# ThinkOmni: Lifting Textual Reasoning to Omni-modal Scenarios via Guidance Decoding
**Authors**: yiranguan, dkliang, xbai}@hust.edu.cn
## Abstract
Omni-modal reasoning is essential for intelligent systems to understand and draw inferences from diverse data sources. While existing Omni-modal Large Language Models (OLLM) excel at perceiving diverse modalities, they lack the complex reasoning abilities of recent Large Reasoning Models (LRM). However, enhancing the reasoning ability of OLLMs through additional training presents significant challenges, including the need for high-quality data, task-specific adaptation, and substantial computational costs. To address these limitations, we propose ThinkOmni, a training-free framework that lifts textual reasoning to omni-modal scenarios. ThinkOmni introduces two key components: 1) LRM-as-a-Guide, which leverages off-the-shelf LRMs to guide the OLLM decoding process; 2) Stepwise Contrastive Scaling, which adaptively balances perception and reasoning signals without manual hyperparameter tuning. Experiments on six multi-modality reasoning benchmarks demonstrate that ThinkOmni consistently delivers performance improvements, with main results achieving $70.2\$ on MathVista and $75.5\$ on MMAU. Overall, ThinkOmni offers a flexible and generalizable solution for omni-modal reasoning and provides new insights into the generalization and application of reasoning capabilities. Project page: https://1ranguan.github.io/thinkomni
footnotetext: Work done at Xiaomi Inc. đ Corresponding author.
## 1 Introduction
The emergence of Large Reasoning Models (LRMs) marks a paradigm shift from the traditional fast thinking of standard LLMs, which rely on immediate intuition, to slow thinking, which emphasizes reflective and iterative reasoning. Recent LRMs, such as DeepSeek-R1 (Guo et al., 2025) and o1 (OpenAI, 2025), have demonstrated exceptional performance in specialized reasoning tasks like mathematical problem-solving and code generation. Nonetheless, their effectiveness remains predominantly constrained to textual inputs, thus limiting their applicability to more complex, omni-modal real-world scenarios involving text, audio, images, and videos (see Fig. 1 (a)).
Omni-modal reasoning is essential for synthesizing diverse data sources and enabling sophisticated inference in context-rich tasks. Strong omni-modal reasoning capabilities have profound implications for practical applications such as advanced virtual assistants (Zhang et al., 2025) and embodied robots (Gan et al., 2020). Although recent advances in Omni-modal Large Language Models (OLLM) (Xu et al., 2025; Li et al., 2025b; Liu et al., 2025c; Fu et al., 2025; Luo et al., 2025) have shown promise in comprehending various input modalities, these models typically fall short when tasked with intricate reasoning across modalities, as illustrated in Fig. 1 (b). Therefore, a fundamental research challenge is how to effectively extend and elevate the reasoning capabilities of models from primarily textual inputs to truly omni-modal scenarios. *[Error downloading image: 2602.23306v1/x1.png]*
Figure 1: Comparison between existing methods and ThinkOmni. We integrate an OLLM with an LRM via guidance decoding, enabling advanced reasoning abilities with omni-modal input.
Actually, this is not a trivial problem, and despite considerable efforts, existing approaches to omni-modal reasoning are still limited in several critical aspects. Specifically, 1) Insufficient modality diversity. Current studies largely focus on specific modalities (e.g., image (Liu et al., 2025b; a; Lin et al., 2025), audio (Li et al., 2025a), or video (Wang et al., 2025)), rather than generalizing across arbitrary combinations of modalities. 2) Task-specific enhancement. Enhancements proposed for existing OLLMs (Zhao et al., 2025; Zhong et al., 2025; Rouditchenko et al., 2025; Yang et al., 2025b) remain confined to particular downstream tasks, lacking broader generalizability. 3) Data scarcity and high training costs. Current methods predominantly rely on extensive supervised finetuning (SFT) (Xu et al., 2024; Yang et al., 2025c) (requiring tens of thousands of reasoning examples) or reinforcement finetuning (RFT) (Shao et al., 2024; Yu et al., 2025) approaches demanding training computational resources (e.g., $8\times$ 40G VRAM for 7B model, $16\times$ 80G VRAM for 32B model). These challenges collectively motivate an important question: Is it possible to overcome the constraints of data and training conditions to bring general reasoning abilities to omni-modal content?
