Uni-LaDiR: Latent Diffusion Unifies Multimodal Reasoning

# Uni-LaDiR: Latent Diffusion UnifiesMultimodal Reasoning
Source: [https://arxiv.org/html/2609.19878](https://arxiv.org/html/2609.19878)
## Uni\-LaDiR: Latent Diffusion Unifies Multimodal Reasoning

Yizhe Zhang2Nikki Lijing Kuang1Yi\-An Ma1Lianhui Qin1Affiliation:1UC San Diego2Meta

###### Abstract

Multimodal reasoning requires models to draw on information from multiple modalities throughout the reasoning process\. Yet existing methods often concatenate or interleave modality\-specific thought tokens in a single sequence, leaving the model to bridge representational differences as it reasons across modalities\. We introduce Uni\-LaDiR \(UnifiedLatentDiffusionReasoner\), a framework that brings these thoughts into a shared latent space for reasoning\. A unified encoder maps teacher reasoning steps from different modalities into shared thought tokens, trained to preserve the information needed for later reasoning steps and the final answer or action\. Because the same context can support multiple valid next steps, we use diffusion to predict the next block of thought tokens from the input and preceding blocks\. Jointly training the encoder and diffusion reasoner with shared model weights encourages thought tokens to be both useful for the task and predictable from the available context\. At inference, the model generates these tokens without teacher observations\. Across eleven vision\-language model \(VLM\) benchmarks and two vision\-language\-action \(VLA\) suites, Uni\-LaDiR achieves relative gains over the strongest evaluated baselines of7\.3%in average mathematical and logical reasoning accuracy and6\.1%in RLBench manipulation success\.

![Refer to caption](https://arxiv.org/html/2609.19878v1/fig1-0912-motion-detail-v76.png)Figure 1:Multimodal reasoning in a shared latent space\.Modality\-specific reasoning \(left\) switches representations between steps\. Uni\-LaDiR \(right\) uses a shared latent space and diffusion to generate thought tokens for answers or actions\.## 1Introduction

A world can be observed in different modalities\. An image shows how it looks like, language describes what is happening, and 3D point clouds reveal the geometry and robot states represent the interaction\. These modalities provide complementary evidence\. Many multimodal problems can only be solved by combining them\([Lu et al\., 2022b](https://arxiv.org/html/2609.19878#bib.bib74);[Lu et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib48);[Hao et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib51)\)and a capable multimodal reasoner should therefore be able to use information from different modalities in its reasoning process\([Gao et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib90);[Gu et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib16)\)\.

Recent work has explored several ways to combine multimodal information in Chain\-of\-Thought reasoning\. For vision\-language models \(VLMs\), prior work interleaves text with visual reasoning steps, where the visual content takes the form of generated images\([Li et al\., 2025c](https://arxiv.org/html/2609.19878#bib.bib91);[Gu et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib16)\)or latent visual tokens\([Li et al\., 2025b](https://arxiv.org/html/2609.19878#bib.bib12);[Yang et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib44)\)\. For vision\-language\-action models \(VLAs\), prior work combines language, visual inputs, and actions\([Brohan et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib95);[Kim et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib31);[Zhao et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib34)\), with some methods also incorporating 3D information and robot states\([Zhen et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib136);[Qu et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib32);[Liu et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib14)\)\. Despite their differences, most methods follow a similar pattern: each modality is encoded into its own tokens or features, which are then concatenated or interleaved in the LLM sequence\([Gao et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib90);[Li et al\., 2025b](https://arxiv.org/html/2609.19878#bib.bib12);[Yang et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib44);[Liu et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib14)\)\. As a result, reasoning still happens over modality\-specific representations\. When intermediate reasoning repeatedly transfers information between modalities, this entangles*cross\-modal representation conversion*with*reasoning dynamics*\. The model needs to determine not only how the underlying reasoning state should change, but also how that state is expressed in the representation of the next modality\.

We argue that*while perception can remain modality\-dependent, reasoning dynamics should operate over a unified reasoning space*\. Each modality is specialized for extracting different evidence from the world, but reasoning only requires the task\-relevant information that this evidence implies\. Once such information is encoded into a shared space, subsequent reasoning can evolve independently of the observation modality\. This design gives subsequent reasoning a shared token interface instead of requiring modality\-specific latent formats\([Kaelbling et al\., 1998](https://arxiv.org/html/2609.19878#bib.bib133);[Hafner et al\., 2019](https://arxiv.org/html/2609.19878#bib.bib134);[Hafner et al\., 2020](https://arxiv.org/html/2609.19878#bib.bib135)\)\.

Based on this idea, we introduce Uni\-LaDiR, a unified latent diffusion framework for multimodal reasoning\. Unlike prior approaches that directly concatenate outputs from modality\-specific encoders, Uni\-LaDiR introduces a unified encoder that maps heterogeneous modality features into a*shared latent reasoning space*\. Specifically, every teacher reasoning step is encoded into a fixed\-width latent block as thought tokens, and each block is trained to support the generation of future reasoning steps and the final task output, rather than to reconstruct its source modality\. This encourages the thought tokens to preserve information according to its utility for subsequent reasoning, while factoring out modality\-specific representational differences that are irrelevant to the task\.

However, learning to predict these thought tokens is not trivial\. First, the next reasoning step is often not unique: the same context may support multiple valid intermediate thoughts or actions\. We therefore use diffusion to model the conditional distribution of the next thought tokens given the input and preceding thought tokens\([Ho et al\., 2020](https://arxiv.org/html/2609.19878#bib.bib129);[Rombach et al\., 2022](https://arxiv.org/html/2609.19878#bib.bib132);[Chi et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib131);[Kang et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib10);[Kang et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib113)\)\. Second, useful thought tokens should satisfy two properties:*reasoning sufficiency*and*context predictability*\. They should preserve the information needed for subsequent reasoning and the final answer, while remaining predictable from the available context at inference time\. We therefore jointly train the unified encoder and the latent diffusion reasoner, allowing the representation space and the reasoning model to adapt to each other\. This encourages thought tokens that retain task\-relevant information and can be generated from the available context\.

We evaluate Uni\-LaDiR on eleven VLM reasoning benchmarks and two VLA manipulation suites\. Compared with the strongest reported modality\-specific latent reasoning baseline in each setting, Uni\-LaDiR improves mean accuracy by7\.3%relative across four mathematical and logical VLM benchmarks and mean success by6\.1%relative across ten RLBench manipulation tasks\. Within the unified latent space, flow matching improves the average by14\.9%over direct L2 prediction and19\.5%over cosine similarity loss \(Figure[6](https://arxiv.org/html/2609.19878#S4.F6)\)\. These results support Uni\-LaDiR’s approach of separating modality\-specific perception from shared latent reasoning, with thought\-token representations and their prediction learned together for the task\.

## 2Preliminaries

#### Problem formulation\.

Given an inputxxcontaining multimodal observations and a query or instruction, the reasoner generatesKKintermediate thought\-token blocksz1:Kz\_\{1:K\}before predicting the task outputyy, which is a text answer for VLMs or a robot action for VLAs\. The intermediate blocks integrate task\-relevant evidence across modalities and mediate prediction ofyy\.

#### Latent multimodal CoT\.

Existing latent multimodal CoT represents intermediate reasoning with modality\-specific latent tokens\. In VLMs, these tokens can encode intermediate visual content\([Li et al\., 2025b](https://arxiv.org/html/2609.19878#bib.bib12);[Yang et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib44)\); in VLAs, they can encode future visual observations, 3D geometry, or robot states\([Liu et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib14)\)\. The resulting blocks are concatenated or interleaved in the model sequence:

pθ\(z1:K,y∣x\)=∏i=1Kpθ​\(zi∣\[x;z<i\]\)⏟intermediate reasoningpθ\(y∣\[x;z1:K\]\)⏟final output\.p\_\{\\theta\}\(z\_\{1:K\},y\\mid x\)=\\underbrace\{\\prod\_\{i=1\}^\{K\}p\_\{\\theta\}\(z\_\{i\}\\mid\[x;z\_\{<i\}\]\)\}\_\{\\text\{intermediate reasoning\}\}\\underbrace\{p\_\{\\theta\}\(y\\mid\[x;z\_\{1:K\}\]\)\}\_\{\\text\{final output\}\}\.\(1\)Here\[;\]\[\\,;\\,\]denotes sequence concatenation\. During training, a modality\-specific encoder maps theii\-th teacher reasoning step, such as an intermediate image or robot state, to featuressis\_\{i\}that supervise the corresponding thought blockziz\_\{i\}\. At inference, the teacher steps are unavailable, and the model must generate eachziz\_\{i\}fromxxand the preceding thought blocks\.

## 3Unified Latent Diffusion Reasoning

Uni\-LaDiR unifies different reasoning steps from different modalities into the same latent space to provide a shared interface for reasoning across modalities\. In this section, we first formalize the requirements of this shared reasoning space \(Section[3\.1](https://arxiv.org/html/2609.19878#S3.SS1)\), then describe how teacher steps are mapped to shared thought tokens \(Section[3\.2](https://arxiv.org/html/2609.19878#S3.SS2)\), how downstream supervision grounds these tokens in subsequent reasoning and task prediction \(Section[3\.3](https://arxiv.org/html/2609.19878#S3.SS3)\), and how latent diffusion learns to generate them from the context available at inference \(Section[3\.4](https://arxiv.org/html/2609.19878#S3.SS4)\)\. Figure[2](https://arxiv.org/html/2609.19878#S3.F2)provides an overview\.

### 3\.1Motivation

Modality\-specific latent CoT couples each reasoning step with a change in representation: predicting the next step requires determining not only its task\-relevant content, but also how that content is expressed in the target modality\. We instead require every intermediate step to inhabit a shared latent reasoning space, independent of the modality from which its supervision originates\.

A useful shared representation must satisfy two complementary properties\.*Reasoning sufficiency*requires each thought block to retain the information needed for subsequent reasoning steps and the final output\.*Context predictability*requires the same block to be recoverable from the input and preceding thought tokens available at inference\. We enforce the former through downstream grounding supervision and the latter through latent diffusion joint training, while sharing the backbone between representation construction and generation\([Duggal et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib137)\)\.

Figure 2:Uni\-LaDiR overview\.During training, a shared backbone\(a\)encodes teacher steps into thought tokens,\(b\)grounds them through downstream predictions, and\(c\)learns their generation by diffusion\. At inference, it generates thought tokens from the task input to predict an answer or action\.
### 3\.2Constructing shared thought tokens

We train a unified encoder to map each teacher reasoning step to the corresponding thought tokenszi⋆z\_\{i\}^\{\\star\}in a unified reasoning space \(Figure[2](https://arxiv.org/html/2609.19878#S3.F2)a\)\. After modality\-specific encoding, we append a fixed\-length block of learnable embeddingsqqto each teacher featuresis\_\{i\}\. Then our unified encoder maps the last hidden states at theqqpositions to the clean thought tokenszi⋆z\_\{i\}^\{\\star\}\.

Every modality uses the sameqq, output projection, and number of thought tokens per block\. During training, we pack multiple teacher steps and their corresponding copies ofqqinto one forward pass\. The construction mask isolates the steps: each copy ofqqattends only to itself and its associated teacher features, preventing information exchange across reasoning steps \(Figure[2](https://arxiv.org/html/2609.19878#S3.F2)a\)\. Thus, all teacher modalities are now encoded into a shared latent space, where downstream supervision determines which information is retained for better subsequent reasoning\.

