MA-Bench: Towards Fine-grained Micro-Action Understanding
Authors: Kun Li, Jihao Gu, Fei Wang, Zhiliang Wu, Hehe Fan, Dan Guo
Organizations: CVLab, College of Information Technology, United Arab Emirates University · University College London · Hefei University of Technology · Institute of Artificial Intelligence, Hefei Comprehensive National Science Center · CCAI, Zhejiang University
With the rapid development of Multimodal Large Language Models (MLLMs), their potential in Micro-Action understanding, a vital role in human emotion analysis, remains unexplored due to the absence of specialized benchmarks. To tackle this issue, we present MA-Bench, a benchmark comprising 1,000 videos and a three-tier evaluation architecture that progressively examines micro-action perception, relational comprehension, and interpretive reasoning. MA-Bench contains 12,000 structured question-answer pairs, enabling systematic assessment of both recognition accuracy and action interpretation. The results of 23 representative MLLMs reveal that there are significant challenges in capturing motion granularity and fine-grained body-part dynamics. To address these challenges, we further construct MA-Bench-Train, a large-scale training corpus with 20.5K videos annotated with structured micro-action captions for fine-tuning MLLMs. The results of Qwen3-VL-8B fine-tuned on MA-Bench-Train show clear performance improvements across micro-action reasoning and explanation tasks. Our work aims to establish a foundation benchmark for advancing MLLMs in understanding subtle micro-action and human-related behaviors. Project Page: https://MA-Bench.github.io
Figures & tables
Figure 1: The pipeline for constructing MA-Bench and MA-Bench-Train for fine-grained micro-action understanding . (1) Micro-motion tracker extracts motion descriptors ( i.e . , motion vectors and coordinates) for each body part. (2) Micro-action benchmark generation leverages motion descriptors and multimodal large language models to create structured micro-action captions, which are then used to generate the benchmarks. (3) MA-Bench enables fine-grained action understanding through well-defined perceptual recognition , relational comprehension , and interpretive reasoning .
Benchmark
#Videos
Topic
#QA
#Annotation
MSVD-QA [ 61 ]
520
Daily Activities
13,157
Auto
MSRVTT-QA [ 61 ]
800
Daily Activities
8,000
Auto
ActivityNet-QA [ 65 ]
2,990
Daily Activities
72,821
Manual
NExT-QA [ 60 ]
1,000
Daily Activities
8, 564
Manual
Video-MME [ 14 ]
900
Daily Activities
2,700
Manual
EgoTaskQA [ 26 ]
2,315
Ego Activities
40,322
Manual
Table 1: Comparison of MA-Bench with existing video understanding benchmarks.
Figure 2: Data statistics of MA-Bench. Action category definitions are consistent with the Micro-Action-52 dataset [ 19 ] .
Figure 3: Example of question-answer pairs from the proposed MA-Bench.
Figure 4: The pipeline for semi-automatic question-answer generation.
Closed-ended
Open-ended
CMAR
FMAR
SAD
MAD
MAS
PPR
MADU
MARE
Methods
Date
Acc.
Acc.
Acc.
Acc.
Acc.
Acc.
AVG
L1
L2
L3
L1
L2
L3
AVG
Full mark
–
100
100
100
100
100
100
100
5
5
5
5
5
5
5
Random
–
14.7
20
50
50
50
50
39.05
–
–
–
–
–
–
–
Proprietary MLLMs
GPT-4o [ 25 ]
2024-11
20.50
30.70
51.30
62.35
49.25
55.10
44.87
0.60
0.65
0.54
1.02
0.91
0.67
0.73
Table 2: The overall performances of 21 MLLMs on MA-Bench, including proprietary and open-source models. “Acc.” denotes the accuracy metric. “AVG” means the average score of all the tasks. The best results among all MLLMs are indicated in bold .
Closed-ended
Open-ended
CMAR
FMAR
SAD
MAD
MAS
PPR
MADU
MARE
Exp.
#Frames
Acc
Acc
Acc
Acc
Acc
Acc
AVG
L1
L2
L3
L1
L2
L3
AVG
A1
4
31.73
36.64
56.11
53.50
47.85
53.83
46.61
0.49
0.51
0.46
1.09
0.93
0.77
0.71
A2
8
32.23
37.11
56.10
53.69
47.94
54.74
46.97
0.50
0.51
0.47
1.18
1.00
0.84
0.75
A3
16
31.49
36.11
56.02
53.71
47.94
54.96
46.71
0.48
0.49
0.44
1.13
0.98
0.82
0.72
Table 3: Ablation results of frames on Qwen3-VL-8B model.
Closed-ended
Open-ended
CMAR
FMAR
SAD
MAD
MAS
PPR
MADU
MARE
Exp.
Methods
SFT (LoRA)
Acc.
Acc.
Acc.
Acc.
Acc.
Acc.
AVG
L1
L2
L3
L1
L2
L3
AVG
B1
Qwen3-VL-8B [ 64 ]
-
32.23
37.11
56.10
53.69
47.94
54.74
46.97
0.50
0.51
0.47
1.18
1.00
0.84
0.75
B2
+ MA-Bench-Train
Vision Encoder
41.20
33.50
58.20
56.90
49.10
52.80
48.62
0.95
1.05
0.98
1.35
1.20
1.10
1.11
B3
+ MA-Bench-Train
LM Decoder
49.10
34.70
58.65
58.80
50.30
52.25
50.63
1.41
1.57
1.45
1.91
1.69
1.56
1.60
B4
+ MA-Bench-Train
All
47.90
32.60
60.30
59.65
50.00
53.60
50.68
1.50
1.67
1.54
1.98
1.78
1.67
1.69
Table 4: Supervised Fine-tuning results on MA-Bench using MA-Bench-Train.
Figure 5: Qualitative example on the task of Micro-Action Reasoning and Explanation (MARE). Green font denotes correct predictions, red indicates completely incorrect ones, yellow marks partial errors, black shows results with no impact, and purple highlights faulty reasoning chains. More examples of other tasks are provided in the supplementary material.
School of Computer Science and Information Engineering, Hefei University of Technology, Hefei, China · School of Information Science and Technology, University of Science and Technology of China, Hefei, China