In this paper, we propose ThinkOmni, a novel training-free framework designed to lift textual reasoning to omni-modal scenarios (see Fig. 1 (d)). Unlike existing approaches (see Fig. 1 (c)) reliant on costly data annotation or additional model training, ThinkOmni directly leverages off-the-shelf LRMs as decoding-time guides for OLLMs. Specifically, we first introduce the LRM-as-a-Guide strategy, enabling the integration of reasoning capabilities from LRMs into OLLMs. We further identify a potential issue: a fixed guidance weight is unsuitable for all the tasks, and manual, task-specific adjustment is impractical. To resolve this, we propose a Stepwise Contrastive Scaling module, adaptively balancing perceptual and reasoning signals based on real-time analysis of model predictions. This module adapts to various task types and facilitates coherent omni-modal reasoning. *[Error downloading image: 2602.23306v1/x2.png]*
Figure 2: Performance comparison.
Extensive experiments conducted on six challenging multi-modal reasoning benchmarks demonstrate the effectiveness of our method. Specifically, our method improves the state-of-the-art open source OLLM Qwen2.5-Omni (Xu et al., 2025) by substantial margins without additional training, as shown in Fig. 2, rivaling or surpassing models that undergo extensive RFT. Additionally, compared to other guidance decoding algorithms (Li et al., 2022; Liu et al., 2024), our method reduces the burden of multi-modal data input, thereby maintaining decoding efficiency.
ThinkOmni provides a flexible framework for lifting textual reasoning to a more diverse and enriched input space. By leveraging the strengths of OLLM and LRM, we explore the effective generalization of reasoning capabilities to omni-modal scenarios in a training-free manner. Besides, our method is not limited to current LRMs. As new LLM technologies emerge (often developing faster than multi-modal variants), our approach can be easily adapted to improve performance across multi-modal variants and other downstream domains.
## 2 Preliminaries
### 2.1 Next Token Prediction
Given an omni-modal input $\mathit{O}$ (e.g., images, audios, videos) and a sequence of text tokens $x_{<t}=(x_{1},x_{2},\ldots,x_{t-1})$ , the OLLM $\mathit{M}$ first computes the logits $z_{t}$ for the next token $x_{t}$ as $z_{t}=\mathit{M}(x_{<t},\mathit{O})$ , where $z_{t}\in\mathbb{R}^{V}$ and $V$ is the vocabulary size. The probability distribution $P$ for $x_{t}$ is then given by
$$
P(x_{t}\mid x_{<t},\mathit{O})=\mathrm{Softmax}(z_{t}). \tag{1}
$$
Then $x_{<t+1}=(x_{1},x_{2},\ldots,x_{t})$ . The model computes a distribution and decodes a token at each step, resulting in an auto-regressive generation process.
### 2.2 Inference-time Guidance Decoding
Finetuning large language models is time-consuming and costly, highlighting the need for methods to modify or control modelsâ behaviors without additional training. In this subsection, we introduce the following works to understand our method better: Contrastive Decoding (Li et al., 2022), Visual Contrastive Decoding (Leng et al., 2024), ProxyTuning (Liu et al., 2024), and ProxyThinker (Xiao et al., 2026). For models within the same family (i.e., sharing the same token vocabulary), these methods guide base model decoding by introducing a contrastive pair at the logits level:
$$
\hat{z}=z^{\mathrm{base}}+\alpha\cdot\underbrace{(z^{+}-z^{-})}_{\mathrm{contrastive~pair}}, \tag{2}
$$
where $\alpha$ controls the influence of the guidance signal. Here, $z^{+}$ and $z^{-}$ represent the logits from the positive and negative references, respectively. These encourage or discourage certain behaviors in the modelâs output. This mechanism is analogous to a differential amplifier circuit, which amplifies the desired signals while suppressing noise. Consequently, the model can reduce hallucinations or achieve preference alignment during inference without additional training. *[Error downloading image: 2602.23306v1/x3.png]*
Figure 3: Guidance decoding methods. âGuid.â denotes the guiding model, and âAmat.â denotes the amateur model.
