Lab-MSP/comparative-reasoning-dpo

Comparative Reasoning — DPO (Qwen2.5-Omni-3B LoRA)

The DPO model from Comparative Reasoning: Making an Audio Language Model Better at Comparing Emotions (Interspeech 2026). Given two utterances, it answers which one has higher arousal, valence, or dominance.

This repository contains a LoRA adapter (rank 64, alpha 128, all linear layers) for Qwen/Qwen2.5-Omni-3B.

Training	DPO (beta 0.2, rpo_alpha 1.0), lr 5e-5, total batch size 32
Training data	10k MSP-Podcast v2.0 preference pairs per attribute (30k total), target: answer only
MSP-Podcast test (A / V / D / Avg)	88.5 / 88.8 / 86.3 / 87.9
BIIC-Podcast / WHiSER (Avg)	75.7 / 91.1

Usage

Merge the adapter and evaluate with the code repository:

swift export --adapters Lab-MSP/comparative-reasoning-dpo --merge_lora true \
  --model Qwen/Qwen2.5-Omni-3B --output_dir outputs/experiments/dpo_merged
bash src/eval.sh msp_test outputs/experiments/dpo_merged

Prompt format (two audios followed by the question):

system: You are a helpful assistant for emotion comparative reasoning.
user:   <audio><audio>You will hear two audio clips. Clip 1 is the first audio clip. Clip 2 is the second audio clip. <attribute definition> Which clip has higher <attribute>?

The model answers <answer> Clip 1 </answer> or ... Clip 2 ....

Citation

@inproceedings{naini2026comparative,
  title     = {Comparative Reasoning: Making an Audio Language Model Better at Comparing Emotions},
  author    = {Naini, Abinay Reddy and Kim, Jaeyeon and Yang, Chao-Han Huck and Watanabe, Shinji and Busso, Carlos},
  booktitle = {Interspeech},
  year      = {2026}
}