Large Language Models (LLMs) have achieved remarkable reliability and advanced capabilities through extended test-time reasoning. However, extending these capabilities to Multi-modal Large Language Models (MLLMs) remains a significant challenge due to a critical scarcity of high-quality, long-chain reasoning data and optimized training pipelines. To bridge this gap, we present a unified multi-agent visual reasoning framework that systematically evolves from our foundational image-centric model, Insight-V, into a generalized spatial-temporal architecture, Insight-V++. We first propose a scalable data generation pipeline equipped with multi-granularity assessment that autonomously synthesizes structured, complex reasoning trajectories across image and video domains without human intervention. Recognizing that directly supervising MLLMs with such intricate data yields sub-optimal results, we design a dual-agent architecture comprising a reasoning agent to execute extensive analytical chains, and a summary agent to critically evaluate and distill final outcomes. While our initial framework utilized Direct Preference Optimization (DPO), its off-policy nature fundamentally constrained reinforcement learning potential. To overcome these limitations, particularly for long-horizon video understanding, Insight-V++ introduces two novel algorithms, ST-GRPO and J-GRPO, which enhance spatial-temporal reasoning and improve evaluative robustness. Crucially, by leveraging reliable feedback from the summary agent, we guide an iterative reasoning path generation process, retraining the entire multi-agent system in a continuous, self-improving loop. Extensive experiments on base models like LLaVA-NeXT and Qwen2.5-VL demonstrate significant performance gains across challenging image and video reasoning benchmarks while preserving strong capabilities on traditional perception-focused tasks.
Figures & tables
Fig. 1 : Illustration and Performance of Insight-V and Insight-V++. Insight-V employs a dual-agent architecture, comprising dedicated reasoning and summarization modules to achieve significant performance gains across various image reasoning benchmarks. Building upon this foundation, Insight-V++ introduces a novel self-evolving training paradigm. By utilizing an iterative loop of Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) to continuously refine its visual reasoning capabilities, Insight-V++ yields superior results on both image and video benchmarks.
Fig. 2 : Data Generation Pipeline of Insight-V Series. The reasoning processes are generated progressively through a reasoning generator, and then fed into a multi-granularity assessment system to ensure high-quality reasoning.
Fig. 3 : Overview of Insight-V and Insight-V++ Model Design. In Insight-V, we derive a multi-agent system from a single base model. By decomposing the task into reasoning and summarization, these two agents collaborate to enhance overall performance. For Insight-V++, we further introduce ST-GRPO and J-GRPO, which are tailored specifically to each agent. Coupled with a self-evolving strategy, this training loop consisting of SFT and customized RL drives the multi-agent system to excel at visual reasoning.
Model
Size
Reasoning
Perception
Average
MMMU val
MMMU-Pro avg
MMBench
ChartQA
MMStar
MathVista testmini
MME
TextVQA
OCRBench
AI2D
Reasoning
Perception
DeepSeek-VL [ 63 ]
7B
35.4
-
73.5
59.1
37.1
36.1
-/-
-
-
65.3
-
-
VILA-1.5 [ 50 ]
8B
38.6
-
75.3
-
39.7
-
1634.9/-
-
-
58.8
-
-
Cambrian-1 [ 91 ]
8B
42.7
-
75.9
73.3
-
49.0
1547.1/-
72.6
-
74.6
-
-
InternLM-XComposer2 [ 14 ]
7B
41.1
-
77.6
71.8
56.2
59.5
2220.4
69.7
-
81.2
-
-
POINTS [ 57 ]
7B
51.4
-
78.0
-
60.9
63.0
2184.1
60.0
-
81.2
-
-
TABLE I : Results on General Image Reasoning and Understanding Tasks. We evaluate 10 benchmarks covering both visual reasoning and perception. Average scores are computed separately for the Reasoning and Perception groups with normalized scales for fair comparison. Insight-V significantly improves reasoning while maintaining strong perception, and Insight-V++ further advances performance on stronger base models.