In Contrastive Decoding (Fig. 3 (a)), the contrastive pair is formed by comparing the responses to the same prompt from the original guiding model and an additional amateur model, with $z^{+}$ set to $z^{\mathrm{base}}$ . In Visual Contrastive Decoding (Fig. 3 (b)), the contrastive pair is created by applying different input conditions to the same model. Specifically, $z^{-}$ is obtained by adding Gaussian noise to the input image and then performing inference. In contrast to these approaches, ProxyTuning and ProxyThinker (Fig. 3 (c)) construct contrastive pairs across different models within the same family, aiming to transfer behaviors from more minor, guiding models to larger, amateur models.
Existing guidance decoding methods are limited to scenarios with consistent input modalities and available expert models. There is often no suitable expert model in omni-modal settings or other downstream tasks, making it technically challenging to construct effective guidance signals. Moreover, the heterogeneity of modalities complicates the alignment and integration of guidance during inference. Our work addresses these challenges by designing a framework for cross-modal guidance decoding, enabling preference alignment without requiring modality-specific expert models. *[Error downloading image: 2602.23306v1/x4.png]*
Figure 4: Overview of ThinkOmni. The framework begins by separating input modalities of the OLLM and introducing the LRM as a guiding model. Stepwise Contrastive Scaling dynamically adjusts guidance parameters based on real-time prediction analysis, enabling adaptive and effective decoding across diverse tasks.
## 3 Method
This section outlines the implementation roadmap of ThinkOmni, starting with a straightforward guidance decoding approach. We first introduce LRM-as-a-Guide, which separates the input modalities of the Omni-modal Large Language Model (OLLM) and incorporates an off-the-shelf Large Reasoning Model (LRM) as a guiding component. While this approach is practical, coordinating fixed guidance decoding hyperparameters remains challenging due to the varying demands for reasoning signals across different tasks and scenarios. To address this shortcoming, we propose Stepwise Contrastive Scaling, a module that dynamically adjusts parameters based on real-time analysis of model predictions, thereby adapting automatically to each decoding scenario. An overview of our framework is provided in Fig. 4.
### 3.1 LRM-as-a-Guide
As discussed in Sec. 2.2, to address the gap where current guidance decoding approaches are limited to models with matched input modalities, we introduce the LRM-as-a-Guide, which lifts advanced textual reasoning into the omni-modal content through collaborative decoding with OLLM.
Let $\mathit{M}_{\mathit{O}}$ denote the OLLM and $\mathit{M}_{\mathit{R}}$ denote the LRM. As shown in Fig. 6 (a), we compute the base logits with full omni-modal input, $z^{\mathrm{base}}=\mathit{M}_{\mathit{O}}(x_{<t},\mathit{O})$ . Then we discard the omni-modal content and feed $\mathit{M}_{\mathit{O}}$ only the textual prefix $x_{<t}$ . The results are treated as the negative logits $z^{-}=\mathit{M}_{\mathit{O}}(x_{<t})$ . The positive logits are produced by the LRM on the same prefix $z^{+}=\mathit{M}_{\mathit{R}}(x_{<t})$ . As formulated in Eq. (2), the token probability distribution will then serve as
$$
\hat{P}=\mathrm{Softmax}\Big[\,\mathit{M}_{\mathit{O}}(x_{<t},\mathit{O})+\alpha\cdot\big(\mathit{M}_{\mathit{R}}(x_{<t})-\mathit{M}_{\mathit{O}}(x_{<t})\big)\,\Big], \tag{3}
$$
where the scalar $\alpha$ determines the extent to which the LRM influences the OLLM. After obtaining the mixed logits, we normalize them to probabilities and then sample the next token as usual.
