Organizations: Computer Network Information Center, Chinese Academy of Sciences, Beijing, China · Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences, Zhejiang, China · LLM Team, Shopee Pte. Ltd., Shanghai, China
Image captioning remains a fundamental task for vision language understanding, yet ground-truth supervision still relies predominantly on human-annotated references. Because human annotations reflect subjective preferences and expertise, ground-truth captions are often incomplete or even incorrect, which in turn limits caption models. We argue that caption quality should be assessed by two objective aspects: completeness (does the caption cover all salient visual facts?) and correctness (are the descriptions true with respect to the image?). To this end, we introduce CCCaption: a dual-reward reinforcement learning framework with a dedicated fine-tuning corpus that explicitly optimizes these properties to generate Complete and Correct Captions. For completeness, we use diverse LVLMs to disentangle the image into a set of visual queries, and reward captions that answer more of these queries, with a dynamic query sampling strategy to improve training efficiency. For correctness, we penalize captions that contain hallucinations by validating the authenticity of sub-caption queries, which are derived from the caption decomposition. Our symmetric dual-reward optimization jointly maximizes completeness and correctness, guiding models toward captions that better satisfy these objective criteria. Extensive experiments across standard captioning benchmarks show consistent improvements, offering a principled path to training caption models beyond human-annotation imitation.
Figures & tables
Figure 1: (a) Comparison of incomplete and incorrect captions generated by MLLMs. An example of incomplete caption with missing information versus an incorrect caption with hallucinated details. (b) Reward and correctness curves during model training, illustrating the progression of reward, correctness (as defined in Section 3 ), and caption length across training steps.
Figure 2: Illustrations of our Complete and Correct Captioning (CCCaption) framework and Complete Caption dataset (CCaption) generation pipeline. (a) The CCCaption reinforcement learning framework. The green and red parts represent the processes for computing the correctness and completeness rewards, respectively. We combine these two rewards using the GRPO algorithm [ 13 ] to perform reinforcement learning training for the captioning model, while employing the dynamic query sampling strategy to enhance training efficiency. (b) The CCaption dataset generation pipeline. By iterating through the entire process, a diverse, complete query set is generated, ensuring thorough coverage of all the image information.
MLLM
Claude
CPMV
Cog2
Avg.
[ 1 ]
[ 36 ]
[ 61 ]
LLaVa-V1.6-34B [ 26 ]
-35.17
-4.17
-26.92
-22.08
Qwen2.5-VL-3B [ 2 ]
-25.33
13.58
-8.92
-6.89
CapRL-3B [ 53 ]
1.33
30.67
20.67
17.56
InternVL3.5-38b [ 49 ]
0.83
35.58
21.17
19.19
Qwen3-VL-2B [ 45 ]
32.50
47.17
44.50
41.39
Table 1: Results of the CapArena evaluation [ 9 ] across different captioning models. Bold numbers indicate the best performance, underlined numbers indicate the runner-up. Qwen3-VL-32B serves as the judge model. “Claude”, “CPMV”, and “Cog2” denote Claude-3.5-Sonnet-2024, MiniCPM-V2.6-8B, and CogVLM2-llama3-chat-19B, respectively.
MLLM
ChartQA
CharXiv
InfoVQA
Verse
MMB
MMMUPro
OCR
COCO
WM2Pro
Avg.
[ 33 ]
[ 50 ]
[ 34 ]
[ 62 ]
[ 29 ]
[ 60 ]
[ 30 ]
[ 15 ]
[ 38 ]
LLaVa-V1.6-34B [ 26 ]
52.60
26.33
34.46
27.45
71.61
33.33
34.95
37.08
32.67
38.94
Qwen2.5-VL-3B [ 2 ]
69.72
31.66
51.81
29.53
72.58
36.97
44.01
38.80
36.88
45.77
CapRL-3B [ 53 ]
73.92
35.11
55.94
33.16
79.16
41.01
50.16
52.53
38.64
51.07
InternVL3.5-38b [ 49 ]
74.84
37.30
59.40
34.91
77.04
41.62
50.81
47.28
38.82
51.34
Qwen3-VL-32B [ 45 ]
75.72
35.74
51.94
35.82
81.34
43.64
49.19
56.82
40.14
52.26
Table 2: Results of the Prism evaluation [ 39 ] across different captioning models. Bold numbers indicate the best performance, underlined numbers indicate the runner-up.
