Multimodal Chain-of-Thought (MCoT) substantially enhances the complex reasoning capabilities of Multimodal Large Language Models (MLLMs) by explicitly modeling intermediate reasoning processes. However, most learning-based MCoT methods follow a stage-centric paradigm, where different reasoning behaviors are assigned to predefined stages and optimized separately. This results in fragmented reasoning and a training-inference reasoning-state gap. To address these limitations, we propose ReState, a unified iterative reasoning framework that reformulates multimodal reasoning from a stage-driven pipeline into a state-driven dynamic process. It comprises two complementary components: Unified Reasoning State Modeling (URSM) and Closed-Loop Reasoning-State Learning (CRSL). URSM unifies reasoning construction, correct reasoning preservation, and erroneous reasoning correction as reasoning-state transitions within a shared state space, enabling these heterogeneous reasoning behaviors to be learned jointly. CRSL further incorporates self-generated historical reasoning states during training and recursively updates the current state during inference, thereby mitigating the reasoning-state gap and enabling iterative self-refinement without requiring an additional reflection module, critic, or independent correction stage. Furthermore, we construct VersaCoT-10K, a high-quality CoT dataset spanning eight visual reasoning domains, which provides diverse supervision for multi-domain reasoning and reasoning-state refinement. Extensive experiments demonstrate that ReState exhibits strong adaptability across different MLLM backbones and achieves strong performance on multiple public reasoning benchmarks, supporting the shift from stage-centric reasoning toward state-centric reasoning. Our data and code are available at https://github.com/YuChuang1205/ReState
Figures & tables
Figure 1: Comparison of reasoning paradigms and performance. The stage-centric paradigm assigns different reasoning behaviors to predefined stages, leading to fragmented modeling and a training-inference reasoning-state gap. In contrast, ReState proposes a state-centric paradigm that recursively refines explicit reasoning states with a unified state-update operator, unifying reasoning construction, preservation, and correction while leveraging model-generated historical states for closed-loop learning. Across three backbones, ReState improves performance by 0.90%-4.60% , 0.83%-2.87% , and 0.18%-13.07% on VersaCoT-10K, ScienceQA, and M 3 CoT, respectively.
Figure 2: Overview of the ReState framework. During training, URSM learns a shared state-update operator from empty, correct, and erroneous reasoning states, while CRSL feeds model-generated historical states back into training to reduce the training-inference reasoning-state mismatch. At inference, the same model recursively updates the current reasoning state and answer, enabling iterative self-refinement without additional critics, reflection modules, or separate correction stages.
Figure 3: Construction pipeline of VersaCoT-10K dataset. It consists of three stages: Data Collection and Standardization , CoT Annotation and Quality Control , and Reasoning-State Augmentation .
Model
Learning
Size
MED
DRIVE
COMMON
DIAG
SPAT
GEO
CHART
MULTI
Avg
Random
-
-
22.00
46.80
26.80
26.27
39.87
22.93
22.00
25.60
29.03
Human
-
-
66.00
89.20
90.40
88.00
92.00
94.40
96.00
85.20
87.65
Open-source Models + Zero-shot
Qwen2.5-VL-7B
Zero-shot
7B
48.00
52.00
88.80
72.40
76.80
70.40
82.40
48.80
67.45
Qwen3-VL-8B
Zero-shot
8B
51.60
57.20
88.80
79.20
89.20
68.80
83.60
54.00
71.55
Qwen3.5-9B
Zero-shot
9B
61.60
59.60
87.60
81.20
83.60
70.80
84.80
54.00
72.90
Table 1: Main results (%) on the VersaCoT-10K dataset. MED = medical, DRIVE = driving and safety, COMMON = visual commonsense, DIAG = scientific diagram, SPAT = spatial perception, GEO = geometric and mathematical, CHART = chart computation, and MULTI = multidisciplinary knowledge. Red denotes the best result, and blue denotes the second best result.
Model
Learning
Size
NAT
SOC
LAN
TXT
IMG
NO
G1-6
G7-12
Avg
Random
-
-
40.28
46.13
29.25
47.45
40.08
33.66
39.35
40.67
39.83
Human
-
-
90.23
84.97
87.48
89.60
87.50
88.10
91.59
82.42
88.40
ViLT
Fine-tune
113M
60.48
63.89
60.27
63.20
61.38
57.00
60.72
61.90
61.14
Patch-TRM
Fine-tune
90M
65.19
46.79
65.55
66.96
55.28
64.95
58.04
67.50
61.42
VisualBERT
Fine-tune
111M
59.33
69.18
61.18
62.71
62.17
58.54
62.96
59.92
61.87
LLaMA-Adapter
Fine-tune
6B
84.37
88.30
84.36
83.72
80.32
86.90
85.83
84.05
85.19
Table 2: Main results (%) on the ScienceQA dataset. NAT = natural science, SOC= social science, LAN = language science, TXT = text context, IMG = image context, NO = no context, G1-6 = grades 1-6, G7-12 = grades 7-12. Red denotes the best result, and blue denotes the second best result.
