Vision-Language Models increasingly rely on textual reasoning to solve advanced multimodal tasks. Despite its effectiveness, this paradigm faces two fundamental limitations. First, relying on textual reasoning alone can limit perceptual ability in vision-centric questions, where accurate answers depend on interpreting task-relevant visual evidence. Second, applying textual reasoning uniformly across inputs overlooks question-dependent variation in both reasoning modality and depth, often producing verbose and unnecessary traces when little logical inference is required. To this end, we propose Adaptive Visual-Textual Reasoning (AVTR), which introduces visual imagination as a complementary reasoning modality to textual thinking, enabling the model to revisit its visual observations and distill the salient evidence through generative visual reflection. Furthermore, we formulate reasoning-modality selection as a learned routing problem, allowing a single policy to adaptively harness the complementary strengths of visual reflection and textual reasoning in response to the perceptual and inferential demands of the input. Across 16 benchmarks, AVTR improves the average score by 4.5 points over Qwen3.5 baseline, while reducing inference latency by 33%. It also scales effectively into a smaller size backbone, outperforming the evaluated 2B-scale baselines.
Figures & tables
Figure 1: Adaptive multi-modal reasoning with visual imagination. Conventional VLMs perceive multimodal inputs but reason entirely in text. AVTR adaptively harnesses the complementary strengths of visual imagination and textual reasoning according to the perceptual and inferential demands of the input. Visual imagination forms a question-conditioned visual state through visual queries with training-time generative supervision, while requiring no pixel generation at inference.
Figure 2: Method overview. (a) Visual imagination is learned in two stages with shared trainable parameters but different inputs and supervision: visual replay reconstructs the input from visual queries, while visual reflection conditions them on the question to generate relevant visual evidence. (b) Imagination-guided SFT jointly optimizes visual-generation and response objectives, enabling the resulting visual state to guide answer generation. The connector and generator provide training-only supervision and are removed at inference. (c) Adaptive reasoning learns a correctness- and cost-aware policy that selects direct answering, visual imagination, textual reasoning, or visual imagination followed by textual reasoning for each input.
HighRes Image
Visual Spatial
Model
V ∗
HR-Bench-4K
HR-Bench-8K
BLINK
ERQA
MMSI-Bench
VSI-Bench
RealWorldQA
Proprietary
GPT-5
72.8
-
-
71.0
65.7
41.9
55.0
76.1
Gemini 2.5 Pro
83.8
87.3
85.4
73.5
55.7
38.0
53.6
68.7
GPT-5 mini
78.6
78.6
74.4
56.7
54.0
34.1
48.7
-
Open foundation models
Table 1: High-resolution image understanding and spatial reasoning.
Video Understanding (Acc)
Temporal Grounding (mIoU)
Model
MVBench
Video-MME
LongVideoBench
MLVU
Charades-STA
QVHighlights
ActivityNet-Cap.
NExT-GQA
Qwen3.5-9B (w/o Think)
69.1
72.4
67.8
79.1
61.8
67.2
49.8
38.4
Qwen3.5-9B (w/ Think)
71.8
75.4
69.1
80.0
55.2
63.5
42.1
34.4
InternVL3.5-8B
72.1
66.0
62.1
70.2
27.8
31.3
31.3
25.1
Video-R1
65.5
61.8
-
-
-
-
-
-
Time-R1
65.9
54.2
55.8
-
58.8
-
52.1
-
Table 2: Video understanding and temporal grounding.
Figure 3: Complementary strengths of visual and textual reasoning. Accuracy differences between the visual imagination and textual reasoning on representative subtasks with large performance gaps. Positive values favor visual imagination, while negative values favor textual reasoning.
Figure 4: Visual reflection for long-video understanding. The rendered image is used only for visualization. The answer is conditioned directly on the continuous visual queries at inference.