MLLM
AI2D
ChartQA
Hallusion
MMB
MME
MMMU
MMStar
OCR
WM2Pro
Avg.
[ 16 ]
[ 33 ]
[ 12 ]
[ 29 ]
[ 5 ]
[ 59 ]
[ 6 ]
[ 30 ]
[ 38 ]
LLaVa-V1.6-34B [ 26 ]
51.03
60.55
74.52
70.22
68.95
74.27
64.01
77.19
64.94
67.30
Qwen2.5-VL-3B [ 2 ]
48.71
61.73
74.76
68.96
65.81
75.77
63.46
78.09
72.66
67.77
CapRL-3B [ 53 ]
48.62
63.86
73.03
67.41
62.65
76.01
61.98
76.19
72.44
66.91
InternVL3.5-38b [ 49 ]
50.86
64.11
75.93
71.10
68.15
78.42
65.32
82.72
75.12
70.19
Qwen3-VL-32B [ 45 ]
52.38
66.82
75.45
69.20
64.38
79.03
64.37
82.35
74.87
69.87
Table 3: Results of the Hallucinations evaluation [ 19 ] across different captioning models. Bold numbers indicate the best performance, underlined numbers indicate the runner-up.
Figure 3: Case analysis across different captioning models. From left to right, the captioning models include Qwen3-VL-2B [ 55 ] , CapRL-3B [ 53 ] , Qwen3-VL-32B [ 55 ] , and CCCaption-2B (Ours), with image caption outputs for different queries labeled for hallucination, forgetfulness, and nitpicking. Both cases are derived from the MMBench dataset [ 29 ] .
Figure 4: Query and image embedding scatters across different datasets. “Vanilla” refers to queries generated by a single MLLM, using the CapRL method [ 53 ] , while “CCaption-44k” represents our approach, which utilizes multiple MLLMs for generation and employs a diversity metric to measure query diversity. The embedding model used is Ops-MM-embedding [ 44 ] , with dimensionality reduction performed using t-SNE [ 32 ] .
Training Data
CharXiv
InfoVQA
MMMU
MMStar
Avg.
[ 50 ]
[ 34 ]
[ 59 ]
[ 6 ]
w/o RL
31.66
51.78
39.64
49.93
43.25
Vanilla
32.92
51.94
40.96
50.47
44.07
CCaption-8k
35.42
53.23
40.80
47.73
44.30
CCaption-16k
35.74
52.91
40.29
50.80
44.94
CCaption-44k
36.99
55.18
41.07
50.53
45.94
Table 4: Ablation of training data under the Prism evaluation [ 39 ] . Bold numbers indicate the best performance, and underlined numbers indicate the runner-up. “Vanilla” refers to the queries generated according to the CapRL method [ 53 ] , while “CCaption” represents our approach. Qwen3-VL 2B [ 45 ] is used for training.
Figure 5: Performance whether the dynamic query sampling strategy is used during training. “w Dynamic” denotes the use of the strategy, while “w/o Dynamic” indicates the absence of the strategy. “Acc.” refers to the accuracy under the Prism evaluation [ 39 ] .
Component
ChartQA [ 33 ]
OCR [ 30 ]
Avg.
Prism
Hall.
Prism
Hall.
w/o RL
71.00
64.53
50.16
82.32
67.00
w/o Corr
74.56
65.84
47.90
81.93
67.56
w/o Comp
71.60
68.11
40.45
83.55
65.93
w/o Dynamic
73.92
66.33
52.43
83.92
69.15
Ours
75.12
66.72
55.34
83.09
70.07
Table 5: Ablation of key components in the CCCaption framework. “Prism” and “Hall.” refer to the results under the Prism [ 39 ] and Hallucinations [ 19 ] evaluations, respectively. “w/o Corr” and “w/o Comp” denote the use of only completeness and correctness rewards, respectively. “w/o Dynamic” indicates without the dynamic sampling strategy. Qwen3-VL 2B [ 45 ] is used for training.