### 3\.3Grounding thought tokens in downstream reasoning

Grounding trains each thought block to preserve information needed by subsequent reasoning\. Givenxxand the prefixz<i⋆z\_\{<i\}^\{\\star\}, the backbone predicts the next teacher targetsiGTs\_\{i\}^\{\\mathrm\{GT\}\}; given the complete sequencez1:K⋆z\_\{1:K\}^\{\\star\}, it predicts the final outputyy\(Figure[2](https://arxiv.org/html/2609.19878#S3.F2)b\)\. For example,z1⋆z\_\{1\}^\{\\star\}contributes to predictings2GTs\_\{2\}^\{\\mathrm\{GT\}\}, while\(z1⋆,z2⋆\)\(z\_\{1\}^\{\\star\},z\_\{2\}^\{\\star\}\)contribute to predictings3GTs\_\{3\}^\{\\mathrm\{GT\}\}\. A grounding attention mask computes these predictions in one forward pass while exposing each prediction only to its permitted thought\-token prefix and hiding later blocks and raw teacher steps \(Appendix[A\.1](https://arxiv.org/html/2609.19878#A1.SS1)\)\.

Lets^i\\hat\{s\}\_\{i\}andy^\\hat\{y\}denote the corresponding predictions\. We optimize

ℒground=∑i=2Kℓ⁡\(s^i,siGT\)⏟reasoning steps\+λy​ℓ⁡\(y^,y\)⏟final output,ℓ⁡\(u^,u\)=\{CE⁡\(u^,u\),u​is text,‖u^−u‖22,u​is continuous\.\\mathcal\{L\}\_\{\\mathrm\{ground\}\}=\\underbrace\{\\sum\_\{i=2\}^\{K\}\\ell\(\\hat\{s\}\_\{i\},s\_\{i\}^\{\\mathrm\{GT\}\}\)\}\_\{\\text\{reasoning steps\}\}\+\\lambda\_\{y\}\\underbrace\{\\ell\(\\hat\{y\},y\)\}\_\{\\text\{final output\}\},\\qquad\\ell\(\\hat\{u\},u\)=\\begin\{cases\}\\operatorname\{CE\}\(\\hat\{u\},u\),&u\\text\{ is text\},\\\\ \\\|\\hat\{u\}\-u\\\|\_\{2\}^\{2\},&\{\\color\[rgb\]\{0,0,0\}u\\text\{ is continuous\}\}\.\\end\{cases\}\(2\)Hereuudenotes either a reasoning\-step target or the final output, andλy\\lambda\_\{y\}weights final\-output supervision\.The continuous reasoning\-step targets are image and 3D point\-cloud features and robot\-state vectors, supervised by the L2 term above\. For VLAs, the final actionyyis instead produced by a flow\-matching action head, so its termℓ⁡\(y^,y\)\\ell\(\\hat\{y\},y\)is that head’s flow\-matching loss rather than an L2 regression\.The targetssiGTs\_\{i\}^\{\\mathrm\{GT\}\}andyyare fixed, and any encoder used to construct target features is frozen\. Gradients fromℒground\\mathcal\{L\}\_\{\\mathrm\{ground\}\}pass throughzi⋆z\_\{i\}^\{\\star\}into the unified encoder and shared backbone, learning the latent thought tokens to optimize their utility for subsequent reasoning\.

### 3\.4Generating thought tokens with joint diffusion training

The diffusion objective trains the backbone to generate each thought block from the context available at inference \(Figure[2](https://arxiv.org/html/2609.19878#S3.F2)c\)\. Diffusion models a conditional distribution over the next thought tokens, allowing different valid thoughts under the same input and preceding thought tokens\([Kang et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib10);[Kang et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib113)\)\. During training, Gaussian noise perturbs each clean blockzi⋆z\_\{i\}^\{\\star\}to producezitz\_\{i\}^\{t\}\. Conditioned onxx, the preceding clean blocksz<i⋆z\_\{<i\}^\{\\star\}, and noise leveltt, the shared backbone predicts the denoising velocityv^i\\hat\{v\}\_\{i\}\.

We again process all reasoning steps in one forward pass\. The diffusion mask restricts each noisy block to its own tokens,xx, and the preceding clean thought blocks, preventing access to future thought blocks \(Appendix[A\.1](https://arxiv.org/html/2609.19878#A1.SS1)\)\. We optimize the flow\-matching objective\([Lipman et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib130)\)

ℒdiff=𝔼t,ϵ​\[∑i=1K‖v^i−vi⋆‖22\],\\mathcal\{L\}\_\{\\mathrm\{diff\}\}=\\mathbb\{E\}\_\{t,\\epsilon\}\\\!\\left\[\\sum\_\{i=1\}^\{K\}\\\|\\hat\{v\}\_\{i\}\-v\_\{i\}^\{\\star\}\\\|\_\{2\}^\{2\}\\right\],\(3\)wherevi⋆v\_\{i\}^\{\\star\}is the target velocity induced by the sampled noiseϵ\\epsilonand timett\. The predicted velocity guides denoising to generate the thought\-token blockz^i\\hat\{z\}\_\{i\}\. Here,ttindexes denoising within a block, whileiiindexes reasoning steps\.

#### Joint training\.

Inspired by[Duggal et al\. \(2026\)](https://arxiv.org/html/2609.19878#bib.bib137), we jointly train the unified encoder and diffusion reasoner with a shared backbone, optimizing the latent thought tokens to be both useful for downstream reasoning and predictable from the available context\. We optimize

ℒ=ℒground\+λ​ℒdiff,\\mathcal\{L\}=\\mathcal\{L\}\_\{\\mathrm\{ground\}\}\+\\lambda\\mathcal\{L\}\_\{\\mathrm\{diff\}\},\(4\)whereλ\\lambdabalances the two objectives\. On the diffusion branch, we stop gradients through the clean thought tokenszi⋆z\_\{i\}^\{\\star\}, including both the conditioning prefix and the noising target\. Thus,ℒground\\mathcal\{L\}\_\{\\mathrm\{ground\}\}updates the unified encoder and shared backbone throughzi⋆z\_\{i\}^\{\\star\}, whereasℒdiff\\mathcal\{L\}\_\{\\mathrm\{diff\}\}updates the shared backbone while treating the current clean tokens as fixed targets\. Because the same backbone constructs and predicts the latent representations, the two objectives jointly improve the thought\-token representations and the model that predicts them\.

#### Inference\.

At inference, the teacher reasoning steps and unified encoding path are not required\. Starting fromxx, the model generatesz1,…,zKz\_\{1\},\\ldots,z\_\{K\}sequentially by denoising one latent block at a time, with each block conditioned on the previously generated blocks \(Figure[2](https://arxiv.org/html/2609.19878#S3.F2), bottom\)\. The complete latent sequence then conditions prediction of the final answer or actionyy\.

## 4Experiments

Our experiments test whether a shared latent reasoning space improves both visual reasoning and robotic manipulation\. Section[4\.1](https://arxiv.org/html/2609.19878#S4.SS1)introduces the benchmarks and baselines; Section[4\.2](https://arxiv.org/html/2609.19878#S4.SS2)reports task performance and manipulation inference efficiency; Section[4\.3](https://arxiv.org/html/2609.19878#S4.SS3)isolates the effect of unifying teacher modalities; and Section[4\.4](https://arxiv.org/html/2609.19878#S4.SS4)examines the training objectives and joint optimization\.

### 4\.1Evaluation protocol

#### Visual reasoning\.

We train on Zebra\-CoT\([Li et al\., 2025a](https://arxiv.org/html/2609.19878#bib.bib15)\), which interleaves text and visual reasoning steps, and use Qwen2\.5\-VL\-7B for the main experiment\. We report accuracy on seven vision\-centric benchmarks: VisualPuzzles\([Song et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib28)\), ChartQA\([Masry et al\., 2022](https://arxiv.org/html/2609.19878#bib.bib22)\), V∗\([Wu and Xie, 2024](https://arxiv.org/html/2609.19878#bib.bib23)\), BLINK\-Jigsaw\([Fu et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib24)\), MMVP\([Tong et al\., 2024b](https://arxiv.org/html/2609.19878#bib.bib25)\), SAT\([Ray et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib26)\), and CV\-Bench\([Tong et al\., 2024a](https://arxiv.org/html/2609.19878#bib.bib27)\)\. We also evaluate four mathematical or logical benchmarks: MathVista\([Lu et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib48)\), MathVision\([Wang et al\., 2024a](https://arxiv.org/html/2609.19878#bib.bib49)\), VisuLogic\([Xu et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib50)\), and EMMA\([Hao et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib51)\)\. Appendix[B\.3](https://arxiv.org/html/2609.19878#A2.SS3)reports results with additional backbones\.

#### VLA manipulation\.

We evaluate LIBERO\([Liu et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib29)\)and RLBench\([James et al\., 2020](https://arxiv.org/html/2609.19878#bib.bib30)\)\. In this setting, future images, 3D point clouds, and robot states serve as training\-only teacher CoTs; continuous actions remain task outputs\. Following[Liu et al\. \(2026\)](https://arxiv.org/html/2609.19878#bib.bib14), our VLA backbone is a mixture\-of\-transformers \(MoT\) that couples two Janus\-Pro\-1\.5B experts \(3\.3B parameters total\), and actions are produced by a flow\-matching action head trained with the same action objective\. We report success rate and, on RLBench, control frequency without action chunking\.

#### Baselines\.

For VLMs, we compare with the Qwen2\.5\-VL base model and explicit reasoners DeepEyes\([Zheng et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib39)\), PixelReasoner\([Wang et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib81)\), Vision\-R1\([Huang and others, 2025b](https://arxiv.org/html/2609.19878#bib.bib80)\), and Open\-Vision\-Reasoner\([Wei et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib138)\)\. Latent reasoning baselines include LVR\([Li et al\., 2025b](https://arxiv.org/html/2609.19878#bib.bib12)\), Mirage\([Yang et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib44)\), CoVT\([Qin et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib45)\), Monet\([Wang et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib46)\),VaLR\-M\([Jeon et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib40)\), SLVR\([Xu et al\., 2026a](https://arxiv.org/html/2609.19878#bib.bib42)\), Mull\-Tokens\([Ray et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib43)\), and ILVR\([Dong et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib17)\)\. For VLAs, we consider three groups\.*Action policies*include OpenVLA\([Kim et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib31)\), SpatialVLA\([Qu et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib32)\), CogACT\([Li et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib33)\),π0\.5\\pi\_\{0\.5\}\([Physical Intelligence et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib35)\), OpenVLA\-OFT\([Kim et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib36)\), HybridVLA\([Liu et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib37)\), and UniVLA\([Bu et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib123)\)\.*Explicit reasoning models*include ManipLLM\([Li et al\., 2023a](https://arxiv.org/html/2609.19878#bib.bib139)\), WorldVLA\([Cen and others, 2025](https://arxiv.org/html/2609.19878#bib.bib122)\), CoT\-VLA\([Zhao et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib34)\), FlowVLA\([Zhong et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib124)\), and FUTURE\-VLA\([Xu et al\., 2026b](https://arxiv.org/html/2609.19878#bib.bib125)\)\.*Latent reasoning models*include FiS\-VLA\([Chen et al\., 2025a](https://arxiv.org/html/2609.19878#bib.bib38)\), ConsisVLA\-4D\([Li et al\., 2026c](https://arxiv.org/html/2609.19878#bib.bib126)\), LaST0\([Liu et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib14)\), AVA\-VLA\([Lei and Shan, 2026](https://arxiv.org/html/2609.19878#bib.bib127)\), and PearlVLA with and without PRL\([Yang and Shan, 2026](https://arxiv.org/html/2609.19878#bib.bib128)\)\. Full settings appear in Appendix[B](https://arxiv.org/html/2609.19878#A2)\.

