DIFFA-2: A Practical Diffusion Large Language Model for General Audio Understanding

Autoregressive (AR) large audio language models (LALMs) such as Qwen-2.5-Omni have achieved strong performance on audio understanding and interaction, but scaling them remains costly in data and computation, and strictly sequential decoding limits inference efficiency. Diffusion large language models (dLLMs) have recently been shown to make effective use of limited training data, and prior work on DIFFA indicates that replacing an AR backbone with a diffusion counterpart can substantially improve audio understanding under matched settings, albeit at a proof-of-concept scale without largescale instruction tuning, preference alignment, or practical decoding schemes. We introduce DIFFA-2, a practical diffusion-based LALM for general audio understanding. DIFFA-2 upgrades the speech encoder, employs dual semantic and acoustic adapters, and is trained with a four-stage curriculum that combines semantic and acoustic alignment, large-scale supervised fine-tuning, and variance-reduced preference optimization, using only fully open-source corpora. Experiments on MMSU, MMAU, and MMAR show that DIFFA-2 consistently improves over DIFFA and is competitive to strong AR LALMs under practical training budgets, supporting diffusion-based modeling is a viable backbone for large-scale audio understanding. Introduction Diffusion large language models (dLLMs) (Nie et al., 2025; Zhu et al., 2025) have recently emerged as a promising alternative to conventional autoregressive (AR) decoders. Instead of generating tokens strictly left-to-right, dLLMs perform iterative denoising over partially masked sequences, enabling any-order token modeling and naturally supporting parallel token updates during decoding. Recent study (Ni et al., 2025) in the text domain further shows that dLLMs can act as strong data learners: when the amount of unique training data is constrained, they continue to improve and can even surpass AR models by leveraging super-dense compute and implicit Monte Carlo-style data augmentation. These properties are particularly appealing for audio understanding, where high-quality audio-text supervision across speech, sound, and music is far more expensive to collect than text-only data, and where the latency of strictly sequential AR decoding becomes a bottleneck for long-form and interactive applications. Despite these advantages, state-of-the-art large audio language models (LALMs) are still predominantly AR-based: systems such as Qwen-3-Omni (Xu et al., 2025b), Qwen-2.5-Omni (Xu et al., 2025a), and Kimi-Audio (Ding et al., 2025) couple powerful speech encoders with AR LLMs and achieve strong results on a wide range of audio understanding and dialogue benchmarks. This raises a natural question: can dLLMs be turned into competitive and practical audio backbones that match these AR LALMs under realistic data and latency budgets? Initial evidence is encouraging. A recent work on diffusion-based large audio-language model (DIFFA) (Zhou et al., 2025) compares an 8B AR backbone with its diffusion counterpart under matched data, adapter design, and training recipe, and finds substantial gains on audio understanding benchmarks such as MMAU (Sakshi et al., 2025) and MMSU (Wang et al., 2025b) after simply replacing the AR backbone. This suggests that the generative paradigm itself can strongly influence audio performance and that dLLMs have significant potential as audio backbones. However, that study remains largely a proof of concept: the model is trained mainly on speech-centric supervision with a relatively small Whisper encoder (Radford et al., 2023) , keeps the diffusion backbone frozen, and does not exploit large-scale instruction data, preference-based objectives, or practical inference acceleration. It therefore does not answer whether dLLMs can be scaled into strong and practical LALMs that reliably compete with state-of-the-art AR models.