Model details
Table | |
|---|
| Base model | Qwen/Qwen3.5-9B |
| Adaptation | LoRA, rank 32, α = 64, dropout 0.05, applied to all linear layers of the language model and the visual merger |
| Released weights | LoRA merged into the base weights (bf16, 4 shards); no adapter loading required |
| Training data | X-AQA-40K: 40,163 samples, 22 sources, 4 domains, one epoch (1,255 steps at global batch 32) |
| Optimization | AdamW, weight decay 0.1, gradient clipping 1.0, peak LR 1e-4 (language model) / 5e-5 (visual merger), cosine decay with 3% warmup and a minimum of 10% of the peak |
| Hardware | 16 × NVIDIA H20, bf16, maximum sequence length 32,768, seed 42 |
Results
X-AQA-Bench
X-AQA-Bench contains 7,101 evaluation rows (5,366 ID, 1,735 OOD) over 31 tasks and 25 source datasets. ID covers sources represented in X-AQA-40K; OOD holds out T3Set, Fitness-7, KIMORE and JIGSAWS. All values are multiplied by 100; ± is the half-width of the 95% bootstrap confidence interval.
Table with columns: Split, Score SRCC↑, Category macro-F1↑, Category κ↑, Text BERTScore↑| Split | Score SRCC↑ | Category macro-F1↑ | Category κ↑ | Text BERTScore↑ |
|---|
| ID | 60.7 ± 3.6 | 59.2 ± 1.8 | 29.3 ± 3.2 | 89.7 ± 0.1 |
| OOD | 24.4 ± 8.5 | 55.7 ± 5.2 | 12.6 ± 9.9 | 87.1 ± 0.1 |
X-AQA-9B outperforms all evaluated open-source MLLMs on all four metrics on both splits. On the ID split it reaches an SRCC of 0.607, compared with 0.465 for GPT-6.1 Sol; on the OOD split its BERTScore is 0.871, compared with 0.862 for Gemini 3.8 Flash.
Dataset-specific evaluation
The same jointly trained model, without dataset-specific fine-tuning, evaluated on the full test sets with each dataset's own metric:
Table with columns: Dataset, Metric, X-AQA-9B, Comparison| Dataset | Metric | X-AQA-9B | Comparison |
|---|
| MTL-AQA | SRCC↑ / relative ℓ2↓ | 0.8862 / 0.6948 | 0.6464 / 1.8750 (Qwen3.5-9B base) |
| Fitness-AQA | F1↑ (knees-forward error) | 0.8130 | 0.8338 (MD specialist) |
| KIMORE (zero-shot) | SRCC↑ on E1–E5 | 0.35 / 0.45 / 0.47 / 0.50 / 0.18 | 0.54 on E4 (Abedi et al., trained on KIMORE) |
| LASANA | Accuracy↑ / balanced accuracy↑ | 0.7706 / 0.7576 | 0.5512 / 0.5669 (Qwen3.5-9B base) |
Code
Dataset preparation, video preprocessing, model training and inference code are available at
github.com/Kelly510/X-AQA. See the repository for the
frame-sampling protocol and serving examples.
Citation
@misc{zheng2026xaqa, title = {X-AQA: Unifying Action Quality Assessment with Multimodal Large Language Models and Multi-Domain Learning}, author = {Kaili Zheng and Wei Ding and Yuchen Ding and Kaiwen Wang and Xun Zhu and Chenyi Guo and Yudong Zhang and Ji Wu}, year = {2026}, eprint = {XXXX.XXXXX}, archivePrefix = {arXiv}}