### 4\.2Main results

#### VLM visual reasoning\.

On the seven vision\-centric benchmarks, Uni\-LaDiR improves average accuracy over Qwen2\.5\-VL\-7B by18\.1%; on the four mathematical and logical benchmarks, the relative gain is34\.7%\(Table[1](https://arxiv.org/html/2609.19878#S4.T1)\)\. It also outperforms the strongest modality\-specific latent baseline in each group:1\.7%over Mirage and7\.3%over ILVR on average\. The gains over modality\-specific baselines reach30\.3%on VisualPuzzles over Mirage, and10\.8%on MathVision and12\.6%on VisuLogic over ILVR\. The advantage over modality\-specific baselines suggests that sharing reasoning representations benefits both visual reasoning and mathematical tasks\.

Table 1:Comparison on VLM reasoning benchmarks\.Accuracy \(%\) of Uni\-LaDiR and baselines with Qwen2\.5\-VL\-7B on\(a\)vision\-centric and\(b\)mathematical and logical benchmarks\. Best results are highlighted in bold\.\(a\) Vision\-Centric Perception, Search, and Spatial Reasoning\.

\(b\) Visual Mathematics and Logic\.

#### VLA manipulation\.

Uni\-LaDiR improves average success over the strongest prior latent policies by0\.45 percentage pointsover PearlVLA \+ PRL on LIBERO and6\.1%relative to LaST0on RLBench \(Tables[3](https://arxiv.org/html/2609.19878#S4.T3)and[2](https://arxiv.org/html/2609.19878#S4.T2)\)\. On RLBench, it improves nine of ten tasks and matches the remaining one; the largest gain is16\.7%on Plants\. These reported means place Uni\-LaDiR among the strongest policies in this comparison\. Without action chunking, Uni\-LaDiR runs at16\.5 Hz, a7\.1%higher inference rate than LaST0and15\.0×\\timesthe rate of explicit CoT\-VLA\.

Table 2:Comparison of Uni\-LaDiR and baselines on RLBench\.We report mean success rates \(%\) and inference speed \(Hz\) without action chunking\.Var\. is the variance of mean success rate across three random seeds\.Best results are highlighted in bold\.Table 3:Comparison of Uni\-LaDiR and baselines on LIBERO\.We report success rates \(%\) across four suites\. Best results are highlighted in bold\.

### 4\.3Analysis

In this section, we test whether progressively unifying teacher modalities reduces the modality\-switching gap and improves reasoning\. All variants use the same teacher modalities, reasoning backbone, step order, latent\-token budget, and training schedule; we only change how teacher signals are represented\. In the separate baseline, each modality bypasses the unified encoder and directly uses its modality\-specific features as latent targets\. In the partially shared variants, only the specified modalities pass through the same unified encoder to form shared thought tokens, while the remaining modalities keep their original modality\-specific targets\. Full sharing routes all available teacher modalities through the unified encoder \(Appendix[B\.2](https://arxiv.org/html/2609.19878#A2.SS2)\)\. As shown in Figure[3](https://arxiv.org/html/2609.19878#S4.F3)a, for visual reasoning, moving text and image from separate targets into the shared space improves mean accuracy by9\.0%relative\. On RLBench, sharing any two of the visual, 3D, and state modalities improves success by1\.8–3\.0%, while sharing all three yields a6\.1%gain over separate targets and a3\.0%gain over the best pair \(Figure[3](https://arxiv.org/html/2609.19878#S4.F3)b\)\. These growining gains with more unified modalities suggest that our method effectively reduce the modality\-switching gap, allowing the reasoning model is able to focus more on evolving the reasoning state rather than adapting between modality\-specific representations\.

Figure 3:Results of unifying more modalities\.\(a\)Qwen2\.5\-VL\-7B mean accuracy on V∗, MMVP, MathVista, and EMMA\.\(b\)RLBench success with different modality groupings\. Error bars denote standard deviation across three seeds\.
### 4\.4Ablation Study

#### Grounding objective\.

We first test how thought tokens should be supervised\. With the model, teachers, latent\-token budget, and training schedule fixed, we replace source reconstruction with prediction of later reasoning steps\. Both variants retain the same final\-output supervision and its gradients into the encoder \(Appendix[B\.2](https://arxiv.org/html/2609.19878#A2.SS2)\)\. Downstream supervision improves the six\-benchmark average by5\.4%relative to reconstruction, with gains of26\.0%on V∗and7\.4%on RLBench \(Figure[4](https://arxiv.org/html/2609.19878#S4.F4)\)\. The gain spans five of six benchmarks, while MathVista decreases by 4\.20 percentage points; downstream grounding improves the aggregate without benefiting every task\.

Figure 4:Results of different grounding objectives\.We compare frozen latent targets, an encoder trained before the reasoner, and joint training under the same architecture and teacher supervision\. Joint training improves the six\-benchmark mean by11\.3%relative to separate stages and21\.9%relative to frozen targets \(Figure[5](https://arxiv.org/html/2609.19878#S4.F5)\)\. Relative to separate stages, it improves all six benchmarks, with gains of8\.8%on EMMA and15\.0%on RLBench\. These results favor learning the representation and its predictor together over fixing the representation before training the reasoner\.

Figure 5:Results of joint versus separate training of the encoder and reasoner\.
#### Diffusion objective\.

Finally, we test whether diffusion improves thought\-token prediction when the latent targets are fixed\. We compare flow matching with direct thought\-token prediction trained using squared L2 or cosine similarity loss \(Appendix[B\.2](https://arxiv.org/html/2609.19878#A2.SS2)\)\. Flow matching improves the six\-benchmark average by14\.9%relative to L2 and19\.5%relative to cosine similarity loss \(Figure[6](https://arxiv.org/html/2609.19878#S4.F6)\)\. These gains support diffusion\-based thought\-token prediction as an effective alternative to direct regression for downstream reasoning and manipulation\.

Figure 6:Results of different latent prediction objectives\.

## 5Related Work

#### Unified multimodal models\.

Early generalist models cast heterogeneous tasks as a shared sequence problem, spanning vision–language understanding, generation, and embodied control\([Wang and others, 2022](https://arxiv.org/html/2609.19878#bib.bib52);[Lu et al\., 2022a](https://arxiv.org/html/2609.19878#bib.bib53);[Lu et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib1);[Reed et al\., 2022](https://arxiv.org/html/2609.19878#bib.bib54);[Mizrahi et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib55);[Bachmann et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib56)\)\. Large\-scale vision–language models then connected pretrained language models to visual inputs through learned interfaces\([Alayrac et al\., 2022](https://arxiv.org/html/2609.19878#bib.bib57);[Chen et al\., 2022](https://arxiv.org/html/2609.19878#bib.bib58);[Chen and others, 2023](https://arxiv.org/html/2609.19878#bib.bib59);[Huang et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib60);[Peng et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib61)\)\. Autoregressive any\-to\-any systems further represented multiple modalities in a common token stream\([Sun et al\., 2023b](https://arxiv.org/html/2609.19878#bib.bib62);[Sun et al\., 2023a](https://arxiv.org/html/2609.19878#bib.bib63);[Yu et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib64);[Ge et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib65);[Jin et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib66);[Wu et al\., 2023b](https://arxiv.org/html/2609.19878#bib.bib67);[Zhan et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib68);[Chameleon Team, 2024](https://arxiv.org/html/2609.19878#bib.bib2);[Wang et al\., 2024b](https://arxiv.org/html/2609.19878#bib.bib69)\), while hybrid autoregressive–diffusion architectures unified discrete and continuous generation\([Tang et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib70);[Zhou et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib3);[Xie et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib71);[Sun et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib4);[Wu et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib72);[Chen et al\., 2025b](https://arxiv.org/html/2609.19878#bib.bib21);[Deng et al\., 2025a](https://arxiv.org/html/2609.19878#bib.bib20);[Xu and others, 2025a](https://arxiv.org/html/2609.19878#bib.bib73);[Bai et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib19)\)\. Whereas these models unify what is read and generated—their inputs, outputs, and task interfaces—Uni\-LaDiR unifies the reasoning in between, learning a single latent space in which the intermediate thoughts from every modality are represented and predicted\.

#### Multimodal reasoning\.

Multimodal chain\-of\-thought methods supervise or elicit textual rationales from visual evidence\([Lu et al\., 2022b](https://arxiv.org/html/2609.19878#bib.bib74);[Zhang et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib5);[Rose and others, 2023](https://arxiv.org/html/2609.19878#bib.bib75);[Xu and others, 2024](https://arxiv.org/html/2609.19878#bib.bib76);[Dong and others, 2024](https://arxiv.org/html/2609.19878#bib.bib77);[Guo and others, 2024](https://arxiv.org/html/2609.19878#bib.bib78);[Yao and others, 2024](https://arxiv.org/html/2609.19878#bib.bib79);[Huang and others, 2025b](https://arxiv.org/html/2609.19878#bib.bib80)\)\. A complementary line externalizes computation through programs, tools, image operations, spatial abstractions, or visual scratchpads\([Gupta and Kembhavi, 2022](https://arxiv.org/html/2609.19878#bib.bib82);[Suris et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib83);[Yang et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib84);[Wu et al\., 2023a](https://arxiv.org/html/2609.19878#bib.bib85);[Hu and others, 2023](https://arxiv.org/html/2609.19878#bib.bib86);[Chen and others, 2024](https://arxiv.org/html/2609.19878#bib.bib87);[Hu et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib88);[Zhou and others, 2024](https://arxiv.org/html/2609.19878#bib.bib89)\)\. Recent work interleaves language with visual thoughts rather than forcing every step into text\([Gao et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib90);[Li et al\., 2025a](https://arxiv.org/html/2609.19878#bib.bib15);[Gu et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib16);[Li et al\., 2026b](https://arxiv.org/html/2609.19878#bib.bib18)\)\. In embodied settings, language plans, visual predictions, and state or action trajectories have been used to guide control\([Ahn et al\., 2022](https://arxiv.org/html/2609.19878#bib.bib92);[Huang et al\., 2022](https://arxiv.org/html/2609.19878#bib.bib93);[Driess et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib94);[Brohan et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib95);[Jiang et al\., 2022](https://arxiv.org/html/2609.19878#bib.bib96);[Li et al\., 2023b](https://arxiv.org/html/2609.19878#bib.bib97);[Kim et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib31);[Li et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib33);[Zhao et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib34);[Zhong et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib98);[Bai et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib13);[Liu et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib14)\)\. These methods still reason by switching between modality\-specific representations—textual rationales, generated images, or fixed domain\-specific latents—whereas Uni\-LaDiR reasons inside one shared interface into which text, images, 3D structure, and robot state are all encoded; actions remain task outputs\.