Model
Size
Science
Commonsense
Mathematics
Lang
Natural
Social
Physical
Social
Temporal
Algebra
Geometry
Theory
Total
Random
-
32.70
30.62
26.71
32.97
22.22
20.33
35.71
27.50
23.81
28.56
Human
-
97.83
92.62
94.31
96.28
92.41
88.71
87.23
88.75
85.71
91.61
Tool-Usage Methods
HuggingGPT
175B
17.57
20.93
10.33
8.70
14.75
9.76
11.35
22.50
9.52
14.60
VisualChatGPT
>175B
30.09
36.28
7.78
43.48
29.92
33.33
21.99
21.25
28.57
25.92
Table 3: Main results (%) on the M 3 CoT dataset. The “Random” and “Human” results are averaged over three trials. Red denotes the best result, and blue denotes the second best result.
Scheme
URSM
CRSL
Qwen2.5-VL-7B
Qwen3-VL-8B
Qwen3.5-9B
SHS
ISR
Baseline
✗
✗
✗
75.15 ( +1.90 )
75.65 ( +3.60 )
77.80 ( +4.10 )
Baseline w/ URSM
✓
✗
✗
75.65 ( +1.40 )
78.20 ( +1.05 )
78.95 ( +2.95 )
Baseline w/ URSM, CRSL(SHS)
✓
✓
✗
76.20 ( +0.85 )
78.50 ( +0.75 )
81.10 ( +0.80 )
Baseline w/ URSM, CRSL(ISR)
✓
✗
✓
76.20 ( +0.85 )
78.65 ( +0.60 )
79.45 ( +2.45 )
ReState (Final)
✓
✓
✓
77.05
79.25
81.90
Table 4: The Break-Down ablation of ReState on the VersaCoT-10K dataset. SHS and ISR denote the training-time self-history state mechanism and inference-time iterative self-refinement in CRSL.
Figure 4: Performance evolution over iterative self-refinement rounds across different MLLMs and datasets. Round 0 denotes the initial reasoning, with most gains achieved in the first few rounds.
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Prompt Design for CoT Annotation during the construction of the VersaCoT-10K dataset. Italicized text represents assignment variables.
Figure 6: Prompt Design for CoT filtering during the construction of the VersaCoT-10K dataset. Italicized text represents assignment variables. A quality score of 8 or higher is deemed acceptable.
Figure 7: Statistics of VersaCoT-10K dataset. Left: data sources and splits across eight reasoning categories. Middle: CoT length distribution. Right: distribution of candidate-option counts.
Scheme
URSM
CRSL
Qwen2.5-VL-7B
Qwen3-VL-8B
Qwen3.5-9B
SHS
ISR
Baseline
✗
✗
✗
79.42 ( +5.57 )
81.62 ( +6.34 )
84.73 ( +4.27 )
Baseline w/ URSM
✓
✗
✗
84.12 ( +0.87 )
86.58 ( +1.38 )
88.65 ( +0.35 )
Baseline w/ URSM, CRSL(SHS)
✓
✓
✗
84.17 ( +0.82 )
86.67 ( +1.29 )
88.83 ( +0.17 )
Baseline w/ URSM, CRSL(ISR)
✓
✗
✓
84.43 ( +0.56 )
87.14 ( +0.82 )
88.65 ( +0.35 )
ReState (Final)
✓
✓
✓
84.99
87.96
89.00
Appendix
Table 5: The Break-Down ablation of ReState on the M 3 CoT dataset. SHS and ISR denote the training-time self-history state mechanism and inference-time iterative self-refinement in CRSL.
Scheme
URSM
CRSL
Qwen2.5-VL-7B
Qwen3-VL-8B
Qwen3.5-9B
SHS
ISR
Baseline
✗
✗
✗
96.11 ( +1.32 )
96.89 ( +1.27 )
95.59 ( +2.78 )
Baseline w/ URSM
✓
✗
✗
97.31 ( +0.12 )
97.95 ( +0.21 )
98.04 ( +0.33 )
Baseline w/ URSM, CRSL(SHS)
✓
✓
✗
97.38 ( +0.05 )
98.07 ( +0.09 )
98.37 ( +0.00 )
Baseline w/ URSM, CRSL(ISR)
✓
✗
✓
97.31 ( +0.12 )
97.95 ( +0.21 )
98.07 ( +0.30 )
ReState (Final)
✓
✓
✓
97.43
98.16
98.37
Appendix
Table 6: The Break-Down ablation of ReState on the ScienceQA dataset. SHS and ISR denote the training-time self-history state mechanism and inference-time iterative self-refinement in CRSL.
