Paper deep dive
Balancing Efficiency and Efficacy: Training-Free Attention-Guided Switching Between Explicit and Latent Thoughts for MLLMs
Haoqian Kang, Liupeng Li, Kuofeng Gao, Jinpeng Wang, Zhenyu Lu, Bin Chen, Ke Chen, Yaowei Wang
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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 this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2608.03450v1
- Canonical: https://arxiv.org/abs/2608.03450v1
Trouble viewing inline? Open PDF directly →
Full Text
56,526 characters extracted from source content.
Expand or collapse full text
Balancing Efficiency and Efficacy: Training-Free Attention-Guided Switching Between Explicit and Latent Thoughts for MLLMs Haoqian Kang ∗ Harbin Institute of Technology, Shenzhen Shenzhen, China 26b951036@stu.hit.edu.cn Liupeng Li ∗ Harbin Institute of Technology, Shenzhen Shenzhen, China 25b951045@stu.hit.edu.cn Kuofeng Gao Tsinghua Shenzhen International Graduate School, Tsinghua University Shenzhen, China gkf24@mails.tsinghua.edu.cn Jinpeng Wang † Harbin Institute of Technology, Shenzhen Shenzhen, China wangjp26@gmail.com Zhenyu Lu Peng Cheng Laboratory Shenzhen, China Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences Shenzhen, China zhenyulu22@m.fudan.edu.cn Bin Chen Harbin Institute of Technology, Shenzhen Shenzhen, China chenbin2021@hit.edu.cn Ke Chen Peng Cheng Laboratory Shenzhen, China chenk02@pcl.ac.cn Yaowei Wang † Harbin Institute of Technology, Shenzhen Shenzhen, China Peng Cheng Laboratory Shenzhen, China yaoweiwang@gmail.com Abstract 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 ex- pensive 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 in- ference strategy for MLLMs that explicitly decouples perception and reasoning. We propose a novel metric, the vision-to-text at- tention 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 ∗ Equal Contribution. † Jinpeng Wang and Yaowei Wang are corresponding authors. This work is licensed under a Creative Commons Attribution 4.0 International License. M ’26, Rio de Janeiro, Brazil. © 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2213-4/2026/11 https://doi.org/10.1145/3767308.3835035 tokens to preserve high-fidelity visual information in the contin- uous 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/M26-AGS. CCS Concepts • Computing methodologies→Computer vision; Natural language processing; Knowledge representation and reasoning. Keywords multimodal large language models; latent reasoning; decoupling perception and reasoning ACM Reference Format: Haoqian Kang, Liupeng Li, Kuofeng Gao, Jinpeng Wang, Zhenyu Lu, Bin Chen, Ke Chen, and Yaowei Wang. 2026. Balancing Efficiency and Efficacy: Training-Free Attention-Guided Switching Between Explicit and Latent Thoughts for MLLMs. In Proceedings of the 34th ACM International Confer- ence on Multimedia (M ’26), November 10–14, 2026, Rio de Janeiro, Brazil. ACM, New York, NY, USA, 10 pages. https://doi.org/10.1145/3767308.3835035 arXiv:2608.03450v1 [cs.M] 4 Aug 2026 M ’26, November 10–14, 2026, Rio de Janeiro, Brazil.Haoqian Kang et al. Without CoT Latent Space Perception token Reasoning token What time of day is it most likely in the photograph? A: Morning B: Afternoon C: Evening D: Night M L L M The image has warm, golden lighting with long shadows and a dim sky, which are typical of sunset or early evening rather than midday or morning. Answer: C Answer: C Explicit CoT What time of day is it most likely in the photograph? A: Morning B: Afternoon C: Evening D: Night M L L M What time of day is it most likely in the photograph? A: Morning B: Afternoon C: Evening D: Night M L L M Answer: C Discrete Space The image has ... ... early evening. Our method Mode Switch Figure 1: Comparison of reasoning paradigms in MLLMs on a VQA example. Without CoT: directly predicts the answer. Explicit CoT: performs step-by-step reasoning in the dis- crete token space before producing the answer. Our proposed method (AGS): interleaves perception and reasoning tokens during inference, dynamically routing visual processing to the latent space to bypass redundant discrete text generation. 1 Introduction Reasoning is a fundamental capability of Multimodal Large Lan- guage Models (MLLMs) [1,18,35,40], enabling complex cross- modal tasks such as scientific diagram analysis and mathematical problem-solving. Unlike pure text-based logic [10,51], multimodal reasoning inherently requires the simultaneous mastery of two distinct cognitive processes: perceiving fine-grained visual details and performing rigorous logical deduction. The dominant reasoning paradigm in current MLLMs is explicit text-based chain-of-thought (CoT) [5,14,49], which relies on the generation of discrete tokens during inference, as illustrated in Fig. 1. However, coercing complex visual concepts into discrete linguistic tokens often incurs significant computational inefficiency and exacerbates visual hallucinations. To mitigate these issues, recent works [17,29,39,50,52] have explored latent reasoning, which interleaves text and image embeddings in a continuous space to emulate “thinking with images" [54,57]. Unfortunately, since such latent behaviors are not naturally aligned during standard pre-training, existing multimodal latent reasoning frameworks typ- ically demand substantial training overhead through high-quality external datasets or strong teacher supervision. To bypass the massive costs of retraining, a natural intuitive step is to adapt training-free latent reasoning frameworks originally designed for LLMs (e.g., SwiReasoning [34]) to the MLLM setting. These methods dynamically alternate between explicit text gen- eration and implicit continuous embeddings based on token-level probability. However, our preliminary empirical evaluations re- veal that such MLLM variants exhibit highly unstable performance, sometimes even underperforming vanilla explicit CoT. Guided by the principle of decoupling perception and reasoning [15,31], we diagnose a critical flaw in these prior arts: the reliance on token- level entropy as the sole metric for mode switching. In multimodal contexts, high output entropy conflates two fundamentally differ- ent challenges, perceptual ambiguity (e.g., struggling to identify a tiny visual object) and reasoning uncertainty (e.g., deliberating over V i s i o n - t o - T e x t A t t e n t i o n R a t i o ... OB is perpendicular ... ... 