#### Latent reasoning\.

Latent\-computation methods internalize explicit rationales, distill hidden reasoning, or allocate additional computation without decoding every step\([Zelikman et al\., 2022](https://arxiv.org/html/2609.19878#bib.bib99);[Goyal et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib100);[Deng and others, 2023](https://arxiv.org/html/2609.19878#bib.bib101);[Deng and others, 2024](https://arxiv.org/html/2609.19878#bib.bib102);[Pfau et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib103);[Zelikman et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib104);[Yu and others, 2024](https://arxiv.org/html/2609.19878#bib.bib105);[Wang and others, 2024](https://arxiv.org/html/2609.19878#bib.bib106)\)\. Recurrent and concept\-level models instead treat depth, loops, or continuous semantic states as the substrate of computation\([Dehghani et al\., 2018](https://arxiv.org/html/2609.19878#bib.bib107);[Giannou et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib108);[Yang and others, 2023](https://arxiv.org/html/2609.19878#bib.bib109);[Barrault and others, 2024](https://arxiv.org/html/2609.19878#bib.bib110);[Kohli and others, 2026](https://arxiv.org/html/2609.19878#bib.bib116);[Jeddi and others, 2026](https://arxiv.org/html/2609.19878#bib.bib117);[Yang and others, 2026](https://arxiv.org/html/2609.19878#bib.bib118)\)\. Continuous chain\-of\-thought approaches directly optimize hidden vectors that replace or compress language rationales\([Hao et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib6);[Shen et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib7);[Xu and others, 2025c](https://arxiv.org/html/2609.19878#bib.bib111);[Xu and others, 2025b](https://arxiv.org/html/2609.19878#bib.bib112);[Deng et al\., 2025b](https://arxiv.org/html/2609.19878#bib.bib8)\)\. Multimodal variants construct continuous thoughts from visual features, dense perceptual teachers, helper images, or task loss\([Pham and Ngo, 2025](https://arxiv.org/html/2609.19878#bib.bib11);[Li et al\., 2025b](https://arxiv.org/html/2609.19878#bib.bib12);[Yang et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib44);[Qin et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib45);[Wang et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib46);[Jeon et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib40);[Viveiros et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib41);[Li et al\., 2026a](https://arxiv.org/html/2609.19878#bib.bib47);[Xu et al\., 2026a](https://arxiv.org/html/2609.19878#bib.bib42);[Ray et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib43);[Hu et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib9)\), while diffusion\-based approaches model multimodal or multi\-step latent trajectories\([Kang et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib10);[Kang et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib113);[He and others, 2023](https://arxiv.org/html/2609.19878#bib.bib114);[Huang and others, 2025a](https://arxiv.org/html/2609.19878#bib.bib115);[Bai et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib13);[Liu et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib14);[Viveiros and others, 2026](https://arxiv.org/html/2609.19878#bib.bib119);[Wu and others, 2026](https://arxiv.org/html/2609.19878#bib.bib120);[Fan and others, 2026](https://arxiv.org/html/2609.19878#bib.bib121)\)\. Prior latent reasoners either compress rationales within a single modality or give each modality its own latent format; Uni\-LaDiR instead learns one representation and one diffusion predictor shared across heterogeneous reasoning modalities, and directly tests whether this sharing outperforms matched modality\-specific designs\.

## 6Conclusion

We introduce Uni\-LaDiR, a unified latent diffusion framework for multimodal reasoning\. A shared encoder maps teacher steps from different modalities into thought tokens, and a jointly trained diffusion reasoner predicts these tokens under downstream reasoning and task supervision\. This gives heterogeneous intermediate steps a unified latent space, allowing the model to reason without generating modality\-specific representations or accessing teacher observations at inference\. Across eleven VLM benchmarks and two VLA suites, Uni\-LaDiR achieves relative improvements of 7\.3% and 6\.1% over strong prior baselines, respectively\.

### AI use statement

OpenAI Codex assisted with drafting, literature search, feedback on method formulation, mathematical objectives and experimental design, result interpretation, figures, plotting code, and numerical checks\. The authors verified all reported experimental results and retain responsibility for the final text, code, citations, figures, and claims\.

### Ethics statement

Experiments use existing visual benchmarks and simulated manipulation tasks\. The method may inherit biases from pretrained models and teacher data; simulation performance does not establish safety or reliability on physical robots\.

### Reproducibility statement

Section[3](https://arxiv.org/html/2609.19878#S3)describes the model and training objectives\. Appendix[A](https://arxiv.org/html/2609.19878#A1)details the attention masks, and Appendix[B](https://arxiv.org/html/2609.19878#A2)reports evaluation protocols and supplementary results\.

