A2R-30B-A3B

🎉 EMNLP 2026 main conference 🎉

📄 arXiv · 🌐 Project page · 💻 Code · 🤗 Dataset

A2R answers questions about who is speaking, not only what is said. It is Qwen/Qwen3-Omni-30B-A3B-Instruct trained with GRPO on CASH-60K.

This is the merged model. Load it directly, no adapter step.

Usage

from transformers import Qwen3OmniMoeForConditionalGeneration, AutoProcessor

model = Qwen3OmniMoeForConditionalGeneration.from_pretrained(
    "PleasedPenguin/A2R-30B-A3B", dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained("PleasedPenguin/A2R-30B-A3B")

vllm serve PleasedPenguin/A2R-30B-A3B --max-model-len 32768

Output format

A2R reasons before it answers, and closes with a structured block:

<reasoning> … </reasoning><answer>{"Answer": "B"}</answer>

Parse the last <answer> block. Scanning the whole string will pick up option letters quoted inside the reasoning trace.

Multiple audio inputs are presented as a single waveform, with any reference voices appended after the main clip. The evaluation harness assembles this for you.

Citation

@misc{lee2026hearsaidwhatunlocking,
  title         = {HEAR Who Said What: Unlocking Speaker-Attributed Reasoning
                   via Counterfactual Voice Grounding},
  author        = {Dongwook Lee and Sangkwon Park and Eunwoo Song and Che Hyun Lee
                   and Youngho Cho and Junho Kim and June Young Yi and Heeseung Kim
                   and Sungroh Yoon},
  year          = {2026},
  eprint        = {2608.29120},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2608.29120}
}