Model
Size
MathVision test
MathVerse testmini
WeMath Loose
LogicVista test
DynaMath worst
CharXiv Reasoning
Average
Closed-source Models
GPT-4o [ 36 ]
-
31.1
40.6
-
64.4
34.5
-
-
GPT-4o mini [ 36 ]
-
27.3
30.0
48.8
41.4
31.6
34.1
-
Open-source Models – SFT
LLaVA-OneVision [ 41 ]
7B
17.6
17.6
-
32.0
9.0
23.6
-
InternVL3-8B [ 139 ]
8B
28.6
33.9
-
43.6
23.0
37.6
-
TABLE II : Results on Advanced Image Reasoning Tasks. We conduct evaluation experiments across 6 visual reasoning benchmarks, covering mathematical, logical, and chart-based advanced reasoning assessments. Insight-V++ achieves the highest average score, outperforming other strong RL-based methods. † means the results are reproduced by OpenMMReasoner [ 126 ] .
Model
Size
VideoMME
VideoMMMU
MMVU MC
VideoMMLU
VideoHolmes
Video-TT MC
Average
Proprietary Models
GPT-4o [ 72 ]
-
71.9
61.2
75.4
49.4
42.0
46.6
57.8
Gemini-1.5-Pro [ 86 ]
-
75.0
53.9
-
-
41.2
42.3
-
Gemini-1.5-Flash [ 86 ]
-
70.3
49.8
-
43.6
-
-
-
Gemini-2.5-Pro [ 87 ]
-
84.3
83.6
-
-
45.0
-
-
Open-Sourced Video MLLMs
TABLE III : Results on Video Reasoning Tasks. We conduct evaluation experiments across 6 video benchmarks, covering both general video understanding and task-specific video reasoning assessments. Insight-V++ demonstrates strong effectiveness and generalizability in video domain as well, achieving consistent improvements across all benchmarks and surpassing competitive open-source video MLLMs at the similar scale by a notable margin.
Model
MMMU
ChartQA
MathVista
MMStar
Avg
Baseline
47.1
75.7
56.9
57.0
59.2
Vanilla - Direct SFT
47.0
79.2
57.6
58.4
60.6
Multi-Turn Supervised
48.1
79.6
57.9
58.2
61.0
Summary Agent Only
47.5
76.3
57.3
57.9
59.8
Multi-Agent
49.7
81.2
58.7
58.6
62.1
TABLE IV : Ablations on the Design Choice of Insight-V Series. The multi-agent design outperforms other configurations, highlighting the critical role of reasoning and summarization decomposition.
Model
MMMU
ChartQA
MathVista
MMStar
Avg
Insight-V (Multi-Agent)
49.7
81.2
58.7
58.6
62.1
+ RLAIF
49.5
81.4
59.1
59.2
62.3
+ DPO
50.8
80.8
59.3
59.9
62.7
+ Iterative DPO
50.2
81.5
59.9
61.5
63.3
TABLE V : Ablations on the DPO training strategy. Iterative DPO progressively enhances the model’s reasoning capabilities, leading to improved performance.
Model
MMMU
MathVision
VideoMME
VideoMMMU
Insight-V++ (Multi-Agent)
59.2
38.7
65.8
49.6
+ ST-GRPO
61.7
42.9
66.7
52.0
+ ST-GRPO/J-GRPO
62.5
44.8
67.2
53.4
Collaborative Inference
63.2
46.7
67.5
53.9
Self-Evolving
64.8
48.6
67.8
55.8
TABLE VI : Ablations on the design choices of Insight-V++. The multi-agent system significantly benefits from the integration of the enhanced GRPO-based RL strategy and the proposed self-evolving mechanism.
Fig. 4 : Ablations on the amount of training data. The reasoning agent benefits from data scaling, providing more valuable insights for the summary agent.