## References

- Ahnet al\.\(2022\)M\. Ahn, A\. Brohan, N\. Brown,et al\.Do as i can, not as i say: grounding language in robotic affordances\.arXiv preprint arXiv:2204\.01691\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Alayracet al\.\(2022\)J\. Alayrac, J\. Donahue, P\. Luc,et al\.Flamingo: a visual language model for few\-shot learning\.Advances in Neural Information Processing Systems\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Bachmannet al\.\(2024\)R\. Bachmann, O\. F\. Kar, D\. Mizrahi,et al\.4M\-21: an any\-to\-any vision model for tens of tasks and modalities\.arXiv preprint arXiv:2406\.09406\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Baiet al\.\(2025\)S\. Bai, Y\. Cai, R\. Chen, K\. Chen, X\. Chen, Z\. Cheng,et al\.Qwen3\-VL technical report\.arXiv preprint arXiv:2511\.21631\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Baiet al\.\(2026\)S\. Bai, J\. Lyu, W\. Zhou, Z\. Li, D\. Wang, L\. Xing, X\. Zhao, P\. Wang, Z\. Wang, C\. Chi, B\. Chen, and S\. ZhangLatent reasoning VLA: latent thinking and prediction for vision\-language\-action models\.arXiv preprint arXiv:2602\.01166\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Barraultet al\.\(2024\)L\. Barraultet al\.Large concept models: language modeling in a sentence representation space\.arXiv preprint arXiv:2412\.08821\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Brohanet al\.\(2023\)A\. Brohan, N\. Brown, J\. Carbajal,et al\.RT\-2: vision\-language\-action models transfer web knowledge to robotic control\.arXiv preprint arXiv:2307\.15818\.Cited by:[§1](https://arxiv.org/html/2609.19878#S1.p2.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Buet al\.\(2025\)Q\. Bu, Y\. Yang, J\. Cai, S\. Gao, G\. Ren, M\. Yao, P\. Luo, and H\. LiUniVLA: learning to act anywhere with task\-centric latent actions\.arXiv preprint arXiv:2505\.06111\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1)\.
- Cenet al\.\(2025\)J\. Cenet al\.WorldVLA: towards autoregressive action world model\.arXiv preprint\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1)\.
- Chameleon Team \(2024\)Chameleon TeamChameleon: mixed\-modal early\-fusion foundation models\.arXiv preprint arXiv:2405\.09818\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Chenet al\.\(2024\)B\. Chenet al\.SpatialVLM: endowing vision\-language models with spatial reasoning capabilities\.arXiv preprint arXiv:2401\.12168\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Chenet al\.\(2025a\)H\. Chen, J\. Liu, C\. Gu, Z\. Liu, R\. Zhang, X\. Li, X\. He, Y\. Guo, C\. Fu, S\. Zhang, and P\. HengFast\-in\-slow: a dual\-system vla model unifying fast manipulation within slow reasoning\.InAdvances in Neural Information Processing Systems,Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1)\.
- Chenet al\.\(2023\)X\. Chenet al\.PaLI\-X: on scaling up a multilingual vision and language model\.arXiv preprint arXiv:2305\.18565\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Chenet al\.\(2022\)X\. Chen, X\. Wang, S\. Changpinyo,et al\.PaLI: a jointly\-scaled multilingual language\-image model\.arXiv preprint arXiv:2209\.06794\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Chenet al\.\(2025b\)X\. Chen, Z\. Wu, X\. Liu, Z\. Pan, W\. Liu, Z\. Xie, X\. Yu, and C\. RuanJanus\-pro: unified multimodal understanding and generation with data and model scaling\.arXiv preprint arXiv:2501\.17811\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Chiet al\.\(2023\)C\. Chi, Z\. Xu, S\. Feng, E\. Cousineau, Y\. Du, B\. Burchfiel, R\. Tedrake, and S\. SongDiffusion policy: visuomotor policy learning via action diffusion\.InProceedings of Robotics: Science and Systems,External Links:[Document](https://dx.doi.org/10.15607/RSS.2023.XIX.026),[Link](https://www.roboticsproceedings.org/rss19/p026.html)Cited by:[§1](https://arxiv.org/html/2609.19878#S1.p5.1)\.
- Dehghaniet al\.\(2018\)M\. Dehghani, S\. Gouws, O\. Vinyals,et al\.Universal transformers\.arXiv preprint arXiv:1807\.03819\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Denget al\.\(2025a\)C\. Deng, D\. Zhu, K\. Li, C\. Gou, F\. Li, Z\. Wang, S\. Zhong, W\. Yu, X\. Nie, Z\. Song, G\. Shi, and H\. FanEmerging properties in unified multimodal pretraining\.arXiv preprint arXiv:2505\.14683\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Denget al\.\(2025b\)J\. Deng, L\. Pang, Z\. Wei, S\. Xu, Z\. Duan, K\. Xu, Y\. Song, H\. Shen, and X\. ChengLLM latent reasoning as chain of superposition\.arXiv preprint arXiv:2510\.15522\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Denget al\.\(2023\)Y\. Denget al\.Implicit chain of thought reasoning via knowledge distillation\.arXiv preprint arXiv:2311\.01460\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Denget al\.\(2024\)Y\. Denget al\.From explicit cot to implicit cot: learning to internalize cot step by step\.arXiv preprint arXiv:2405\.14838\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Donget al\.\(2026\)S\. Dong, S\. Wang, X\. Liu, C\. Li, H\. Hou, and Z\. WeiInterleaved latent visual reasoning with selective perceptual modeling\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 29316–29335\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1)\.
- Donget al\.\(2024\)Y\. Donget al\.Insight\-V: exploring long\-chain visual reasoning with multimodal large language models\.arXiv preprint arXiv:2411\.14432\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Driesset al\.\(2023\)D\. Driess, F\. Xia, M\. S\. M\. Sajjadi,et al\.PaLM\-E: an embodied multimodal language model\.arXiv preprint arXiv:2303\.03378\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Duggalet al\.\(2026\)S\. Duggal, X\. Bai, Z\. Wu, R\. Zhang, E\. Shechtman, A\. Torralba, P\. Isola, and W\. T\. FreemanEnd\-to\-end training for unified tokenization and latent denoising\.arXiv preprint arXiv:2603\.22283\.External Links:[Link](https://arxiv.org/abs/2603.22283)Cited by:[§3\.1](https://arxiv.org/html/2609.19878#S3.SS1.p2.1),[§3\.4](https://arxiv.org/html/2609.19878#S3.SS4.SSS0.Px1.p1.1)\.
- Fanet al\.\(2026\)Y\. Fanet al\.Bridging the gap between latent and explicit reasoning with looped transformers\.arXiv preprint arXiv:2606\.31779\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Fuet al\.\(2024\)X\. Fu, Y\. Hu, B\. Li, Y\. Feng, H\. Wang, X\. Lin, D\. Roth, N\. A\. Smith, W\. Ma, and R\. KrishnaBLINK: multimodal large language models can see but not perceive\.InEuropean Conference on Computer Vision,Cited by:[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px1.p1.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px1.p1.1)\.
- Gaoet al\.\(2024\)J\. Gao, Y\. Li, Z\. Cao, and W\. LiInterleaved\-modal chain\-of\-thought\.arXiv preprint arXiv:2411\.19488\.Cited by:[§1](https://arxiv.org/html/2609.19878#S1.p1.1),[§1](https://arxiv.org/html/2609.19878#S1.p2.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Geet al\.\(2023\)Y\. Ge, Y\. Ge, Z\. Zeng,et al\.Planting a SEED of vision in large language model\.arXiv preprint arXiv:2307\.08041\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Giannouet al\.\(2023\)A\. Giannou, S\. Rajput, J\. Sohn, K\. Lee, and D\. PapailiopoulosLooped transformers as programmable computers\.arXiv preprint arXiv:2301\.13196\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Goyalet al\.\(2023\)S\. Goyal, Z\. Ji, A\. S\. Rawat,et al\.Think before you speak: training language models with pause tokens\.arXiv preprint arXiv:2310\.02226\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Guet al\.\(2025\)J\. Gu, Y\. Hao, H\. W\. Wang, L\. Li, M\. Q\. Shieh, Y\. Choi, R\. Krishna, and Y\. ChengThinkMorph: emergent properties in multimodal interleaved chain\-of\-thought reasoning\.arXiv preprint arXiv:2510\.27492\.Cited by:[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px6.p1.1),[§1](https://arxiv.org/html/2609.19878#S1.p1.1),[§1](https://arxiv.org/html/2609.19878#S1.p2.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Guoet al\.\(2024\)J\. Guoet al\.MAmmoTH\-VL: eliciting multimodal reasoning with instruction tuning at scale\.arXiv preprint arXiv:2412\.05237\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Gupta and Kembhavi \(2022\)T\. Gupta and A\. KembhaviVisual programming: compositional visual reasoning without training\.arXiv preprint arXiv:2211\.11559\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Hafneret al\.\(2020\)D\. Hafner, T\. Lillicrap, J\. Ba, and M\. NorouziDream to control: learning behaviors by latent imagination\.InInternational Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=S1lOTC4tDS)Cited by:[§1](https://arxiv.org/html/2609.19878#S1.p3.1)\.
- Hafneret al\.\(2019\)D\. Hafner, T\. Lillicrap, I\. Fischer, R\. Villegas, D\. Ha, H\. Lee, and J\. DavidsonLearning latent dynamics for planning from pixels\.InProceedings of the 36th International Conference on Machine Learning,Vol\.97,pp\. 2555–2565\.External Links:[Link](https://proceedings.mlr.press/v97/hafner19a.html)Cited by:[§1](https://arxiv.org/html/2609.19878#S1.p3.1)\.
- Haoet al\.\(2024\)S\. Hao, S\. Sukhbaatar, D\. Su, X\. Li, Z\. Hu, J\. Weston, and Y\. TianTraining large language models to reason in a continuous latent space\.arXiv preprint arXiv:2412\.06769\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Haoet al\.\(2025\)Y\. Hao, J\. Gu, H\. W\. Wang, L\. Li, Z\. Yang, L\. Wang, and Y\. ChengCan MLLMs reason in multimodality? EMMA: an enhanced multimodal reasoning benchmark\.arXiv preprint arXiv:2501\.05444\.Cited by:[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px1.p1.1),[§1](https://arxiv.org/html/2609.19878#S1.p1.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px1.p1.1)\.
- Heet al\.\(2023\)L\. Heet al\.Multi\-modal latent space learning for chain\-of\-thought reasoning in language models\.arXiv preprint arXiv:2312\.08762\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Hoet al\.\(2020\)J\. Ho, A\. Jain, and P\. AbbeelDenoising diffusion probabilistic models\.InAdvances in Neural Information Processing Systems,Vol\.33\.External Links:[Link](https://arxiv.org/abs/2006.11239)Cited by:[§1](https://arxiv.org/html/2609.19878#S1.p5.1)\.
- Huet al\.\(2026\)L\. Hu, S\. Qin, Z\. Liao, Q\. Guo, L\. Wan, W\. Feng, and Y\. LiuCoLT: teaching multi\-modal models to think with chain of latent thoughts\.arXiv preprint arXiv:2606\.31986\.Cited by:[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px2.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Huet al\.\(2023\)Y\. Huet al\.Visual program distillation: distilling tools and programmatic reasoning into vision\-language models\.arXiv preprint arXiv:2312\.03052\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Huet al\.\(2024\)Y\. Hu, W\. Shi, X\. Fu,et al\.Visual sketchpad: sketching as a visual chain of thought for multimodal language models\.arXiv preprint arXiv:2406\.09403\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Huanget al\.\(2025a\)C\. Huanget al\.ThinkAct: vision\-language\-action reasoning via reinforced visual latent planning\.arXiv preprint arXiv:2507\.16815\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Huanget al\.\(2023\)S\. Huang, L\. Dong, W\. Wang,et al\.Language is not all you need: aligning perception with language models\.arXiv preprint arXiv:2302\.14045\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Huanget al\.\(2022\)W\. Huang, F\. Xia, T\. Xiao,et al\.Inner monologue: embodied reasoning through planning with language models\.arXiv preprint arXiv:2207\.05608\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Huanget al\.\(2025b\)W\. Huanget al\.Vision\-R1: incentivizing reasoning capability in multimodal large language models\.arXiv preprint arXiv:2503\.06749\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Jameset al\.\(2020\)S\. James, Z\. Ma, D\. R\. Arrojo, and A\. J\. DavisonRLBench: the robot learning benchmark and learning environment\.IEEE Robotics and Automation Letters5\(2\),pp\. 3019–3026\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px2.p1.1)\.
- Jeddiet al\.\(2026\)A\. Jeddiet al\.LoopFormer: elastic\-depth looped transformers for latent reasoning via shortcut modulation\.arXiv preprint arXiv:2602\.11451\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Jeonet al\.\(2026\)B\. Jeon, Y\. Jeong, H\. Lee, M\. Cho, and J\. ShinVision\-aligned latent reasoning for multi\-modal large language models\.arXiv preprint arXiv:2602\.04476\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Jianget al\.\(2022\)Y\. Jiang, A\. Gupta, Z\. Zhang,et al\.VIMA: general robot manipulation with multimodal prompts\.arXiv preprint arXiv:2210\.03094\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Jinet al\.\(2023\)Y\. Jin, K\. Xu, K\. Xu,et al\.LaVIT: unified language\-vision pretraining in llm with dynamic discrete visual tokenization\.arXiv preprint arXiv:2309\.04669\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Kaelblinget al\.\(1998\)L\. P\. Kaelbling, M\. L\. Littman, and A\. R\. CassandraPlanning and acting in partially observable stochastic domains\.Artificial Intelligence101\(1–2\),pp\. 99–134\.External Links:[Document](https://dx.doi.org/10.1016/S0004-3702%2898%2900023-X)Cited by:[§1](https://arxiv.org/html/2609.19878#S1.p3.1)\.
- Kanget al\.\(2026\)H\. Kang, Y\. Zhang, N\. L\. Kuang, Y\. Ma, and L\. QinLaDi\-RL: latent diffusion reasoning prevents entropy collapse in reinforcement learning\.arXiv preprint arXiv:2602\.01705\.Cited by:[§1](https://arxiv.org/html/2609.19878#S1.p5.1),[§3\.4](https://arxiv.org/html/2609.19878#S3.SS4.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Kanget al\.\(2025\)H\. Kang, Y\. Zhang, N\. L\. Kuang, N\. Majamaki, N\. Jaitly, Y\. Ma, and L\. QinLaDiR: latent diffusion enhances LLMs for text reasoning\.arXiv preprint arXiv:2510\.04573\.Cited by:[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px3.p1.1),[§1](https://arxiv.org/html/2609.19878#S1.p5.1),[§3\.4](https://arxiv.org/html/2609.19878#S3.SS4.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Kimet al\.\(2025\)M\. J\. Kim, C\. Finn, and P\. LiangFine\-tuning vision\-language\-action models: optimizing speed and success\.arXiv preprint arXiv:2502\.19645\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1)\.
- Kimet al\.\(2024\)M\. J\. Kim, K\. Pertsch, S\. Karamcheti, T\. Xiao, A\. Balakrishna, S\. Nair, R\. Rafailov, E\. Foster, G\. Lam, P\. Sanketi,et al\.OpenVLA: an open\-source vision\-language\-action model\.arXiv preprint arXiv:2406\.09246\.Cited by:[§1](https://arxiv.org/html/2609.19878#S1.p2.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Kohliet al\.\(2026\)H\. Kohliet al\.Loop, think, & generalize: implicit reasoning in recurrent\-depth transformers\.arXiv preprint arXiv:2604\.07822\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Lei and Shan \(2026\)D\. Lei and L\. ShanThink less, act early: reinforced latent reasoning with early exit in vision\-language\-action models\.arXiv preprint arXiv:2606\.15099\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1)\.
- Liet al\.\(2025a\)A\. Li, C\. Wang, D\. Fu, K\. Yue, Z\. Cai, W\. B\. Zhu, O\. Liu, P\. Guo, W\. Neiswanger, F\. Huang, T\. Goldstein, and M\. GoldblumZebra\-cot: a dataset for interleaved vision language reasoning\.arXiv preprint arXiv:2507\.16746\.Cited by:[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px2.p1.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px1.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Liet al\.\(2025b\)B\. Li, X\. Sun, J\. Liu, Z\. Wang, J\. Wu, X\. Yu, H\. Chen, E\. Barsoum, M\. Chen, and Z\. LiuLatent visual reasoning\.arXiv preprint arXiv:2509\.24251\.Cited by:[§1](https://arxiv.org/html/2609.19878#S1.p2.1),[§2](https://arxiv.org/html/2609.19878#S2.SS0.SSS0.Px2.p1.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Liet al\.\(2025c\)C\. Li, W\. Wu, H\. Zhang, Y\. Xia, S\. Mao, L\. Dong, I\. Vulić, and F\. WeiImagine while reasoning in space: multimodal visualization\-of\-thought\.arXiv preprint arXiv:2501\.07542\.Cited by:[§1](https://arxiv.org/html/2609.19878#S1.p2.1)\.
- Liet al\.\(2026a\)K\. Li, C\. Shang, L\. Karlinsky, R\. Feris, T\. Darrell, and R\. HerzigLatent implicit visual reasoning\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Liet al\.\(2026b\)P\. Li, Z\. Gao, L\. Zhang, M\. Huang, Y\. Li, Z\. Yang, F\. Xu, and J\. LiuVisual\-opsd: cross\-modal on\-policy self\-distillation for efficient unified multimodal reasoning\.arXiv preprint arXiv:2606\.18974\.Cited by:[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px6.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Liet al\.\(2024\)Q\. Li, Y\. Liang, Z\. Wang, L\. Luo, X\. Chen, M\. Liao, F\. Wei, Y\. Deng, S\. Xu, Y\. Zhang,et al\.CogACT: a foundational vision\-language\-action model for synergizing cognition and action in robotic manipulation\.arXiv preprint arXiv:2411\.19650\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Liet al\.\(2026c\)W\. Li, J\. Liu, L\. Yixing, J\. Tong, R\. Shao, and L\. NieConsisVLA\-4d: advancing spatiotemporal consistency in efficient 3d\-perception and 4d\-reasoning for robotic manipulation\.arXiv preprint arXiv:2605\.05126\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1)\.
- Liet al\.\(2023a\)X\. Li, M\. Zhang, Y\. Geng, H\. Geng, Y\. Long, Y\. Shen, R\. Zhang, J\. Liu, and H\. DongManipLLM: embodied multimodal large language model for object\-centric robotic manipulation\.arXiv preprint arXiv:2312\.16217\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1)\.
- Liet al\.\(2023b\)X\. Li, M\. Liu, H\. Zhang,et al\.Vision\-language foundation models as effective robot imitators\.arXiv preprint arXiv:2311\.01378\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Lipmanet al\.\(2023\)Y\. Lipman, R\. T\. Q\. Chen, H\. Ben\-Hamu, M\. Nickel, and M\. LeFlow matching for generative modeling\.InInternational Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=PqvMRDCJT9t)Cited by:[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px3.p2.1),[§3\.4](https://arxiv.org/html/2609.19878#S3.SS4.p2.1)\.
- Liuet al\.\(2023\)B\. Liu, Y\. Zhu, C\. Gao, Y\. Feng, Q\. Liu, Y\. Zhu, and P\. StoneLIBERO: benchmarking knowledge transfer for lifelong robot learning\.InAdvances in Neural Information Processing Systems,Vol\.36\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px2.p1.1)\.