36° angle ... ... OB and OC ... ... triangle OBC ... Generation process Based on the geometric property that a tangent is perpendicular to the radius at the point of tangency, radius OB is perpendicular to tangent AB, which means in the right-angled triangle ABO, angle AOB can be calculated as 54° by subtracting the known 36° angle and the 90° right angle from the triangle's total 180°. Next, observing the inner triangle OBC, since OB and OC are both radii of the same circle, they are equal in length, making it an isosceles triangle where the base angles OBC and C are equal. Finally, applying the Exterior Angle Theorem to triangle OBC, the exterior angle AOB (54°) is equal to the sum of the two non-adjacent interior angles (angle OBC and angle C), which means 54° equals twice the measure of angle C, ultimately yielding 27° for angle C and leading to the correct choice, D. Figure 2: Visualization of the proposed vision-to-text atten- tion ratio (푅 퐴 ) during generation on a geometric reasoning example. The bar chart demonstrates how attention ratio dynamically fluctuates throughout the inference phase. No- tably, the distinct spikes in the ratio (orange bars) align pre- cisely with the generation of perception-heavy visual con- cepts from the image, such as identifying spatial relation- ships and specific angles (highlighted in red in the text). a complex mathematical step). Treating these distinct cognitive bottlenecks uniformly leads to suboptimal routing decisions. To address this entanglement, we argue that mode switching must dynamically adapt to whether the model is in a perception- dominant or logic-dominant phase during generation. Moving be- yond the opaque nature of output probabilities, we delve into the model’s internal representations and propose a novel, highly inter- pretable metric: the vision-to-text attention ratio. Concretely, the vision-to-text attention ratio is defined as the ratio of the attention allocated to visual tokens versus text tokens. This metric directly reflects the functional role of the current decoding step, accurately distinguishing perceptual tokens from reasoning tokens. As visu- alized in Fig. 2, tokens exhibiting higher vision-to-text attention ratios consistently align with perception-heavy concepts, empiri- cally validating our hypothesis that this metric captures the true cognitive state of the model. Empowered by the proposed metric, we introduce Attention- Guided Switching (AGS), a novel training-free inference strategy that performs asymmetric token routing. Since visual information is naturally dense and continuous, coercing it into textual tokens during perception-heavy phases inevitably causes information loss. Therefore, for perceptual tokens (where the model heavily attends to the image), AGS adaptively triggers implicit latent reasoning. Specifically, it forms a probability-weighted mixture of token em- beddings as inputs, enabling plug-and-play operation at inference time. This bypasses the linguistic bottleneck, allowing the model to process high-fidelity, continuous visual representations with- out forcing premature textualization that causes hallucinations. Conversely, since logical deduction relies on strict sequential de- pendencies, for reasoning tokens, the framework enables explicit text generation to provide strong logical anchoring and structural rigor, thereby minimizing deduction errors. Extensive experiments demonstrate the superiority of AGS. Com- pared to explicit text-based CoT and entropy-based SwiR-MLLM Training-Free Attention-Guided Switching Between Explicit and Latent Thoughts for MLLMsMM ’26, November 10–14, 2026, Rio de Janeiro, Brazil. baselines, our method achieves state-of-the-art performance while remarkably accelerating the inference process. Our contributions are summarized as follows: •We identify entropy entanglement in multimodal reasoning, where token-level entropy conflates perceptual ambiguity with logical uncertainty, resulting in unstable mode switching. • We propose the vision-to-text attention ratio, an interpretable token-level metric to distinguish perception-dominant and logic- dominant decoding steps. •Based on this metric, we develop AGS, a training-free strat- egy that adaptively routes perceptual tokens to the continuous latent space and logical tokens to the explicit text space. •Extensive evaluations of challenging multimodal benchmarks validate that our method achieves superior accuracy while sig- nificantly reducing autoregressive decoding steps and overall inference latency. 2 Related Works 2.1 MLLM Reasoning Recent advances in Multimodal Large Language Models (MLLMs) [2, 6,16,26–28,37,48] have extended visual question answering to- ward complex multimodal reasoning. Early methods [18,21] mainly follow the “thinking about images” paradigm, where visual inputs are encoded once and statically consumed by the LLM. This static design may weaken visual grounding over long reasoning chains and increase hallucinations. Recent works [7, 33, 56] instead pursue “thinking with images,” treating visual inputs as an interactive reasoning workspace. Rep- resentative approaches include tool-augmented visual search [4,19, 44], programmatic visual manipulation [12], interpretable reasoning– perception alignment [25], and visual imagination [52]. These ap- proaches improve fidelity but often require extra tools, computation, or training. Broader efficiency studies compress visual tokens [58] or replace attention in vision backbones [32], yet do not optimize MLLM reasoning trajectories. Long outputs further increase MLLM inference cost [8,41]. In contrast, our training-free method dy- namically switches between explicit and latent reasoning, reducing generation steps without modifying model parameters. 2.2 Decoupling Perception and Reasoning Coupling visual perception with symbolic reasoning can undermine reasoning stability in complex multimodal tasks. Early methods adopt macro-level decoupling through isolated modules, such as LLM–MLLM collaboration [15] or two-stage pipelines [9,31]. Al- though these designs can reduce hallucinations, they incur addi- tional latency and cannot fully capture the interleaved nature of multimodal cognition. Recent approaches [13,20,22] instead perform token-level decou- pling within a single autoregressive trajectory, distinguishing per- ceptual and logical tokens through signals such as visual sensitivity, distributional shifts, or hidden-state similarity. However, they still require additional supervision, curated data, or training overhead. Our method instead uses a training-free white-box metric based on cross-modal attention ratios to dynamically distinguish and switch between perception-dominant and logic-dominant states during inference. 