Fig. 5 : Qualitative Results of Insight-V Series. We present qualitative comparisons of Insight-V with Chain-of-Thought and learning Insight-V with direct SFT (Vanilla). For the Insight-V system, the reasoning agent delivers a more coherent and structured reasoning process, guiding the summary agent toward the correct answer, whereas other methods struggle with complex reasoning tasks and fail to solve such challenging problems.
Multimodal large language models (MLLMs) have recently made strong progress in vision-language reasoning, yet their performance often degrades as generations grow longer. A key factor is that they frequently lose track of earlier visual evidence and intermediate constraints under a monolithic growing context. Inspired by how humans separately recall what they see and what they infer when solving complex tasks, we propose DLMR, a parameter-efficient mechanism that equips MLLMs with Dual Latent Memories: a visual memory that compresses image evidence and a reasoning memory that tracks intermediate conclusions and constraints. A Router then dynamically decides which memory and how much to reuse during inference, preserving visual grounding while maintaining coherent long-horizon reasoning. DLMR is trained in three stages, from latent memory construction to selective router learning, while keeping the base MLLM frozen, yielding substantial gains on both general and reasoning benchmarks with only a small number of additional trainable parameters. Analyses further show interpretable, state-dependent routing with specialized memory roles and reduced decoding tokens over long generations. Code is available at https://github.com/Hunter-Wrynn/DLMR.
Hao-Xuan Ma, Jin-Fei Qi, Yicheng Xiao +1
School of Artificial Intelligence, Nanjing University, China · National Key Laboratory for Novel Software Technology, Nanjing University, China · Institute of Automation, Chinese Academy of Sciences, Beijing, China
Multimodal Large Language Models (MLLMs) achieve strong multimodal reasoning performance, yet we identify a recurring failure mode in long-form generation: as outputs grow longer, models progressively drift away from image evidence and fall back on textual priors, resulting in ungrounded reasoning and hallucinations. Interestingly, Based on attention analysis, we find that MLLMs have a latent capability for late-stage visual verification that is present but not consistently activated. Motivated by this observation, we propose Visual Re-Examination (VRE), a self-evolving training framework that enables MLLMs to autonomously perform visual introspection during reasoning without additional visual inputs. Rather than distilling visual capabilities from a stronger teacher, VRE promotes iterative self-improvement by leveraging the model itself to generate reflection traces, making visual information actionable through information gain. Extensive experiments across diverse multimodal benchmarks demonstrate that VRE consistently improves reasoning accuracy and perceptual reliability, while substantially reducing hallucinations, especially in long-chain settings. Code is available at https://github.com/Xiaobu-USTC/VRE.
Shuai Lv, Chang Liu, Feng Tang +5
University of Science and Technology of China, Hefei, China · Huawei Foundation Model Department, Shanghai, China · The Chinese University of Hong Kong, Hong Kong, China +1
Large Vision-Language Models (LVLMs) have shown strong promise for multimodal reasoning, yet often struggle with tasks requiring concepts beyond what is directly observable in the input image. Existing methods generate intermediate images or latent visual tokens to guide reasoning, but these representations can introduce errors and increasingly interfere with textual reasoning as reasoning progresses. We propose Visual-to-Text Chain-of-Thought Distillation (V2T), a framework that enables LVLMs to internalize visual reasoning without generating intermediate visual representations at inference time. V2T first trains a teacher LVLM using interleaved visual and textual chains of thought, and then uses knowledge distillation to train a student LVLM using the teacher's logits and cross-entropy supervision from ground-truth textual reasoning. When reasoning images can be mapped to the original image, V2T can additionally distill the teacher's attention to corresponding regions, while ground-truth bounding boxes can further guide a subsequent reinforcement learning stage. Experiments across multiple multimodal reasoning benchmarks show that V2T consistently outperforms the teacher and existing baselines, improving average accuracy by 14.3% on a held-out set and 2.7% on the broader visual evaluation suite. Moreover, lightweight SFT and substantially reduced RL make V2T up to 42x faster to train than state-of-the-art baselines.
Wenhan Yang, Nilay Naharas, Ali Payani +1
University of California, Los Angeles · Cisco Research