- Liuet al\.\(2025\)J\. Liu, H\. Chen, P\. An, Z\. Liu, R\. Zhang, C\. Gu, X\. Li, Z\. Guo, S\. Chen, M\. Liu,et al\.HybridVLA: collaborative diffusion and autoregression in a unified vision\-language\-action model\.arXiv preprint arXiv:2503\.10631\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1)\.
- Liuet al\.\(2026\)Z\. Liu, J\. Liu, H\. Chen, J\. Yu, Z\. Guo, C\. Hou, C\. Gu, X\. Mi, R\. Zhang, K\. Wu, Z\. Che, J\. Tang, P\. Heng, and S\. ZhangLaST0\{\}\_\{0\}: latent spatio\-temporal chain\-of\-thought for robotic vision\-language\-action model\.arXiv preprint arXiv:2601\.05248\.Cited by:[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px4.p1.1),[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px5.p1.1),[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px6.p1.1),[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px7.p1.1),[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px8.p1.1),[§1](https://arxiv.org/html/2609.19878#S1.p2.1),[§2](https://arxiv.org/html/2609.19878#S2.SS0.SSS0.Px2.p1.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px2.p1.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Luet al\.\(2023\)J\. Lu, C\. Clark, S\. Lee, Z\. Zhang, S\. Khosla, R\. Marten, D\. Hoiem, and A\. KembhaviUnified\-IO 2: scaling autoregressive multimodal models with vision, language, audio, and action\.arXiv preprint arXiv:2312\.17172\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Luet al\.\(2022a\)J\. Lu, C\. Clark, R\. Zellers, R\. Mottaghi, and A\. KembhaviUnified\-io: a unified model for vision, language, and multi\-modal tasks\.arXiv preprint arXiv:2206\.08916\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Luet al\.\(2024\)P\. Lu, H\. Bansal, T\. Xia, J\. Liu, C\. Li, H\. Hajishirzi, H\. Cheng, K\. Chang, M\. Galley, and J\. GaoMathVista: evaluating mathematical reasoning of foundation models in visual contexts\.InInternational Conference on Learning Representations,Cited by:[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px1.p1.1),[§1](https://arxiv.org/html/2609.19878#S1.p1.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px1.p1.1)\.
- Luet al\.\(2022b\)P\. Lu, S\. Mishra, T\. Xia,et al\.Learn to explain: multimodal reasoning via thought chains for science question answering\.Advances in Neural Information Processing Systems\.Cited by:[§1](https://arxiv.org/html/2609.19878#S1.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Masryet al\.\(2022\)A\. Masry, D\. X\. Long, J\. Q\. Tan, S\. Joty, and E\. HoqueChartQA: a benchmark for question answering about charts with visual and logical reasoning\.InFindings of the Association for Computational Linguistics: ACL 2022,Cited by:[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px1.p1.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px1.p1.1)\.
- Mizrahiet al\.\(2023\)D\. Mizrahi, R\. Bachmann, O\. F\. Kar, T\. Yeo, M\. Gao, A\. Dehghan, and A\. Zamir4M: massively multimodal masked modeling\.arXiv preprint arXiv:2312\.06647\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Penget al\.\(2023\)Z\. Peng, W\. Wang, L\. Dong,et al\.Kosmos\-2: grounding multimodal large language models to the world\.arXiv preprint arXiv:2306\.14824\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Pfauet al\.\(2024\)J\. Pfau, W\. Merrill, and S\. R\. BowmanLet’s think dot by dot: hidden computation in transformer language models\.arXiv preprint arXiv:2404\.15758\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Pham and Ngo \(2025\)T\. Pham and C\. NgoMultimodal chain of continuous thought for latent\-space reasoning in vision\-language models\.arXiv preprint arXiv:2508\.12587\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Physical Intelligenceet al\.\(2025\)Physical Intelligence, K\. Black, N\. Brown, J\. Darpinian, K\. Dhabalia, D\. Driess, A\. Esmail, M\. Equi, C\. Finn,et al\.π0\.5\\pi\_\{0\.5\}: A vision\-language\-action model with open\-world generalization\.arXiv preprint arXiv:2504\.16054\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1)\.
- Qinet al\.\(2025\)Y\. Qin, B\. Wei, J\. Ge, K\. Kallidromitis, S\. Fu, T\. Darrell, and X\. WangChain\-of\-visual\-thought: teaching VLMs to see and think better with continuous visual tokens\.arXiv preprint arXiv:2511\.19418\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Quet al\.\(2025\)D\. Qu, H\. Song, Q\. Chen, Y\. Yao, X\. Ye, Y\. Ding, Z\. Wang, J\. Gu,et al\.SpatialVLA: exploring spatial representations for visual\-language\-action model\.arXiv preprint arXiv:2501\.15830\.Cited by:[§1](https://arxiv.org/html/2609.19878#S1.p2.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1)\.
- Rayet al\.\(2025\)A\. Ray, A\. Abdelkader, C\. Mao, B\. A\. Plummer, K\. Saenko, R\. Krishna, L\. Guibas, and W\. ChuMull\-tokens: modality\-agnostic latent thinking\.arXiv preprint arXiv:2512\.10941\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Rayet al\.\(2024\)A\. Ray, J\. Duan, E\. Brown, R\. Tan, D\. Bashkirova, R\. Hendrix, K\. Ehsani, A\. Kembhavi, B\. A\. Plummer, R\. Krishna, K\. Zeng, and K\. SaenkoSAT: dynamic spatial aptitude training for multimodal language models\.arXiv preprint arXiv:2412\.07755\.Cited by:[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px1.p1.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px1.p1.1)\.
- Reedet al\.\(2022\)S\. Reed, K\. Zolna, E\. Parisotto,et al\.A generalist agent\.Transactions on Machine Learning Research\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Rombachet al\.\(2022\)R\. Rombach, A\. Blattmann, D\. Lorenz, P\. Esser, and B\. OmmerHigh\-resolution image synthesis with latent diffusion models\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,External Links:[Link](https://arxiv.org/abs/2112.10752)Cited by:[§1](https://arxiv.org/html/2609.19878#S1.p5.1)\.
- Roseet al\.\(2023\)D\. Roseet al\.Visual chain of thought: bridging logical gaps with multimodal infillings\.arXiv preprint arXiv:2305\.02317\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Shenet al\.\(2025\)Z\. Shen, H\. Yan, L\. Zhang, Z\. Hu, Y\. Du, and Y\. HeCODI: compressing chain\-of\-thought into continuous space via self\-distillation\.arXiv preprint arXiv:2502\.21074\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Songet al\.\(2025\)Y\. Song, T\. Ou, Y\. Kong, Z\. Li, G\. Neubig, and X\. YueVisualPuzzles: decoupling multimodal reasoning evaluation from domain knowledge\.arXiv preprint arXiv:2504\.10342\.Cited by:[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px1.p1.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px1.p1.1)\.
- Sunet al\.\(2023a\)Q\. Sun, Y\. Cui, X\. Zhang,et al\.Emu2: generative multimodal models are in\-context learners\.arXiv preprint arXiv:2312\.13286\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Sunet al\.\(2023b\)Q\. Sun, Y\. Cui, X\. Zhang,et al\.Emu: generative pretraining in multimodality\.arXiv preprint arXiv:2307\.05222\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Sunet al\.\(2024\)Y\. Sun, H\. Bao, W\. Wang, Z\. Peng, L\. Dong, S\. Huang, J\. Wang, and F\. WeiMultimodal latent language modeling with next\-token diffusion\.arXiv preprint arXiv:2412\.08635\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Suriset al\.\(2023\)D\. Suris, S\. Menon, and C\. VondrickViperGPT: visual inference via python execution for reasoning\.arXiv preprint arXiv:2303\.08128\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Tanget al\.\(2023\)Z\. Tang, Z\. Yang, C\. Zhu, M\. Zeng, and M\. BansalAny\-to\-any generation via composable diffusion\.arXiv preprint arXiv:2305\.11846\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Tonget al\.\(2024a\)S\. Tong, E\. Brown, P\. Wu, S\. Woo, M\. Middepogu, S\. C\. Akula, J\. Yang, S\. Yang, A\. Iyer, X\. Pan, Z\. Wang, R\. Fergus, Y\. LeCun, and S\. XieCambrian\-1: a fully open, vision\-centric exploration of multimodal LLMs\.InAdvances in Neural Information Processing Systems,Cited by:[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px1.p1.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px1.p1.1)\.
- Tonget al\.\(2024b\)S\. Tong, Z\. Liu, Y\. Zhai, Y\. Ma, Y\. LeCun, and S\. XieEyes wide shut? exploring the visual shortcomings of multimodal LLMs\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,Cited by:[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px1.p1.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px1.p1.1)\.
- Viveiroset al\.\(2026\)A\. G\. Viveiros, N\. Gonçalves, M\. Lindemann, and A\. MartinsLanteRn: latent visual structured reasoning\.arXiv preprint arXiv:2603\.25629\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Viveiroset al\.\(2026\)A\. G\. Viveiroset al\.What is holding back latent visual reasoning?\.arXiv preprint arXiv:2605\.18445\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Wanget al\.\(2024\)B\. Wanget al\.Grokked transformers are implicit reasoners: a mechanistic journey to the edge of generalization\.arXiv preprint arXiv:2405\.15071\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Wanget al\.\(2025\)H\. Wang, A\. Su, W\. Ren, F\. Lin, and W\. ChenPixel reasoner: incentivizing pixel\-space reasoning with curiosity\-driven reinforcement learning\.arXiv preprint arXiv:2505\.15966\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1)\.
- Wanget al\.\(2024a\)K\. Wang, J\. Pan, W\. Shi, Z\. Lu, H\. Ren, A\. Zhou, M\. Zhan, and H\. LiMeasuring multimodal mathematical reasoning with the MATH\-Vision dataset\.InAdvances in Neural Information Processing Systems Datasets and Benchmarks Track,Cited by:[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px1.p1.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px1.p1.1)\.
- Wanget al\.\(2022\)P\. Wanget al\.OFA: unifying architectures, tasks, and modalities through a simple sequence\-to\-sequence learning framework\.arXiv preprint arXiv:2202\.03052\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Wanget al\.\(2026\)Q\. Wang, Y\. Shi, Y\. Wang, Y\. Zhang, P\. Wan, K\. Gai, X\. Ying, and Y\. WangMonet: reasoning in latent visual space beyond images and language\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Wanget al\.\(2024b\)X\. Wang, X\. Zhang, Z\. Luo,et al\.Emu3: next\-token prediction is all you need\.arXiv preprint arXiv:2409\.18869\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Weiet al\.\(2025\)Y\. Wei, L\. Zhao, J\. Sun, K\. Lin, J\. Yin, J\. Hu, Y\. Zhang, E\. Yu, H\. Lv, Z\. Weng, J\. Wang, C\. Han, Y\. Peng, Q\. Han, Z\. Ge, X\. Zhang, D\. Jiang, and V\. M\. PatelOpen vision reasoner: transferring linguistic cognitive behavior for visual reasoning\.arXiv preprint arXiv:2507\.05255\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1)\.
- Wuet al\.\(2023a\)C\. Wu, S\. Yin, W\. Qi,et al\.Visual chatgpt: talking, drawing and editing with visual foundation models\.arXiv preprint arXiv:2303\.04671\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Wuet al\.\(2024\)C\. Wu, X\. Chen, Z\. Wu,et al\.Janus: decoupling visual encoding for unified multimodal understanding and generation\.arXiv preprint arXiv:2410\.13848\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Wu and Xie \(2024\)P\. Wu and S\. XieV\*: guided visual search as a core mechanism in multimodal LLMs\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,Cited by:[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px1.p1.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px1.p1.1)\.
- Wuet al\.\(2023b\)S\. Wu, H\. Fei, L\. Qu, W\. Ji, and T\. ChuaNExT\-GPT: any\-to\-any multimodal llm\.arXiv preprint arXiv:2309\.05519\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Wuet al\.\(2026\)Y\. Wuet al\.Continuous reasoning for vision\-language\-action\.arXiv preprint arXiv:2606\.00229\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Xieet al\.\(2024\)J\. Xie, W\. Mao, Z\. Bai,et al\.Show\-o: one single transformer to unify multimodal understanding and generation\.arXiv preprint arXiv:2408\.12528\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Xuet al\.\(2024\)G\. Xuet al\.LLaVA\-CoT: let vision language models reason step\-by\-step\.arXiv preprint arXiv:2411\.10440\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Xuet al\.\(2025a\)J\. Xuet al\.Qwen2\.5\-Omni technical report\.arXiv preprint arXiv:2503\.20215\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Xuet al\.\(2026a\)T\. Xu, Y\. Sun, Q\. Wang, J\. Lu, Y\. Wang, T\. Zhang, L\. Guo, F\. Rao, J\. Lyu, F\. Chen, and J\. LiuSemantic\-enriched latent visual reasoning\.arXiv preprint arXiv:2605\.19342\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Xuet al\.\(2025\)W\. Xu, J\. Wang, W\. Wang, Z\. Chen, W\. Zhou, A\. Yang, L\. Lu, H\. Li, X\. Wang, X\. Zhu, W\. Wang, J\. Dai, and J\. ZhuVisuLogic: a benchmark for evaluating visual reasoning in multi\-modal large language models\.arXiv preprint arXiv:2504\.15279\.Cited by:[§B\.1](https://arxiv.org/html/2609.19878#A2.SS1.SSS0.Px1.p1.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px1.p1.1)\.
- Xuet al\.\(2026b\)X\. Xu, H\. Li, J\. Ye, Y\. Chen, J\. Zeng, X\. Chen, L\. Xu, D\. Lin, W\. Li, and J\. PangFuturevla: joint visuomotor prediction for vision\-language\-action model\.arXiv preprint arXiv:2603\.10712\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1)\.
- Xuet al\.\(2025b\)Y\. Xuet al\.SoftCoT\+\+: test\-time scaling with soft chain\-of\-thought reasoning\.arXiv preprint arXiv:2505\.11484\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Xuet al\.\(2025c\)Y\. Xuet al\.SoftCoT: soft chain\-of\-thought for efficient reasoning with llms\.arXiv preprint arXiv:2502\.12134\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Yang and Shan \(2026\)B\. Yang and L\. ShanPearlVLA: progressive embodied action\-plan refinement in latent space\.arXiv preprint arXiv:2606\.17924\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1)\.
- Yanget al\.\(2023\)L\. Yanget al\.Looped transformers are better at learning learning algorithms\.arXiv preprint arXiv:2311\.12424\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Yanget al\.\(2026\)X\. Yanget al\.Stabilizing recurrent dynamics for test\-time scalable latent reasoning in looped language models\.arXiv preprint arXiv:2605\.26733\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Yanget al\.\(2025\)Z\. Yang, X\. Yu, D\. Chen, M\. Shen, and C\. GanMachine mental imagery: empower multimodal reasoning with latent visual tokens\.arXiv preprint arXiv:2506\.17218\.Cited by:[§1](https://arxiv.org/html/2609.19878#S1.p2.1),[§2](https://arxiv.org/html/2609.19878#S2.SS0.SSS0.Px2.p1.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Yanget al\.\(2023\)Z\. Yang, L\. Li, K\. Lin,et al\.MM\-REACT: prompting chatgpt for multimodal reasoning and action\.arXiv preprint arXiv:2303\.11381\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Yaoet al\.\(2024\)H\. Yaoet al\.Mulberry: empowering mllm with o1\-like reasoning and reflection via collective monte carlo tree search\.arXiv preprint arXiv:2412\.18319\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Yuet al\.\(2023\)L\. Yu, B\. Shi, R\. Pasunuru,et al\.Scaling autoregressive multi\-modal models: pretraining and instruction tuning\.arXiv preprint arXiv:2309\.02591\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Yuet al\.\(2024\)P\. Yuet al\.Distilling system 2 into system 1\.arXiv preprint arXiv:2407\.06023\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Zelikmanet al\.\(2024\)E\. Zelikman, G\. Harik, Y\. Shao,et al\.Quiet\-star: language models can teach themselves to think before speaking\.arXiv preprint arXiv:2403\.09629\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Zelikmanet al\.\(2022\)E\. Zelikman, Y\. Wu, J\. Mu, and N\. D\. GoodmanSTaR: bootstrapping reasoning with reasoning\.Advances in Neural Information Processing Systems\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px3.p1.1)\.
- Zhanet al\.\(2024\)J\. Zhan, J\. Dai, J\. Ye,et al\.AnyGPT: unified multimodal llm with discrete sequence modeling\.arXiv preprint arXiv:2402\.12226\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Zhanget al\.\(2024\)Z\. Zhang, A\. Zhang, M\. Li, H\. Zhao, G\. Karypis, and A\. SmolaMultimodal chain\-of\-thought reasoning in language models\.Transactions on Machine Learning Research\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Zhaoet al\.\(2025\)Q\. Zhao, Y\. Lu, M\. J\. Kim, Z\. Fu, Z\. Zhang, Y\. Wu, Z\. Li, Q\. Ma, S\. Han, C\. Finn, A\. Handa, M\. Liu, D\. Xiang, G\. Wetzstein, and T\. LinCoT\-vla: visual chain\-of\-thought reasoning for vision\-language\-action models\.arXiv preprint arXiv:2503\.22020\.Cited by:[§1](https://arxiv.org/html/2609.19878#S1.p2.1),[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1),[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Zhenet al\.\(2024\)H\. Zhen, X\. Qiu, P\. Chen, J\. Yang, X\. Yan, Y\. Du, Y\. Hong, and C\. Gan3D\-VLA: a 3D vision\-language\-action generative world model\.arXiv preprint arXiv:2403\.09631\.Cited by:[§1](https://arxiv.org/html/2609.19878#S1.p2.1)\.
- Zhenget al\.\(2025\)Z\. Zheng, M\. Yang, J\. Hong, C\. Zhao, G\. Xu, L\. Yang, C\. Shen, and X\. YuDeepEyes: incentivizing thinking with images via reinforcement learning\.arXiv preprint arXiv:2505\.14362\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1)\.
- Zhonget al\.\(2026\)Z\. Zhong, J\. Li, J\. He, H\. Yan, X\. Gong, G\. Zhao, Y\. Cai, J\. Gao, X\. Yan, B\. Liu, Y\. Chen, L\. Yang, and H\. LiDualCoT\-VLA: visual\-linguistic chain of thought via parallel reasoning for vision\-language\-action models\.arXiv preprint arXiv:2603\.22280\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.
- Zhonget al\.\(2025\)Z\. Zhong, H\. Yan, J\. Li, X\. Liu, X\. Gong, W\. Song, J\. Chen, and H\. LiFlowVLA: visual chain of thought\-based motion reasoning for vision\-language\-action models\.arXiv preprint arXiv:2508\.18269\.Cited by:[§4\.1](https://arxiv.org/html/2609.19878#S4.SS1.SSS0.Px3.p1.1)\.
- Zhouet al\.\(2024\)C\. Zhou, L\. Yu, A\. Babu, K\. Tirumala, M\. Yasunaga, L\. Shamis, J\. Kahn, X\. Ma, L\. Zettlemoyer, and O\. LevyTransfusion: predict the next token and diffuse images with one multi\-modal model\.arXiv preprint arXiv:2408\.11039\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px1.p1.1)\.
- Zhouet al\.\(2024\)Q\. Zhouet al\.Image\-of\-thought prompting for visual reasoning refinement in multimodal large language models\.arXiv preprint arXiv:2405\.13872\.Cited by:[§5](https://arxiv.org/html/2609.19878#S5.SS0.SSS0.Px2.p1.1)\.