2.3 Latent Reasoning Explicit Chain-of-Thought (CoT) [42] improves reasoning but in- curs substantial latency by autoregressively generating interme- diate text. Latent reasoning [11,36,55] reduces this cost by per- forming intermediate computation in continuous representation spaces. Originating in LLMs [45,46,55], existing methods typi- cally propagate latent reasoning through previous hidden states, probability-weighted vocabulary embeddings, or dedicated latent tokens. Latent reasoning has recently been extended to MLLMs [17,29, 39,50,52] to support “thinking with images.” However, existing multimodal methods generally rely on resource-intensive retraining and fine-grained supervision to align latent trajectories. Directly applying training-free LLM routing methods, such as entropy-based SwiReasoning [34], is also suboptimal because output entropy con- flates perceptual ambiguity with logical uncertainty. Our frame- work addresses this gap with an attention-guided, asymmetric, and training-free latent reasoning strategy for MLLMs. 3 Method Overview. In this section, we present the AGS framework. We first provide preliminaries on standard and latent MLLM reasoning formulations in Section 3.1. Next, in Section 3.2, we introduce the vision-to-text attention ratio to quantitatively decouple perceptual and logical tokens. Finally, Section 3.3 details the token-level dy- namic routing mechanism that seamlessly switches between latent and explicit reasoning modes. 3.1 Preliminary: MLLM Reasoning Vision and Language Inputs. Let I and푄denote the image and question fed into a Multimodal Large Language Model (MLLM). The image I is first processed by a vision encoder to extract visual features, which are subsequently mapped into the LLM embedding spaceR 푑 via a cross-modal projector. This yields a sequence of푁 visual embeddings, denoted as E 푣 = [ e 푣,1 , e 푣,2 , . . ., e 푣,푁 ] ∈ R 푁×푑 , where each e 푣,푖 ∈ R 푑 . Concurrently, the textual input푄is tok- enized and embedded into a sequence of푀textual embeddings E 푡 =[e 푡,1 ,e 푡,2 , . . .,e 푡,푀 ] ∈ R 푀×푑 . Finally, the visual and textual em- beddings are concatenated along the sequence dimension to form a unified joint representation E=[E 푣 ; E 푡 ]=[ e 1 , e 2 , . . ., e 퐿 ] ∈ R 퐿×푑 , where퐿= 푁 + 푀. This consolidated sequence E serves as the foundational input for subsequent reasoning within the pre-trained LLM. Autoregressive Forward Propagation. The reasoning backbone of the MLLM is a pre-trained LLM, parameterized by휃and denoted asM 휃 . In standard autoregressive generation, the previously gen- erated sequence is maintained entirely as a sequence of continuous embeddings, denoted as E <푡 =[e 1 ,e 2 , . . .,e 푡−1 ] ∈ R (푡−1)×푑 . At each generation step 푡 , conditioned on the initial multimodal context E and the generated sequence E <푡 , the model predicts the probability distribution of the t-th token: p 푡 =M 휃 E, E <푡 ∈Δ |V|−1 ,(1) whereV is the vocabulary of the LLM, andΔ |V|−1 represents the corresponding probability simplex. Based on this distribution, a hard discrete token푦 푡 ∈Vis determined via a decoding strategy M ’26, November 10–14, 2026, Rio de Janeiro, Brazil.Haoqian Kang et al. Output Logits Output Token . . . Weighted Embedding Selected Embedding Layer 1 Explicit to Latent Decoding As shown in the figure, AB is tangent to circle O at point B, and the extended line of AO intersects circle O at point C. Connect BC, if angle A = 36.0, then angle C is equal to () A: 36° B: 54° C: 60° D: 27° Layer 2 Layer N · · · · · · Visual TokensQuery Tokens . . . . . . . . . · · . . . . . . · · Reasoning Process Internal Attention Map Current Step 퐀 VisualQueryGenerated Generation Step 퐀 퐀 trigger Generation Step 퐀 퐀 trigger Reasoning Token Perception Token Latent to Explicit 퐀 퐀, 퐀 퐀 퐀, 퐀 퐀 퐀, 퐀 = 퐀 퐀, 퐀 퐀 퐀, 퐀 퐀 퐀, 퐀 Visual Attention: Text Attention: 퐀 퐀, 퐀 Attention Ratio: Figure 3: Illustration of AGS. The overall inference process (left) shows the model dynamically shifting its reasoning mode based on the calculated vision-to-text attention ratio,푅 퐴,푡 . High푅 퐴,푡 value triggers a switch from explicit logic to latent perception, while low푅 퐴,푡 value causes a transition from latent perception back to explicit logic, as visualized in the bottom-right plots. The specific calculation for the 푅 퐴,푡 is detailed in the top-right panel. D (e.g., greedy search or nucleus sampling): 푦 푡 =D(p 푡 ).(2) Subsequently, this discrete token푦 푡 is mapped back into the contin- uous representation space via the model’s vocabulary embedding functionEmb:V → R 푑 . The continuous embedding for the cur- rent step is obtained as: e 푡 = Emb(푦 푡 ).(3) Finally, this selected embedding e 푡 is appended to the continuous sequence, forming E <푡+1 =[ E <푡 ; e 푡 ] , which acts as the input state for the next autoregressive step 푡 + 1. Latent Reasoning Forward. Unlike standard autoregressive gen- eration that relies on discrete token decoding, latent reasoning propagates information entirely within the continuous embedding space. Instead of sampling a hard discrete token푦 푡 and executing a discrete look-up (e 푡 = Emb(푦 푡 )), this approach utilizes the pre- dicted probability distribution p 푡 to directly compute an expected, “soft” token embedding. Let W∈ R |V|×푑 denote the vocabulary weight matrix corresponding to the embedding functionEmb(·). The continuous latent representation ̃ e 푡 ∈ R 푑 for the current step is computed as the probability-weighted sum of all vocabulary embeddings: ̃ e 푡 = W ⊤ p 푡 .(4) This continuous embedding ̃ e 푡 is then appended to the sequence in place of e 푡 , forming the updated continuous context ̃ E <푡+1 = [E <푡 ; ̃ e 푡 ]for the subsequent reasoning step푡 +1. This fully contin- uous formulation effectively bypasses the information bottleneck of discrete token selection, allowing the aggregated embedding to carry and propagate richer distributional semantics within a single forward pass. 3.2 Decoupling Perception and Reasoning As discussed earlier, tokens play distinct functional roles during inference. To quantitatively capture this, we propose a novel metric, the attention ratio푅 퐴 , to characterize the specific role of each token at any given generation step. Specifically, let푁 푙 and푁 ℎ denote the total number of attention layers and attention heads within the LLM, respectively. LetI 푣 and I 푡 represent the index sets of the visual and textual tokens in the current context. As the autoregressive generation proceeds, the size of the visual index set|I 푣 |remains constant, whereas the textual index setI 푡 dynamically expands. Let훼 (푙,ℎ) 푡,푗 denote the attention weight assigned by the current query token at step푡to the key token 푗 , computed by the ℎ-th attention head in the 푙 -th layer. To mitigate the bias introduced by the inherent length disparity between visual and textual contexts, we first calculate the average attention allocated at the current generation step푡to visual tokens, denoted as퐴 푉,푡 , and textual tokens, denoted as퐴 푇,푡 . These are averaged across all layers and heads: 퐴 푉,푡 = 1 |I 푣 |푁 푙 푁 ℎ 푁 푙 ∑︁ 푙=1 푁 ℎ ∑︁ ℎ=1 ∑︁ 푗∈I 푣 훼 (푙,ℎ) 푡,푗 , 퐴 푇,푡 = 1 |I 푡 |푁 푙 푁 ℎ 푁 푙 ∑︁ 푙=1 푁 ℎ ∑︁ ℎ=1 ∑︁ 푘∈I 푡 훼 (푙,ℎ) 푡,푘 . (5) Based on these normalized attention values, we define the overall attention ratio at step 푡 as: 푅 퐴,푡 = 퐴 푉,푡 퐴 푇,푡 .