Scheme
ϕ
R+
R−
Rh
Qwen2.5-VL-7B
Qwen3-VL-8B
Qwen3.5-9B
ReState (Final) w/o ϕ
✗
✓
✓
✓
75.70 ( +1.35 )
77.95 ( +1.30 )
80.15 ( +1.75 )
ReState (Final) w/o R+
✓
✗
✓
✓
75.65 ( +1.40 )
78.50 ( +0.75 )
79.50 ( +2.40 )
ReState (Final) w/o R−
✓
✓
✗
✓
76.45 ( +0.60 )
78.90 ( +0.35 )
81.75 ( +0.15 )
ReState (Final) w/o Rh
✓
✓
✓
✗
76.20 ( +0.85 )
78.65 ( +0.60 )
79.45 ( +2.45 )
ReState (Final)
✓
✓
✓
✓
77.05
79.25
81.90
Appendix
Table 7: Ablation of different reasoning states in ReState on VersaCoT-10K. ϕ , R+ , R− , and Rh denote empty, positive, negative, and self-generated historical reasoning states, respectively.
Model
Learning
Size
MED
DRIVE
COMMON
DIAG
SPAT
GEO
CHART
MULTI
Avg
Random
-
-
22.00
46.80
26.80
26.27
39.87
22.93
22.00
25.60
29.03
Human
-
-
66.00
89.20
90.40
88.00
92.00
94.40
96.00
85.20
87.65
Open-source Models + Zero-shot
Qwen2.5-VL-7B
Zero-shot
7B
48.00
52.00
88.80
72.40
76.80
70.40
82.40
48.80
67.45
Qwen3-VL-8B
Zero-shot
8B
51.60
57.20
88.80
79.20
89.20
68.80
83.60
54.00
71.55
Qwen3-VL-32B
Zero-shot
32B
60.00
51.20
87.20
81.20
77.60
80.40
92.80
53.60
73.00
Appendix
Table 8: Main results (%) on the VersaCoT-10K dataset. Learning = Learning and training methods. Size = backbone model size. MED = medical, DRIVE = driving and safety, COMMON = visual commonsense, DIAG = scientific diagram, SPAT = spatial and perceptual, GEO = geometric and mathematical, CHART = chart and quantitative, and MULTI = multidisciplinary. Red denotes the best result, and blue denotes the second best result.
Model
Learning
Size
NAT
SOC
LAN
TXT
IMG
NO
G1-6
G7-12
Avg
Random
-
-
40.28
46.13
29.25
47.45
40.08
33.66
39.35
40.67
39.83
Human
-
-
90.23
84.97
87.48
89.60
87.50
88.10
91.59
82.42
88.40
MCAN
Fine-tune
95M
56.08
46.23
58.09
59.43
51.17
55.40
51.65
59.72
54.54
Top-Down
Fine-tune
70M
59.50
54.33
61.82
62.90
54.88
59.79
57.27
62.16
59.02
BAN
Fine-tune
112M
60.88
46.57
66.64
62.61
52.60
65.51
56.83
63.94
59.37
DFAF
Fine-tune
74M
64.03
48.82
63.55
65.88
54.49
64.11
57.12
67.17
60.72
Appendix
Table 9: Main results (%) on the ScienceQA dataset. Learning = Learning and training methods. Size=backbone model size. NAT = natural science, SOC= social science, LAN = language science, TXT = text context, IMG = image context, NO = no context, G1-6 = grades 1-6, G7-12 = grades 7-12. Red denotes the best result, and blue denotes the second best result.
Science
Commonsense
Mathematics
Model
Size
Lang
Natural
Social
Physical
Social
Temporal
Algebra
Geometry
Theory
Total
Random
-
32.70
30.62
26.71
32.97
22.22
20.33
35.71
27.50
23.81
28.56
Human
-
97.83
92.62
94.31
96.28
92.41
88.71
87.23
88.75
85.71
91.61
Tool-Usage Methods
HuggingGPT
175B
17.57
20.93
10.33
8.70
14.75
9.76
11.35
22.50
9.52
14.60
VisualChatGPT
>175B
30.09
36.28
7.78
43.48
29.92
33.33
21.99
21.25
28.57
25.92
Appendix
Table 10: Main results (%) on the M 3 CoT dataset. The “Random” and “Human” results are averaged over three trials. Red denotes the best result, and blue denotes the second best result.