## Appendix AMethod Details

### A\.1Attention masks for parallel block training

Figure[7](https://arxiv.org/html/2609.19878#A1.F7)illustrates the three masks forKKreasoning steps packed into each pass; ellipses indicate omitted intermediate blocks\. The restrictions apply at every layer, preventing masked information from reaching predictions through other hidden states\.

Figure 7:Attention masks for parallel block training\.Rows attend to columns; each cell represents a token block\. Grounding interleaves clean thought blocks with target streams; diffusion interleaves noisy and clean thought blocks\. Clean blocks cannot read target or noisy streams\.#### Construction\.

The packed sequence contains pairs\[si;q\]\[s\_\{i\};q\]for all teacher steps\. Attention is isolated within each pair: the learnable block reads its own teacher features and itself, but neither the task input nor other pairs\.

#### Grounding\.

The packed sequence is\[x;z1⋆;s2;z2⋆;s3;…;zK⋆;y\]\[x;z\_\{1\}^\{\\star\};s\_\{2\};z\_\{2\}^\{\\star\};s\_\{3\};\\ldots;z\_\{K\}^\{\\star\};y\], wheresis\_\{i\}andyylabel target prediction streams\. A prediction ofsiGTs\_\{i\}^\{\\mathrm\{GT\}\}can readxxandz<i⋆z\_\{<i\}^\{\\star\}; the final\-output prediction can readxxand allz1:K⋆z\_\{1:K\}^\{\\star\}\. The conditioning stream preserves this prefix order and cannot read the target streams\. Different target streams are isolated from one another\. Within a text target, standard causal attention permits earlier tokens of that target; continuous predictions cannot read their ground\-truth values\.

#### Diffusion\.

The packed sequence is\[x;z1t;z1⋆;z2t;z2⋆;…;zKt;zK⋆\]\[x;z\_\{1\}^\{t\};z\_\{1\}^\{\\star\};z\_\{2\}^\{t\};z\_\{2\}^\{\\star\};\\ldots;z\_\{K\}^\{t\};z\_\{K\}^\{\\star\}\]\. Each noisy blockzitz\_\{i\}^\{t\}can readxx, the clean prefixz<i⋆z\_\{<i\}^\{\\star\}, and all tokens within itself\. It cannot read its own clean target, later clean blocks, or other noisy blocks\. The clean conditioning stream follows block\-causal attention and cannot read noisy blocks\.