(6) As illustrated in Figure 3 (top-right),푅 퐴,푡 quantifies the dynamic information-seeking preference of the model during generation. A high푅 퐴,푡 indicates a dominant attention allocation toward the Training-Free Attention-Guided Switching Between Explicit and Latent Thoughts for MLLMsMM ’26, November 10–14, 2026, Rio de Janeiro, Brazil. visual tokens, signifying that the model is actively extracting vi- sual information. We define such tokens as serving a perception role. Conversely, a low푅 퐴,푡 reflects a structural shift in reliance toward the textual context for semantic deduction and logical pro- gression. Accordingly, we characterize these tokens as functioning in a reasoning role. 3.3 Reasoning Mode Switching Building upon the token role decoupling, AGS dynamically routes each generation step to either the latent or explicit reasoning mode. Since the absolute magnitude of visual attention can vary across samples and reasoning phases due to differences in image com- plexity and multimodal context, a single global threshold may not generalize well. We therefore adopt a phase-wise dynamic thresh- old, using a phase-specific reference to capture relative attention changes within each reasoning phase. Specifically, let푡 푘 denote the step at which the most recent mode switch occurred. The threshold for the current reasoning phase, denoted as휏 푘 , is dynamically anchored to the attention ratio of the very first token generated in this phase, i.e.,휏 푘 = 푅 퐴,푡 푘 . Then, when 푅 퐴,푡 ≥ 휏 푘 , the model focuses on visual perception and operates in the latent reasoning mode to propagate information smoothly within the continuous embedding space. Conversely, when푅 퐴,푡 < 휏 푘 , it indicates a structural shift towards textual reliance, prompting the model to switch to the explicit reasoning mode to produce discrete semantic tokens. To ensure generation quality and logical coherence, we impose two critical constraints on the switching mechanism: Minimum Maintenance Window (푊): To prevent fragmented logical chains caused by overly frequent mode oscillation, we enforce a minimum duration for the explicit reasoning mode. If the model switches to explicit reasoning at step푡 exp , it must remain in this state for at least푊consecutive steps, temporarily overriding the dynamic푅 퐴,푡 condition. Maximum Switch Budget (퐶): Unconstrained reason- ing loops may lead to non-termination. We track the cumulative number of mode transitions, denoted as푐 푡 . Once푐 푡 reaches a prede- fined maximum budget퐶, we force the termination of the thinking process by appending the</think>token. Thereafter, the model exclusively adopts standard explicit reasoning until the generation concludes. Formally, let푆 푡 ∈ 0,1denote the reasoning state at step푡, where 0 indicates latent reasoning and 1 indicates explicit reason- ing. Let푐 푡−1 denote the cumulative number of mode transitions prior to step푡. Accounting for the dynamic threshold, the window constraint, and the maximum switch budget, the state indicator is determined by: 푆 푡 = I 푅 퐴,푡 < 휏 푘 ∨(푆 푡−1 = 1∧푡 −푡 exp < 푊)∨푐 푡−1 ≥ 퐶 , (7) whereI(·)is the indicator function. Following this state determina- tion, the transition tracker is updated as푐 푡 = 푐 푡−1 + I(푆 푡 ≠ 푆 푡−1 ). Let e ∗ 푡 denote the final selected embedding to be appended to the sequence for the subsequent step. To strictly enforce the stopping mechanism without causing an infinite loop, we inject the termina- tion identifier exactly once at the step the budget is reached. The complete forward routing is formulated as: e ∗ 푡 = e stop ,if 푐 푡 =퐶 and 푐 푡−1 < 퐶 ̃ e 푡 ,if 푐 푡 < 퐶 and 푆 푡 = 0 Emb( ˆ 푦 푡 ),otherwise ,(8) where e stop ∈ R 푑 is the predefined continuous embedding of the ter- mination identifier (e.g.,</think>), ̃ e 푡 is the probability-weighted soft embedding from Eq. 4, andEmb( ˆ 푦 푡 )is the standard discrete embedding anchored by the greedy decoding token ˆ 푦 푡 = arg maxp 푡 . This routing design empowers the MLLM to autonomously allocate computational bandwidth between continuous visual perception and discrete logical deduction, while mathematically guaranteeing deterministic termination. 4 Experiments 4.1 Experimental Setup Models. We evaluate AGS on two state-of-the-art MLLM families: Qwen3-VL-Thinking [1] and InternVL3.5 [40], each covering 2B, 4B, and 8B variants. Qwen3-VL-Thinking typically produces long deliberative chains, whereas InternVL3.5 follows a substantially more concise reasoning pattern. Their distinct architectures, rea- soning behaviors, and parameter scales allow us to assess whether AGS generalizes across both computation-intensive and compact inference regimes, rather than relying on a particular model family or chain-length distribution. Evaluation Benchmarks. We evaluate our method on six challeng- ing benchmarks spanning three multimodal reasoning domains: •Mathematical Reasoning: MathVista [23] evaluates multi- skill visual mathematics, MathVision [38] contains competition- level problems, MathVerse [53] reduces textual bias to assess visual reasoning, and WeMath [30] focuses on multi-step math- ematical reasoning. •STEM Reasoning: ScienceQA [24] evaluates multimodal scien- tific reasoning with background knowledge. •General Multimodal Reasoning: M 3 CoT [3] evaluates gen- eral multimodal multi-step reasoning across diverse scenarios. For the additional baseline comparison, we also include V ∗ [43], which evaluates fine-grained visual perception. Baselines. We primarily compare AGS with Explicit CoT, which generates all intermediate reasoning steps as discrete text before producing the final answer. To evaluate the proposed routing met- ric, we additionally compare it with the entropy-based strategy used in SwiReasoning [34] in the ablation study. For multimodal latent reasoning, we compare with LEAD [47]. Both LEAD and AGS use the same Qwen3-VL-4B-Thinking checkpoint, enabling a checkpoint-controlled comparison that more directly isolates the effect of the latent reasoning strategy. Metrics. We report Task Accuracy for answer correctness and Inference Efficiency as the average number of autoregressive gen- eration steps. Unlike visible output-token counts, generation steps also account for latent reasoning steps, which perform forward propagation in the continuous embedding space without emitting M ’26, November 10–14, 2026, Rio de Janeiro, Brazil.Haoqian Kang et al. Table 1: Quantitative evaluation across diverse multimodal reasoning tasks. We report both Accuracy (Acc, %) and Efficiency (Eff), measured by the average number of autoregressive generation steps. Best results are highlighted in bold. Model & Method Math ReasoningSTEM and General Avg MathVerseMathVisionMathVistaWeMathM3CoTScienceQA Acc↑Eff↓Acc↑Eff↓Acc↑Eff↓Acc↑Eff↓Acc↑Eff↓Acc↑Eff↓ Acc↑ Eff↓ InternVL3.5 Series InternVL3.5-2B35.63659.449423.225351.637250.628877.119341.3328 InternVL3.5-2B (+ AGS)36.936210.349023.023653.234852.827181.817643.0314 InternVL3.5-4B47.834115.948636.026065.828664.826888.415853.1300 InternVL3.5-4B (+ AGS)50.033716.947235.725765.927466.027287.914453.7293 InternVL3.5-8B48.331518.867139.025170.834167.020790.313355.7320 InternVL3.5-8B (+ AGS)48.830218.243438.521970.926368.020990.512155.8258 Qwen3-VL Series Qwen3-VL-2B-Thinking34.132145.8394127.2215632.4336448.6243469.3141336.22754 Qwen3-VL-2B-Thinking (+ AGS)42.614819.6197037.1111248.6151254.2122573.880044.31350 Qwen3-VL-4B-Thinking44.1313611.1386833.3209243.1321757.5225581.3124245.12635 Qwen3-VL-4B-Thinking (+ AGS)60.0114417.7160239.691569.4111465.898588.264356.81067 Qwen3-VL-8B-Thinking55.9257410.7369330.5174662.9248962.7188187.589851.72214 Qwen3-VL-8B-Thinking (+ AGS)64.7100218.9159935.386176.886767.486189.852958.8953 Table 2: Evaluation on the POPE Benchmark. InternVL3.5 Base+ AGS Qwen3-VL Base+ AGS Acc↑ Eff↓ Acc↑ Eff↓Acc↑ Eff↓ Acc↑ Eff↓ 2B83.67884.5792B-Think86.811287.2104 4B81.67482.2714B-Think87.013087.1127 8B82.87983.9768B-Think86.79386.888 discrete tokens. This metric therefore provides a fairer estimate of the actual inference cost. Implementation Details. We implement AGS in PyTorch and eval- uate it on NVIDIA RTX A6000 GPUs, following the official Hug- gingFace inference pipelines for Qwen3-VL and InternVL3.5. The maximum sequence length is 4096. We set the minimum explicit- reasoning maintenance window to푊=512 and the maximum switch budget to퐶=4 for all experiments. Explicit tokens are generated using multinomial sampling withdo_sample=Trueand a temperature of 0.6. 4.2 Main Result Table 1 reports results across two MLLM families and six reasoning benchmarks. Overall, AGS improves both accuracy and inference efficiency, alleviating the accuracy–efficiency trade-off of Explicit CoT. Table 2 further shows consistent POPE improvements, indi- cating reduced object hallucinations without sacrificing efficiency. The gains are particularly pronounced for Qwen3-VL-Thinking, which typically produces long reasoning chains. On the 8B model, AGS reduces the average generation cost from 2214 to 953 steps (∼57%) while improving accuracy from 51.7% to 58.8%. The 2B and 4B variants similarly reduce generation steps by approximately 50% and 60%, with absolute accuracy gains of 8.1 and 11.7 percentage points, respectively. The improvements are especially strong on visual-mathematical tasks. On WeMath, AGS raises Qwen3-VL-4B-Thinking accuracy from 43.1% to 69.4%; on MathVerse, the 4B and 8B models gain 15.9 and 8.8 percentage points. These results suggest that latent routing avoids redundant textualization while preserving visual information. AGS also generalizes to the more concise InternVL3.5 family, whose baselines average about 300 generation steps. For InternVL3.5- 8B, it reduces the cost from 320 to 258 steps while maintaining com- parable average accuracy (55.7% versus 55.8%) and achieving 90.5% on ScienceQA. Results across two model families and scales from 2B to 8B demonstrate generalizability without additional training or model-specific tuning. On POPE, AGS improves accuracy for all six model variants by 0.1–1.1 percentage points and reduces generation steps in five settings. InternVL3.5-8B improves from 82.8% to 83.9% while re- ducing steps from 79 to 76, and Qwen3-VL-2B-Thinking improves from 86.8% to 87.2% while reducing steps from 112 to 104. The only exception is InternVL3.5-2B, whose cost increases marginally from 78 to 79 steps while accuracy rises from 83.6% to 84.5%. Overall, the efficiency gains do not systematically degrade visual grounding and modestly improve object-hallucination robustness. 4.3 Comparison with LEAD Table 5 compares AGS with LEAD using the same Qwen3-VL-4B- Thinking checkpoint. AGS requires fewer generation steps on all four benchmarks and achieves higher accuracy on V ∗ , ScienceQA, and M 3 CoT, while LEAD performs better on WeMath. Specifically, AGS reduces the cost from 413 to 346 steps on V ∗ , from 1706 to 1114 on WeMath, from 871 to 643 on ScienceQA, and from 1511 to 985 on M 3 CoT. These gains span fine-grained perception, scientific reason- ing, and general multimodal multi-step reasoning, demonstrating Training-Free Attention-Guided Switching Between Explicit and Latent Thoughts for MLLMsMM ’26, November 10–14, 2026, Rio de Janeiro, Brazil. Table 3: Ablation study on the reasoning mode transition metric using the Qwen3-VL-8B-Thinking model on MathVerse, WeMath, M3CoT, and ScienceQA. Method / Routing Metric MathVerseWeMathM3CoTScienceQAAvg Acc↑Eff↓Acc↑Eff↓Acc↑Eff↓Acc↑Eff↓ Acc↑ Eff ↓ Explicit CoT (No Routing)55.9257462.9248962.7188187.589867.31961 Token Entropy58.8106868.9100166.485088.853570.7864 Attention Ratio (Ours)64.7100276.886767.486189.852974.7815 Table 4: Ablation study on the explicit reasoning mainte- nance window size (푊) on the M 3 CoT benchmark. Our de- fault setting (푊= 512) is highlighted. Window Size (푊 ) M3CoT Acc↑Eff↓ 6452.2209 12858.9412 25663.6609 51267.4861 102462.91766 Table 5: Comparison results with LEAD and our AGS. Method V ∗ WeMath ScienceQAM 3 CoT Acc↑ Eff↓ Acc↑ Eff↓ Acc↑ Eff↓ Acc↑ Eff↓ Explicit CoT 82.746243.1321781.3124257.52255 LEAD80.6413 72.5170673.687164.51511 AGS (Ours)81.134669.4111488.264365.8985 a more favorable overall accuracy–efficiency trade-off under the same checkpoint. 4.4 Ablation Study Metric for Reasoning Mode Switching. We compare the proposed vision-to-text attention ratio with Explicit CoT and token-level en- tropy routing. As shown in Table 3, entropy routing reduces the average cost from 1961 to 864 steps but provides limited accuracy gains, consistent with the uncertainty conflation issue: entropy can- not distinguish perceptual ambiguity from logical difficulty and may therefore select inappropriate reasoning modes. By directly tracking cross-modal attention flow, our metric better separates perception-dominant and logic-dominant phases. It improves aver- age accuracy over entropy routing by 4.0 percentage points (70.7% to 74.7%) while further reducing the cost to 815 steps. The gains reach 7.9 and 5.9 points on WeMath and MathVerse, demonstrating a more effective and interpretable routing signal. Maintenance Window Size. We evaluate the sensitivity to the minimum maintenance window using Qwen3-VL-8B-Thinking on M 3 CoT, with푊ranging from 64 to 1024 generation steps. As shown in Table 4, performance follows a clear inverted U-shaped trend. Small windows such as푊=64 or 128 provide high efficiency but substantially reduce accuracy, reaching only 52.2% at푊=64. Such settings return the model to latent perception before a co- herent explicit deduction is completed, resulting in fragmented 23456 Maximum Switch