Figure 5: Comparison between visual imagination and textual reasoning. On this spatial task, textual reasoning produces a lengthy reasoning trace but still predicts the wrong camera-motion direction, whereas visual imagination reaches the correct answer with a compact visual state.
Figure 6: Efficiency analysis. Mean score over the 16 benchmarks versus average latency per query.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Visual Reconstruction Quality. Each example shows an input image, its associated image caption, and its reconstruction image.
Figure 8: Reconstruction quality across visual-query budgets. Each row shows the input followed by reconstructions with N=16 , 32 , and 64 queries.
Visual queries ( N )
8
16
32
64
128
SSIM
0.2563
0.3934
0.4217
0.5235
0.5632
Appendix
Table 8: Effect of the visual-query budget on reconstruction. Reconstruction fidelity measured by SSIM ( Wang et al., 2004 ) improves consistently as the number of continuous visual queries increases.
Figure 9: Visual reflection for temporal grounding. The rendered image is used only for visualization. The answer is conditioned directly on the continuous visual queries at inference.
Figure 10: Comparison between visual imagination and textual reasoning. On BLINK Visual Similarity, visual imagination preserves the fine-grained shape evidence and answers correctly, while the long textual trace reaches the wrong choice.
Figure 11: Comparison between visual imagination and textual reasoning. On BLINK Counting, textual reasoning returns the correct count, whereas visual imagination changes count-relevant details and produces an incorrect answer.
Figure 12: Generative visual imagination improves both accuracy and efficiency. (a) Accuracy on four high-resolution benchmarks; numbers indicate gains from enabling visual reasoning. (b) Average accuracy over the four benchmarks versus latency.
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
Visual reasoning models (VRMs) have recently shown strong cross-modal reasoning capabilities by integrating visual perception with language reasoning. However, they often suffer from overthinking, producing unnecessarily long reasoning chains for any tasks. We attribute this issue to Reasoning Path Redundancy in visual reasoning: many visual questions do not require the full reasoning process. To address this, we propose AVR, an adaptive visual reasoning framework that decomposes visual reasoning into three cognitive functions: visual perception, logical reasoning, and answer application. It further enables models to dynamically choose among three response formats: Full Format, Perception-Only Format, and Direct Answer. AVR is trained with FS-GRPO, an adaptation of Group Relative Policy Optimization that encourages the model to select the most efficient reasoning format while preserving correctness. Experiments on multiple vision-language benchmarks show that AVR reduces token usage by 50--90% while maintaining overall accuracy, especially in perception-intensive tasks. These results demonstrate that adaptive visual reasoning can effectively mitigate overthinking in VRMs. Code and data are available at: https://github.com/RunRiotComeOn/AVR.
Yixu Huang, Tinghui Zhu, Muhao Chen
♠Fudan University · ♣University of California, Davis
The introduction of long chain-of-thought (CoT) reasoning and external tools has substantially improved the performance of multimodal large language models (MLLMs) on general and reasoning tasks. However, existing fast-slow reasoning paradigms mainly regulate textual reasoning length, while visual computation introduced by external tools or visual tokens is often treated as an auxiliary component. This separation prevents MLLMs from jointly optimizing visual perception and textual reasoning. To address this issue, we propose L2T (adaptive Look, flexible Thought), a unified reasoning framework that jointly optimizes the form and depth of computation. L2T first discretizes tool-produced visual states into visual tokens, bringing visual perception and textual reasoning into a shared autoregressive space. Based on this unified representation, L2T introduces a modality-aware routing reward to select appropriate reasoning modes and a difficulty-aware fast-slow strategy to adapt reasoning depth according to task complexity. Experiments on general perception and reasoning-intensive tasks demonstrate the effectiveness of L2T. Compared with explicit tool-calling approaches, L2T maintains competitive performance while achieving 1.7--3.6 times higher throughput.
Quanlin Chen, Xiuwei Chen, Kun Xiang +5
Sun Yat-sen University · Shenzhen Loop Area Institute · Beijing Zhongguancun Academy +2