## Appendix BAdditional Experiments

### B\.1Experimental settings

#### Benchmarks and scoring\.

All methods in Tables[1](https://arxiv.org/html/2609.19878#S4.T1)and[5](https://arxiv.org/html/2609.19878#A2.T5)use the same evaluation examples within each benchmark\. Table[4](https://arxiv.org/html/2609.19878#A2.T4)lists the evaluation splits, sample counts, and scoring units for the eleven visual benchmarks\([Song et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib28);[Masry et al\., 2022](https://arxiv.org/html/2609.19878#bib.bib22);[Wu and Xie, 2024](https://arxiv.org/html/2609.19878#bib.bib23);[Fu et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib24);[Tong et al\., 2024b](https://arxiv.org/html/2609.19878#bib.bib25);[Ray et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib26);[Tong et al\., 2024a](https://arxiv.org/html/2609.19878#bib.bib27);[Lu et al\., 2024](https://arxiv.org/html/2609.19878#bib.bib48);[Wang et al\., 2024a](https://arxiv.org/html/2609.19878#bib.bib49);[Xu et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib50);[Hao et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib51)\)\. VisualPuzzles and VisuLogic use their released evaluation collections, even when the hosting service names the splittrain\. VisuLogic evaluation uses the original 1,000\-question benchmark\([Xu et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib50)\), rather than the current 1,003\-row hosted collection\.

Table 4:Shared visual evaluation sets\.Splits, example counts, and scoring units used for the benchmark comparisons\.
#### VLM training\.

We train on Zebra\-CoT\([Li et al\., 2025a](https://arxiv.org/html/2609.19878#bib.bib15)\), using its interleaved text and image steps as teacher supervision\. We use AdamW with a peak learning rate of10−510^\{\-5\}, cosine decay, a global batch size of 64, and one epoch of training, informed by prior multimodal fine\-tuning recipes\([Li et al\., 2025a](https://arxiv.org/html/2609.19878#bib.bib15);[Hu et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib9)\)\. We use bf16, a maximum sequence length of 8,192, 3% warmup, weight decay of 0\.01, and gradient clipping at 1\.0\. Each backbone retains its native image processor\. The unified encoder and shared reasoning backbone are optimized jointly; teacher traces are omitted at evaluation\.

#### Thought tokens and latent sampling\.

We use four 512\-dimensional thought tokens per block, with learned projections to and from the backbone width, and loss weightsλy=1\\lambda\_\{y\}=1andλ=5\\lambda=5, taking the token budget and weights from a latent\-diffusion reference configuration\([Kang et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib10)\)\. Training retains the annotated reasoning\-step sequence; the inference budget is four blocks, or 16 thought tokens, followed by greedy answer decoding with a 512\-token limit for VLMs\. The VLA reasoning horizon is four steps\.

For flow matching, we drawϵ∼𝒩⁡\(0,I\)\\epsilon\\sim\\mathcal\{N\}\(0,I\)andt∼𝒰⁡\[0,1\]t\\sim\\mathcal\{U\}\[0,1\], setzit=\(1−t\)​ϵ\+t​zi⋆z\_\{i\}^\{t\}=\(1\-t\)\\epsilon\+tz\_\{i\}^\{\\star\}, and usevi⋆=zi⋆−ϵv\_\{i\}^\{\\star\}=z\_\{i\}^\{\\star\}\-\\epsilonas the velocity target\([Lipman et al\., 2023](https://arxiv.org/html/2609.19878#bib.bib130)\)\. At inference, we integrate fromt=0t=0to11per block with one sampled trajectory and no classifier\-free guidance\.

#### RLBench\.

We evaluate ten tasks with a Franka Panda in CoppeliaSim, using 100 waypoint\-generated training demonstrations per task and keyframe sampling\([Liu et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib14)\)\. Inputs include a front\-view RGB image resized to384×384384\\times 384, a point cloud sampled to 1,024 points, the task instruction, and proprioception\. We evaluate the final checkpoint with 20 rollouts per task and seed, giving 600 rollouts across three seeds\.

#### LIBERO\.

We train a separate policy for each of the four suites, each containing ten tasks and 500 training demonstrations\. Inputs use third\-person and wrist RGB views resized to384×384384\\times 384\. We omit the point\-cloud teacher because this modality is unavailable, and sample intermediate teacher steps every eight frames\([Liu et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib14)\)\. We evaluate the final checkpoint with 50 rollouts per task and seed, giving 6,000 rollouts across the four suites and three seeds\. Both benchmarks use their native success conditions\.

#### Seeds and aggregation\.

Our experiments use three random seeds\. For question\-level accuracy or rollout success onNNexamples per seed, the mean iss¯=1003​N​∑r=13∑j=1Ncr​j\\bar\{s\}=\\frac\{100\}\{3N\}\\sum\_\{r=1\}^\{3\}\\sum\_\{j=1\}^\{N\}c\_\{rj\}, wherecr​j∈\{0,1\}c\_\{rj\}\\in\\\{0,1\\\}indicates correctness or success\. Grouped metrics retain the benchmark’s aggregation rule: CV\-Bench averages its 2D and 3D scores\. For MMVP, we report question\-level accuracy over 300 questions, matching the VLMEvalKit protocol used by the relevant comparison methods\([Gu et al\., 2025](https://arxiv.org/html/2609.19878#bib.bib16);[Li et al\., 2026b](https://arxiv.org/html/2609.19878#bib.bib18)\)\. Each answer is scored independently against its correct option; the reported result is the mean of the three seed\-level accuracies\. Benchmark averages give equal weight to each reported benchmark or suite\. Relative gains are100​\(sours/sbaseline−1\)100\(s\_\{\\rm ours\}/s\_\{\\rm baseline\}\-1\); rounding follows aggregation\. RLBench Table[2](https://arxiv.org/html/2609.19878#S4.T2)retains the baseline paper’s variance convention\([Liu et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib14)\)\. Reported gains describe differences in mean scores; they do not denote statistical significance\.

#### VLA training\.

We train for 300 epochs with AdamW, using a peak learning rate of10−410^\{\-4\}, cosine decay to zero, no weight decay or warmup, and gradient clipping at 1\.0\([Liu et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib14)\)\. Training uses eight NVIDIA A800 GPUs with bf16 and DeepSpeed ZeRO\-1\. The batch size is eight per device and 64 globally, without gradient accumulation\. Actions are seven\-dimensional, with ten flow\-matching steps in the action head and a four\-step reasoning horizon\. We mix reasoning\-to\-action ratios of 1:1, 1:2, and 1:4 during training and use 1:4 at evaluation\. The latent diffusion objective is defined in Section[3](https://arxiv.org/html/2609.19878#S3)\.

#### Inference speed\.

All inference rates in Table[2](https://arxiv.org/html/2609.19878#S4.T2), including those of the baselines, are measured on the same single NVIDIA RTX 4090, without action chunking, using one reasoning update per four action decisions\([Liu et al\., 2026](https://arxiv.org/html/2609.19878#bib.bib14)\)\. We use batch size one and bf16, with 100 warmup decisions followed by 1,000 timed decisions per seed\. Wall\-clock timing is synchronized with CUDA before and after the timed sequence and includes observation preprocessing, host\-to\-device transfer, reasoning updates, and action prediction; simulator stepping and rendering are excluded\. Each method uses the same inputs and timing boundary\. Throughput is the number of decisions divided by elapsed time, averaged across three seeds\. The reported 16\.5 Hz corresponds to 60\.6 ms per decision\.

### B\.2Ablation protocols

#### Modality grouping\.

Figure[3](https://arxiv.org/html/2609.19878#S4.F3)varies how many teacher modalities share one latent space\. Every variant encodes each modality with the same learned encoder, which reuses the reasoning backbone’s weights \(Section[3\.4](https://arxiv.org/html/2609.19878#S3.SS4)\); what changes is only the latent interface—the query tokensqqand output projection—that defines each space\. In the separate variant, every modality has its own query tokens and projection, so each is encoded into a private latent target space\. For each RLBench pairwise variant, the named pair shares one set of query tokens and one projection, forming a common space, while the remaining modality keeps its own; full sharing maps all three modalities into a single shared space\. The visual\-reasoning comparison likewise gives text and image either separate spaces or one shared space\. All variants retain every teacher modality, the reasoning backbone, step order, latent\-token budget, and training schedule\. Because the learned encoder and its capacity are identical across variants, this study isolates the effect of unifying the latent interface rather than of introducing learned compression\.

#### Grounding objective\.

In Figure[4](https://arxiv.org/html/2609.19878#S4.F4), reconstruction predicts the source teacher features from each thought block; downstream grounding predicts later teacher steps\. Both variants retain the same final\-output lossλy​ℓ​\(y^,y\)\\lambda\_\{y\}\\ell\(\\hat\{y\},y\)and propagate its gradients through the thought tokens into the unified encoder and shared backbone\. Text CE is averaged over supervised tokens, and continuous L2 loss is averaged over target elements, using the same reduction and loss weights in both variants\. Teacher targets remain fixed\. Thus, the comparison changes intermediate supervision while preserving the encoder’s access to final\-task supervision\.

#### Thought\-token prediction\.

Figure[6](https://arxiv.org/html/2609.19878#S4.F6)uses fixed clean thought\-token targets for all three objectives\. L2 and cosine regression directly predict the next clean thought block\. Flow matching predicts the velocity along the noisy\-to\-clean path, using the sampler specified in Appendix[B\.1](https://arxiv.org/html/2609.19878#A2.SS1)\. Each method conditions on the same input and preceding thought blocks and supplies its predicted tokens to the downstream reasoner\. This comparison reports downstream performance for the respective prediction procedures\.

### B\.3Additional results

#### Reasoning across backbones\.

We evaluate Uni\-LaDiR with BAGEL\-7B, Qwen3\-VL\-8B, and Janus\-Pro at 1\.5B and 7B to test whether the gains extend beyond the main backbone \(Table[5](https://arxiv.org/html/2609.19878#A2.T5)\)\. On the seven vision\-centric benchmarks, the average improves over the corresponding base models by29\.5%,12\.4%,35\.9%, and51\.5%, respectively\. With BAGEL, Uni\-LaDiR also improves the average over Visual\-OPSD by2\.5%, leading on six of seven benchmarks; with Qwen3\-VL, it improves over CoLT by13\.3%\. On BAGEL’s mathematical and logical benchmarks, the average gains are27\.9%over the base model and20\.0%over Uni\-CoT, including a28\.3%gain on EMMA over Uni\-CoT\. These improvements across model families and sizes suggest that shared latent reasoning is useful beyond a single backbone\.

Table 5:Reasoning across backbones\.Accuracy \(%\) on\(a\)seven vision\-centric benchmarks and\(b\)four mathematical and logical benchmarks\. Bold marks the best result within each backbone\.

#### Qualitative manipulation comparison\.

Figure[8](https://arxiv.org/html/2609.19878#A2.F8)compares a LaST0rollout with a successful Uni\-LaDiR rollout from the same initial scene\. LaST0approaches the phone but does not secure a grasp within ten decisions\. Uni\-LaDiR grasps the phone, transfers it to the base, and releases it, reaching success at7\.057\.05seconds of simulation time\. This selected episode illustrates a difference at the grasping stage; aggregate success rates are reported in Table[2](https://arxiv.org/html/2609.19878#S4.T2)\.

![Refer to caption](https://arxiv.org/html/2609.19878v1/rlbench-uniladir-aligned-v7.png)Figure 8:Phone placement in RLBench\.LaST0does not grasp the phone within ten recorded decisions \(top\), while Uni\-LaDiR completes placement \(bottom\)\. Columns share simulation timestamps, excluding inference latency; the final frame is held after each recording ends\.