://mlsanity.com/paper/2502.13124/) (2025)
- [Baichuan-Audio: A Unified Framework for End-to-End Speech Interaction](https://mlsanity.com/paper/2502.17239/) (2025)
- [MiMo-Audio: Audio Language Models are Few-Shot Learners](https://mlsanity.com/paper/2512.23808/) (2025)
- [LongCat-Next: Lexicalizing Modalities as Discrete Tokens](https://mlsanity.com/paper/2603.27538/) (2026)
- [Closing the Gap Between Text and Speech Understanding in LLMs](https://mlsanity.com/paper/2510.13632/) (2025)
- [X-OPD: Cross-Modal On-Policy Distillation for Capability Alignment in Speech LLMs](https://mlsanity.com/paper/2603.24596/) (2026)
- [Understanding the Modality Gap: An Empirical Study on the Speech-Text Alignment Mechanism of Large Speech Language Models](https://mlsanity.com/paper/2510.12116/) (2025)
- [Towards Spoken Mathematical Reasoning: Benchmarking Speech-based Models over Multi-faceted Math Problems](https://mlsanity.com/paper/2505.15000/) (2025)
- [CORD: Bridging the Audio-Text Reasoning Gap via Weighted On-policy Cross-modal Distillation](https://mlsanity.com/paper/2601.16547/) (2026)
- [Closing the Modality Reasoning Gap for Speech Large Language Models](https://mlsanity.com/paper/2601.05543/) (2026)
- [From Text to Talk: Audio-Language Model Needs Non-Autoregressive Joint Training](https://mlsanity.com/paper/2509.20072/) (2025)
- [X3-OPD: Distilling Reasoning into Large Audio-Language Models via On-Policy Alignment](https://mlsanity.com/paper/2607.21550/) (2026)
- [Minimizing Modality Gap from the Input Side: Your Speech LLM Can Be a Prosody-Aware Text LLM](https://mlsanity.com/paper/2605.05927/) (2026)

## Abstract (arXiv)

Autoregressive generation of interleaved text and acoustic tokens is a common approach to spoken-response generation in speech large language models. Although this design enables streaming generation with explicit textual guidance, generated acoustic tokens become part of the context for subsequent text predictions. Given identical speech inputs, we observe markedly lower answer accuracy for the internal text generated in speech-to-text-and-speech (S2TS) mode than for speech-to-text (S2T) responses. We term this discrepancy the output-mode gap (OMG). To reduce OMG, we propose Joint-Output On-Policy Distillation (JO-OPD), which distills the model's stronger S2T policy into joint generation using student-generated S2TS trajectories. At each text position, the S2T teacher provides soft targets from a text-only projection of the student's preceding outputs, while the student predicts from the corresponding full interleaved history. A preservation objective further regularizes native non-text predictions. Experiments on Step-Audio-2-mini and Baichuan-Audio-Instruct reveal OMG across two interleaved generation architectures. On Step-Audio-2-mini, JO-OPD reduces OMG from 42.87 to 16.26 percentage points on Spoken-MQA and from 29.72 to 13.04 points on speech-rendered GSM8K, with little change in S2T accuracy and substantially larger reductions than matched SFT baselines. ASR-based evaluation further shows a 7.49-point improvement in spoken-answer accuracy on Spoken-MQA.