Figure 8: Qualitative comparison on an example of small loan dilemma reasoning in VersaCoT-10K.
Figure 9: Qualitative comparison on an example of rock cycle reasoning in VersaCoT-10K.
Figure 10: Qualitative comparison on an example of fossil age reasoning in ScienceQA.
Figure 11: Qualitative comparison on an example of solution concentration reasoning in ScienceQA.
Figure 12: Qualitative comparison on an example of species recognition in M 3 CoT.
Figure 13: Qualitative comparison on an example of tangram recognition in M 3 CoT.
Figure 14: Iterative reasoning-state refinement of ReState on VersaCoT-10K.
Figure 15: Iterative reasoning-state refinement of ReState on ScienceQA.
Figure 16: Iterative reasoning-state refinement of ReState on M 3 CoT.
Setting
VersaCoT-10K → ScienceQA
VersaCoT-10K → M 3 CoT
Qwen2.5-VL-7B
Qwen3-VL-8B
Qwen3.5-9B
Qwen2.5-VL-7B
Qwen3-VL-8B
Qwen3.5-9B
Origin (No SFT)
84.39 ( +0.66 )
90.62 ( -0.52 )
91.25 ( +0.69 )
59.79 ( +1.38 )
61.43 ( +0.69 )
62.81 ( +2.29 )
Origin (Direct SFT)
84.70 ( +0.35 )
89.55 ( +0.55 )
91.54 ( +0.40 )
58.93 ( +2.24 )
57.98 ( +4.14 )
63.11 ( +1.99 )
ReState (Ours)
85.05
90.10
91.94
61.17
62.12
65.10
Appendix
Table 11: Cross-dataset generalization of ReState from VersaCoT-10K to ScienceQA and M 3 CoT.
Model
Overall
Correctness
Relevance
Coherence
Solving Process
Origin (No SFT)
7.30 ( +0.82 )
5.85 ( +0.97 )
8.44 ( +0.61 )
8.62 ( +0.34 )
6.29 ( +1.36 )
Origin (Direct SFT)
7.82 ( +0.30 )
6.44 ( +0.38 )
8.84 ( +0.21 )
8.61 ( +0.35 )
7.38 ( +0.27 )
Multimodal-CoT
7.56 ( +0.56 )
6.29 ( +0.53 )
8.37 ( +0.68 )
8.40 ( +0.56 )
7.20 ( +0.45 )
MIND
8.00 ( +0.12 )
6.63 ( +0.19 )
9.01 ( +0.04 )
8.88 ( +0.08 )
7.49 ( +0.16 )
ReState (Ours)
8.12
6.82
9.05
8.96
7.65
Appendix
Table 12: CoT quality evaluation on VersaCoT-10K with Qwen2.5-VL-7B.
Model
Overall
Correctness
Relevance
Coherence
Solving Process
Origin (No SFT)
8.21 ( +0.64 )
6.93 ( +0.80 )
9.19 ( +0.28 )
9.15 ( +0.28 )
7.55 ( +1.23 )
Origin (Direct SFT)
8.49 ( +0.36 )
7.25 ( +0.48 )
9.28 ( +0.19 )
9.13 ( +0.30 )
8.29 ( +0.49 )
Multimodal-CoT
8.27 ( +0.58 )
7.28 ( +0.45 )
8.87 ( +0.60 )
8.85 ( +0.58 )
8.06 ( +0.72 )
MIND
8.35 ( +0.50 )
7.57 ( +0.16 )
9.13 ( +0.34 )
8.65 ( +0.78 )
8.04 ( +0.74 )
ReState (Ours)
8.85
7.73
9.47
9.43
8.78
Appendix
Table 13: CoT quality evaluation on VersaCoT-10K with Qwen3-VL-8B.
Model
Overall
Correctness
Relevance
Coherence
Solving Process
Origin (No SFT)
7.97 ( +0.54 )
6.68 ( +0.64 )
9.01 ( +0.31 )
8.78 ( +0.51 )
7.42 ( +0.71 )
Origin (Direct SFT)
8.25 ( +0.26 )
6.81 ( +0.51 )
9.12 ( +0.20 )
9.06 ( +0.23 )
8.00 ( +0.13 )
Multimodal-CoT
8.17 ( +0.34 )
7.04 ( +0.28 )
8.81 ( +0.51 )
8.91 ( +0.38 )
7.93 ( +0.20 )
MIND
8.42 ( +0.09 )
7.22 ( +0.10 )
9.25 ( +0.07 )
9.25 ( +0.04 )
7.98 ( +0.15 )
ReState (Ours)
8.51
7.32
9.32
9.29
8.13
Appendix
Table 14: CoT quality evaluation on VersaCoT-10K with Qwen3.5-9B.