Budget (C) 0 200 400 600 800 1000 1200 1400 Generation Steps Efficiency (Default C = 4) Efficiency (Others) Accuracy 58 60 62 64 66 68 Accuracy (%) Figure 4: Ablation study on the maximum mode transition budget (퐶) using the Qwen3-VL-8B-Thinking model on the M 3 CoT. The bar chart (left axis) reports inference efficiency in generation steps, while the line plot (right axis) represents reasoning accuracy. reasoning chains. Conversely,푊=1024 keeps the model in explicit generation for too long, increasing the cost to 1766 steps while reducing accuracy to 62.9%. This suggests that overly extended tex- tual reasoning can drift away from continuous visual evidence and become dominated by verbose linguistic priors. We therefore select 푊=512, which provides the best balance between coherent logical anchoring and timely return to high-fidelity visual perception. Maximum Switch Budget. We further analyze the effect of the maximum switch budget퐶. As shown in Figure 4, accuracy follows an inverted U-shaped trend, while excessive switching increases computational cost. With퐶=2, inference requires fewer than 500 steps, but accuracy drops to approximately 58%, as insufficient perception–reasoning interactions prevent the model from resolv- ing complex multi-hop dependencies. Increasing the budget beyond 퐶=4 introduces redundant transitions without further gains. At 퐶=6, accuracy decreases to around 65%, while the generation cost exceeds 1400 steps, indicating computational bloat, overthinking, and logical drift. We therefore set퐶=4, which achieves the highest accuracy of approximately 67.5% with bounded generation cost. M ’26, November 10–14, 2026, Rio de Janeiro, Brazil.Haoqian Kang et al. <perception process> First, I should visualize the coordinate system. Let's assume point A is at the origin (0,0). Since he walks north to B, which is 5 meters north, so B would be at (0,5). Then from B, he walks east 5 meters to C. East is the positive x-direction, so C would be at (5,5). Wait, but wait, north is usually the positive y-direction, right? So moving north from A (0,0) to B (0,5), then east from B (0,5) to C (5,5). So point C is at (5,5) relative to A. Now, the question is asking for the direction of C from A. So from A, looking towards C. The coordinates of C are (5,5), so it's 5 meters east and 5 meters north from A. Wait, no. Wait, <perception process> the options given are like "North by east 45°" etc. Let me recall the standard notation. In navigation, directions are often expressed as "North by East" or "East by North", but the standard is usually "North [direction] angle". Wait, maybe it's similar to bearings. </think> Mike starts at point A. Walking 5 meters north to point B places B directly north of A. Then walking 5 meters east from B to point C places C 5 meters east and 5 meters north of A. This forms a right triangle with legs of 5 meters each. The direction from A to C is northeast, specifically 45° east of north. The standard directional notation for this is "North by east 45°", which matches option C. Answer: C Generation Steps: 627 As shown in the figure, Mike's initial position is at point A. He walks 5 meters north to reach point B, then walks 5 meters east to reach point C. In which direction is point C from point A? A. North by west 45° B. South by east 45° C. North by east 45° D. South by west 45° E. No correct answer Okay, let's see. So Mike starts at point A. He walks 5 meters north to point B, then 5 meters east to point C. We need to figure out the direction of point C from point A. ...... Wait, but if AB is 5m north and BC is 5m east, then the coordinates would be: let's assign A as the origin (0,0). Then B is at (0,5) because north is positive y-axis. Then C is at (5,5) because moving east from B (which is at (0,5)) 5 meters would take us to (5,5). So from A (0,0) to C (5,5), the direction would be northeast, but specifically, what angle? ...... The displacement components are east 5m (x-axis) and north 5m (y-axis). The angle theta satisfies tan(theta) = east / north = 5/5 = 1. So theta = 45 degrees. ...... Based on the problem description and the given figure, Mike starts at point A, walks 5 meters north to point B, then 5 meters east to point C. This forms a right-angled triangle ABC with AB = 5 meters (north) and BC = 5 meters (east), making triangle ABC an isosceles right triangle with a 45° angle at point A. ...... Answer: C Generation Steps: 2349 As shown in the figure, Mike's initial position is at point A. He walks 5 meters north to reach point B, then walks 5 meters east to reach point C. In which direction is point C from point A? A. North by west 45° B. South by east 45° C. North by east 45° D. South by west 45° E. No correct answer Explicit CoT Hybrid CoT Figure 5: Qualitative comparison between standard Explicit CoT (left) and our proposed Hybrid CoT (right) on a spatial reasoning sample from the WeMath dataset. 050100150200 Generation Step t 1.0 0.5 0.0 0.5 1.0 Cosine Sim. S t Step-wise Latent Vector Cosine Similarity cos(e t , e t1 ) Latent Explicit Figure 6: Step-wise cosine similarity between consecutive in- put embeddings,cos(푒 푡 ,푒 푡−1 ), for latent and explicit reasoning steps in a representative WeMath example using Qwen3-VL- 4B-Thinking. Latent similarities remain below 1 and fluctu- ate like explicit steps, indicating stable propagation. 4.5 Case Study Figure 5 compares Explicit CoT with AGS on a WeMath spatial- reasoning example. Explicit CoT requires 2349 steps to repeatedly textualize visual coordinates and displacements, resulting in ver- bose reasoning and potential logical loops. Guided by the attention ratio, AGS routes visual perception through the continuous latent space (<perception process>) while retaining explicit generation for logical deduction. It reaches the correct answer (Option C) in only 627 steps, reducing generation cost by approximately 73%. This demonstrates that selective latent routing avoids redundant visual textualization without sacrificing reasoning accuracy. 4.6 Stability of Latent Trajectories We assess latent propagation using the step-wise cosine similarity between consecutive input embeddings,cos(푒 푡 ,푒 푡−1 ). As shown in Figure 6, latent similarities remain below 1 and fluctuate through- out generation, indicating evolving rather than collapsed repre- sentations. Their ranges and variation are also comparable to ex- plicit steps, suggesting stable representation dynamics during mode switching. 