Chain-of-thought (CoT) reasoning has enabled multi-modal large language models (MLLMs) to tackle complex visual reasoning tasks by generating explicit intermediate reasoning steps in natural language. However, this text-based reasoning paradigm is inherently slow at inference time with even thousands of tokens and fundamentally constrained by the expressiveness of natural language. In this paper, we propose CoLT, (Chain of Latent Thoughts), a novel framework that teaches multi-modal models to reason through a chain of latent thought representations instead of verbose text tokens, which can perform thinking with as few as 3 steps. Naively forcing the model to think with latent states easily produces meaningless semantics and makes training unstable. To effectively regulate the latent reasoning process, we introduce a lightweight external decoder that provides step-level supervision for each latent reasoning step in two complementary directions: a forward mode that decodes latent thoughts into the textual reasoning of the next step, and a backward mode that aligns decoder hidden states with the model's latent thoughts given preceding textual context. We further incorporate internal supervision that encourages coherent step-by-step latent transitions. The decoder and internal supervision are removed during inference to maintain high efficiency of latent reasoning. Extensive experiments on eight benchmarks demonstrate that CoLT not only outperforms existing latent reasoning methods such as CODI and SIM-CoT, but also surpasses latent visual reasoning approaches that rely on auxiliary images with costly annotation requirements. Compared to text CoT methods, CoLT can notably reduce the inference time by 10.1× and text decoding time by 22.6×. Code is released at https://github.com/hulianyuyy/CoLT.
Lianyu Hu, Shengqian Qin, Zeqin Liao +4
Nanyang Technological University, Singapore · Shanghai Jiao Tong University, China · NKIARI, Shenzhen Futian, China +2
The effectiveness of Chain-of-Thought (CoT) prompting in Multimodal Large Language Models (MLLMs) remains uncertain: across several visual reasoning benchmarks, CoT prompting often degrades performance compared to direct prompting. In this paper, we provide a systematic analysis of CoT behavior in three modern MLLM families across model scales on datasets requiring step-wise visual evidence. Our analysis identifies two recurring failure modes: premature answer commitment and limited direct visual-token access during rationale generation. We further find that standard CoT-style Supervised Fine-Tuning (CoT-SFT) can mitigate these issues only partially, while often increasing reliance on textual priors and reducing counterfactual visual dependence. Motivated by these findings, we propose Attentive-CoT (Att-CoT), an attention-guided fine-tuning objective that encourages CoT trajectories to delay answer commitment while maintaining sustained visual-token access. Att-CoT can be plugged into any CoT-SFT training run without architectural changes. Experiments on three visual reasoning benchmarks across six MLLMs show that Att-CoT enhances CoT performance over standard fine-tuning.
Reasoning in Multimodal Large Language Models (MLLMs) requires both fine-grained visual perception and rigorous logical deduction. Explicit text-based Chain-of-Thought (CoT) is computationally expensive and prone to visual hallucinations, while existing latent reasoning methods typically require costly training. Furthermore, directly adapting training-free LLM reasoning mechanisms to the multimodal setting yields unstable performance. We identify that this failure stems from their reliance on token-level entropy, which fundamentally conflates perceptual ambiguity (e.g., unclear visual details) with logical uncertainty (e.g., complex reasoning steps). To overcome this bottleneck, we present a novel training-free inference strategy for MLLMs that explicitly decouples perception and reasoning. We propose a novel metric, the vision-to-text attention ratio, to dynamically gauge the model's cognitive focus. Guided by this metric, our proposed framework, Attention-Guided Switching (AGS), adaptively triggers latent reasoning for perceptual tokens to preserve high-fidelity visual information in the continuous space, while enforcing explicit text generation for logical tokens to maintain structural anchoring. Extensive experiments demonstrate that our method achieves state-of-the-art performance, significantly improving both accuracy and inference efficiency by reducing autoregressive steps and latency. Code is released at https://github.com/swordAndSnow/MM26-AGS.
Haoqian Kang, Liupeng Li, Kuofeng Gao +5
Harbin Institute of Technology, Shenzhen Shenzhen, China · Tsinghua Shenzhen International Graduate School, Tsinghua University Shenzhen, China · Peng Cheng Laboratory Shenzhen, China +1