5 Conclusions In this paper, we identify a key limitation of entropy-based latent reasoning for MLLMs: perceptual ambiguity and reasoning uncer- tainty are entangled. To address this issue, we introduce the in- terpretable vision-to-text attention ratio to distinguish perception- dominant from logic-dominant decoding phases. Based on this metric, our training-free strategy routes perceptual tokens through the continuous latent space to preserve visual information, while retaining explicit generation for structured reasoning. Extensive ex- periments across diverse model families and multimodal reasoning benchmarks show that our method improves reasoning accuracy, reduces visual hallucinations, and substantially lowers autoregres- sive decoding steps and inference latency, without architectural modification or task-specific supervision. This provides an efficient alternative to costly model retraining. Acknowledgments We want to thank the anonymous reviewers and the meta-reviewer for their valuable comments and suggestions. This work is sup- ported in part by the National Science and Technology Major Project 2025ZD1601300, the National Natural Science Foundation of China under Grants 624B2088, 62536003 and 62301189, and by the project of Peng Cheng Laboratory (PCL2025A14). Training-Free Attention-Guided Switching Between Explicit and Latent Thoughts for MLLMsMM ’26, November 10–14, 2026, Rio de Janeiro, Brazil. References [1]Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al.2025. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025). [2]Hardy Chen, Haoqin Tu, Fali Wang, Hui Liu, Xianfeng Tang, Xinya Du, Yuyin Zhou, and Cihang Xie. 2025. Sft or rl? an early investigation into training r1-like reasoning large vision-language models. arXiv preprint arXiv:2504.11468 (2025). [3]Qiguang Chen, Libo Qin, Jin Zhang, Zhi Chen, Xiao Xu, and Wanxiang Che. 2024. M3cot: A novel benchmark for multi-domain multi-step multi-modal chain-of-thought. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 8199–8221. [4] Zihui Cheng, Qiguang Chen, Xiao Xu, Jiaqi Wang, Weiyun Wang, Hao Fei, Yidong Wang, Alex Jinpeng Wang, Zhi Chen, Wanxiang Che, et al.2025. Visual thoughts: A unified perspective of understanding multimodal chain-of-thought. arXiv preprint arXiv:2505.15510 (2025). [5]Yihe Deng, Hritik Bansal, Fan Yin, Nanyun Peng, Wei Wang, and Kai-Wei Chang. 2025. Openvlthinker: An early exploration to complex vision-language reasoning via iterative self-improvement. arXiv e-prints (2025), arXiv–2503. [6]Kaixuan Fan, Kaituo Feng, Haoming Lyu, Dongzhan Zhou, and Xiangyu Yue. 2025. Sophiavl-r1: Reinforcing mllms reasoning with thinking reward. arXiv preprint arXiv:2505.17018 (2025). [7] Jun Gao, Yongqi Li, Ziqiang Cao, and Wenjie Li. 2025. Interleaved-modal chain-of- thought. In Proceedings of the Computer Vision and Pattern Recognition Conference. 19520–19529. [8] Kuofeng Gao, Yang Bai, Jindong Gu, Shu-Tao Xia, Philip Torr, Zhifeng Li, and Wei Liu. 2024. Inducing high energy-latency of large vision-language models with verbose images. In ICLR. [9]Yunhao Gou, Kai Chen, Zhili Liu, Lanqing Hong, Xin Jin, Zhenguo Li, James T Kwok, and Yu Zhang. 2025. Reasoning-Aligned Perception Decoupling for Scal- able Multi-modal Reasoning. arXiv preprint arXiv:2506.04559 (2025). [10]Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. 2025. Deepseek-r1: Incen- tivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 (2025). [11]Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. 2024. Training large language models to reason in a continuous latent space. arXiv preprint arXiv:2412.06769 (2024). [12]Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. 2024. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. Advances in Neural Information Processing Systems 37 (2024), 139348–139379. [13]Siyuan Huang, Xiaoye Qu, Yafu Li, Yun Luo, Zefeng He, Daizong Liu, and Yu Cheng. 2025. Spotlight on token perception for multimodal reinforcement learn- ing. arXiv preprint arXiv:2510.09285 (2025). [14] Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Zhe Xu, Xu Tang, Yao Hu, and Shaohui Lin. 2025. Vision-r1: Incentivizing reasoning capability in multimodal large language models. arXiv preprint arXiv:2503.06749 (2025). [15]Hongrui Jia, Chaoya Jiang, Shikun Zhang, and Wei Ye. 2025. Decoupling Rea- soning and Perception: An LLM-LMM Framework for Faithful Visual Reasoning. arXiv preprint arXiv:2509.23322 (2025). [16] Xin Lai, Junyi Li, Wei Li, Tao Liu, Tianjian Li, and Hengshuang Zhao. 2025. Mini- o3: Scaling up reasoning patterns and interaction turns for visual search. arXiv preprint arXiv:2509.07969 (2025). [17]Bangzheng Li, Ximeng Sun, Jiang Liu, Ze Wang, Jialian Wu, Xiaodong Yu, Hao Chen, Emad Barsoum, Muhao Chen, and Zicheng Liu. 2025. Latent visual rea- soning. arXiv preprint arXiv:2509.24251 (2025). [18]Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al.2024. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326 (2024). [19]Liupeng Li, Haoqian Kang, Zhenyu Lu, Jinpeng Wang, Bin Chen, Ke Chen, and Yaowei Wang. 2026. CVSearch: Empowering Multimodal LLMs with Cognitive Vi- sual Search for High-Resolution Image Perception. arXiv preprint arXiv:2605.23655 (2026). [20]Yunheng Li, Hangyi Kuang, Hengrui Zhang, Jiangxia Cao, Zhaojie Liu, Qibin Hou, and Ming-Ming Cheng. 2026. Rethinking Token-Level Policy Optimization for Multimodal Chain-of-Thought. arXiv preprint arXiv:2603.22847 (2026). [21]Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual in- struction tuning. Advances in neural information processing systems 36 (2023), 34892–34916. [22]Jinda Lu, Junkang Wu, Jinghan Li, Kexin Huang, Shuo Yang, Guoyin Wang, Jiancan Wu, Xiang Wang, and Xiangnan He. 2026. Bridging Perception and Reasoning: Token Reweighting for RLVR in Multimodal LLMs. arXiv preprint arXiv:2603.25077 (2026). [23] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. 2023. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255 (2023). [24]Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Taf jord, Peter Clark, and Ashwin Kalyan. 2022. Learn to explain: Multi- modal reasoning via thought chains for science question answering. Advances in neural information processing systems 35 (2022), 2507–2521. [25]Zhenyu Lu, Liupeng Li, Jinpeng Wang, Haoqian Kang, Yan Feng, Ke Chen, and Yaowei Wang. 2026. SegCompass: Exploring Interpretable Alignment with Sparse Autoencoders for Enhanced Reasoning Segmentation. arXiv preprint arXiv:2605.22658 (2026). [26] Zhenyu Lu, Liupeng Li, Jinpeng Wang, Haoqian Kang, Manyuan Zhang, Yan Feng, Ke Chen, and Yaowei Wang. 2026. AffIn-Space: Learning Affine-Invariant Repre- sentations for 3D Spatial Understanding with MLLMs. In Forty-third International Conference on Machine Learning. https://openreview.net/forum?id=ZRymA5ujIZ [27]Ruilin Luo, Zhuofan Zheng, Yifan Wang, Yiyao Yu, Xinzhe Ni, Zicheng Lin, Jin Zeng, and Yujiu Yang. 2025. Ursa: Understanding and verifying chain-of-thought reasoning in multimodal mathematics. arXiv e-prints (2025), arXiv–2501. [28]Fanqing Meng, Lingxiao Du, Zongkai Liu, Zhixiang Zhou, Quanfeng Lu, Daocheng Fu, Botian Shi, Wenhai Wang, Junjun He, Kaipeng Zhang, et al.2025. Mm-eureka: Exploring visual aha moment with rule-based large-scale reinforce- ment learning. arXiv preprint arXiv:2503.07365 (2025). [29]Tan-Hanh Pham and Chris Ngo. 2025. Multimodal chain of continuous thought for latent-space reasoning in vision-language models. arXiv preprint arXiv:2508.12587 (2025). [30]Runqi Qiao, Qiuna Tan, Guanting Dong, MinhuiWu MinhuiWu, Chong Sun, Xi- aoshuai Song, Jiapeng Wang, Zhuoma Gongque, Shanglin Lei, Yifan Zhang, et al. 2025. We-math: Does your large multimodal model achieve human-like mathe- matical reasoning?. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 20023–20070. [31] Yuxuan Qiao, Haodong Duan, Xinyu Fang, Junming Yang, Lin Chen, Songyang Zhang, Jiaqi Wang, Dahua Lin, and Kai Chen. 2024. Prism: A framework for decoupling and assessing the capabilities of vlms. Advances in Neural Information Processing Systems 37 (2024), 111863–111898. [32] Carmelo Scribano, Mohammad Mahdi, Nedyalko Prisadnikov, Yuqian Fu, Giorgia Franchini, Danda Pani Paudel, Marko Bertogna, and Luc Van Gool. 2026. Accel- erating Vision Foundation Models with Drop-in Depthwise Convolution. arXiv preprint arXiv:2605.22132 (2026). [33]Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. 2024. Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought rea- soning. Advances in Neural Information Processing Systems 37 (2024), 8612–8642. [34]Dachuan Shi, Abedelkadir Asi, Keying Li, Xiangchi Yuan, Leyan Pan, Wenke Lee, and Wen Xiao. 2025. SwiReasoning: Switch-Thinking in Latent and Explicit for Pareto-Superior Reasoning LLMs. arXiv preprint arXiv:2510.05069 (2025). [35]Zhaochen Su, Peng Xia, Hangyu Guo, Zhenhua Liu, Yan Ma, Xiaoye Qu, Jiaqi Liu, Yanshu Li, Kaide Zeng, Zhengyuan Yang, et al.2025. Thinking with images for multimodal reasoning: Foundations, methods, and future frontiers. arXiv preprint arXiv:2506.23918 (2025). [36]Wenhui Tan, Jiaze Li, Jianzhong Ju, Zhenbo Luo, Ruihua Song, and Jian Luan. 2025. Think silently, think fast: Dynamic latent compression of llm reasoning chains. arXiv preprint arXiv:2505.16552 (2025). [37]Jintao Tong, Shilin Yan, Hongwei Xue, Xiaojun Tang, Kunyu Shi, Guannan Zhang, Ruixuan Li, and Yixiong Zou. 2026. SwimBird: Eliciting Switchable Reasoning Mode in Hybrid Autoregressive MLLMs. arXiv preprint arXiv:2602.06040 (2026). [38]Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. 2024. Measuring multimodal mathematical reasoning with math-vision dataset. Advances in Neural Information Processing Systems 37 (2024), 95095–95169. [39]Qixun Wang, Yang Shi, Yifei Wang, Yuanxing Zhang, Pengfei Wan, Kun Gai, Xianghua Ying, and Yisen Wang. 2025. Monet: Reasoning in latent visual space beyond images and language. arXiv preprint arXiv:2511.21395 (2025). [40]Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, et al.2025. Internvl3. 5: Ad- vancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265 (2025). [41]Xiasi Wang, Tianliang Yao, Simin Chen, Runqi Wang, Lei Ye, Kuofeng Gao, Yi Huang, and Yuan Yao. 2025. VLMInferSlow: Evaluating the Efficiency Robustness of Large Vision-Language Models as a Service. In ACL. [42]Yaoting Wang, Shengqiong Wu, Yuecheng Zhang, Shuicheng Yan, Ziwei Liu, Jiebo Luo, and Hao Fei. 2025. Multimodal chain-of-thought reasoning: A comprehensive survey. arXiv preprint arXiv:2503.12605 (2025). [43] Penghao Wu and Saining Xie. 2024. V*: Guided visual search as a core mechanism in multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 13084–13094. [44] Qiong Wu, Xiangcong Yang, Yiyi Zhou, Chenxin Fang, Baiyang Song, Xiaoshuai Sun, and Rongrong Ji. 2025. Grounded chain-of-thought for multimodal large language models. arXiv preprint arXiv:2503.12799 (2025). [45]Yige Xu, Xu Guo, Zhiwei Zeng, and Chunyan Miao. 2025. Softcot: Soft chain- of-thought for efficient reasoning with llms. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). M ’26, November 10–14, 2026, Rio de Janeiro, Brazil.Haoqian Kang et al. 23336–23351. [46]Yige Xu, Xu Guo, Zhiwei Zeng, and Chunyan Miao. 2025. Softcot++: Test-time scaling with soft chain-of-thought reasoning. arXiv preprint arXiv:2505.11484 (2025). [47]Zhongxing Xu, Zhonghua Wang, Zhe Qian, Dachuan Shi, Feilong Tang, Ming Hu, Shiyan Su, Xiaocheng Zou, Wei Feng, Dwarikanath Mahapatra, et al.2026. Thinking in uncertainty: Mitigating hallucinations in mlrms with latent entropy- aware decoding. arXiv preprint arXiv:2603.13366 (2026). [48] Wenhao Yang, Yu Xia, Jinlong Huang, Shiyin Lu, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, Yuanyu Wan, and Lijun Zhang. 2025. Deep But Reliable: Advancing Multi-turn Reasoning for Thinking with Images. arXiv preprint arXiv:2512.17306 (2025). [49] Yi Yang, Xiaoxuan He, Hongkun Pan, Xiyan Jiang, Yan Deng, Xingtao Yang, Haoyu Lu, Dacheng Yin, Fengyun Rao, Minfeng Zhu, et al.2025. R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 2376– 2385. [50] Zeyuan Yang, Xueyang Yu, Delin Chen, Maohao Shen, and Chuang Gan. 2025. Machine mental imagery: Empower multimodal reasoning with latent visual tokens. arXiv preprint arXiv:2506.17218 (2025). [51]Edward Yeo, Yuxuan Tong, Morry Niu, Graham Neubig, and Xiang Yue. 2025. Demystifying long chain-of-thought reasoning in llms. arXiv preprint arXiv:2502.03373 (2025). [52]Huanyu Zhang, Wenshan Wu, Chengzu Li, Ning Shang, Yan Xia, Yangyu Huang, Yifan Zhang, Li Dong, Zhang Zhang, Liang Wang, et al.2025. Latent sketchpad: Sketching visual thoughts to elicit multimodal reasoning in mllms. arXiv preprint arXiv:2510.24514 (2025). [53]Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Yu Qiao, et al.2024. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems?. In European Conference on Computer Vision. Springer, 169–186. [54] Yi-Fan Zhang, Xingyu Lu, Shukang Yin, Chaoyou Fu, Wei Chen, Xiao Hu, Bin Wen, Kaiyu Jiang, Changyi Liu, Tianke Zhang, et al.2025. Thyme: Think beyond images. arXiv preprint arXiv:2508.11630 (2025). [55]Zhen Zhang, Xuehai He, Weixiang Yan, Ao Shen, Chenyang Zhao, Shuohang Wang, Yelong Shen, and Xin Eric Wang. 2025. Soft thinking: Unlocking the reason- ing potential of llms in continuous concept space. arXiv preprint arXiv:2505.15778 (2025). [56]Kesen Zhao, Beier Zhu, Qianru Sun, and Hanwang Zhang. 2025. Unsupervised visual chain-of-thought reasoning via preference optimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 2303–2312. [57]Ziwei Zheng, Michael Yang, Jack Hong, Chenxiao Zhao, Guohai Xu, Le Yang, Chao Shen, and Xing Yu. 2025. Deepeyes: Incentivizing" thinking with images" via reinforcement learning. arXiv preprint arXiv:2505.14362 (2025). [58]Sifan Zhou, Yichao Cao, Jiahao Nie, Yuqian Fu, Ziyu Zhao, Xiaobo Lu, and Shuo Wang. 2026. Comptrack: Information bottleneck-guided low-rank dynamic token compression for point cloud tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 13773–13781.