Although the LRM cannot access omni-modal information, we mitigate this disadvantage and amplify the reasoning preference through the logits contrastive. During the generation process, the OLLM and LRM collaborate in a complementary manner. The OLLM, serving as the primary agent, extracts and integrates omni-modal clues, while the LRM provides deeper reasoning over the textual trace. As decoding progresses, the LRM can compensate for the lack of omni-modal information by leveraging the already decoded tokens, and the OLLM achieves logical reasoning through the reasoning preferences supplied by the LRM. Their strengths are seamlessly fused through logit mixing, resulting in a unified decoding framework that effectively integrates perception and reasoning.
### 3.2 Stepwise Contrastive Scaling
*[Error downloading image: 2602.23306v1/x5.png]*
Figure 5: Case studies from (a) MMAU and (b) MathVista. (Sakshi et al., 2025; Lu et al., 2024) Tasks require different levels of LRM involvement. Using a fixed $\alpha$ limits the ability of the model to optimally adapt to task-specific needs, highlighting the need for a more flexible approach.
While LRM-as-a-Guide effectively enables collaboration between the LRM and OLLM, there remains room for improvement regarding the choice of the fixed guidance weight $\alpha$ . A fixed $\alpha$ may not consistently achieve the optimal balance between perception and reasoning across different tasks. The OLLM prefers a smaller $\alpha$ to emphasize omni-modal cues, while the LRM benefits from a larger $\alpha$ to strengthen its guidance. Additionally, since $z^{+}$ and $z^{-}$ do not possess comprehensive omni-modal content, excessive reliance on them (adopting a large $\alpha$ ) can lead to recognition bias such as hallucination (Fig. 5 (a)). Conversely, setting $\alpha$ too low may diminish the effectiveness of guidance, thereby constraining the logical reasoning capabilities (see Fig. 5 (b)). Motivated by this, we propose Stepwise Contrastive Scaling, which dynamically apportions a tokenâs prediction budget between perception and reasoning through online analysis of logits. *[Error downloading image: 2602.23306v1/x6.png]*
Figure 6: Detailed process of ThinkOmni. (a) The OLLM handles multi-modal inputs, while the LRM focuses on textual reasoning. By mixing their logits, the system effectively integrates perception and reasoning during token generation. (b) Each decoding step balances perception and reasoning dynamically by comparing logit distributions under different conditions and models.
We introduce a stepwise influence metric to determine whether each decoding step is dominated by perception or reasoning. Specifically, all the generated logits are first transformed into probability distributions with a softmax function, let $P_{\mathit{O}}$ , $P_{\mathit{R}}$ , $P$ denote the corresponding distributions for $\mathit{M}_{\mathit{O}}(x_{<t},\mathit{O})$ , $\mathit{M}_{\mathit{R}}(x_{<t})$ , and $\mathit{M}_{\mathit{O}}(x_{<t})$ , respectively. The pairwise distances between these distributions are then quantified by the JensenâShannon divergence, which is employed in DoLa (Chuang et al., 2024) to measure the disagreement between two logits. This metric is symmetric, bounded, and numerically stable, making it well-suited for our purposes:
$$
D_{R}=\operatorname{JS}\bigl(P_{\mathit{R}}\parallel P\bigr),\qquad D_{P}=\operatorname{JS}\bigl(P_{\mathit{O}}\parallel P\bigr). \tag{4}
$$
Intuitively, $D_{R}$ reflects the unique influence of reasoning preference, whereas $D_{P}$ captures the contribution from perceptual omni-modalities. A larger pairwise distance signifies that the corresponding factor (perception or reasoning) impacts the current decoding step more. Building on this metric, we proceed to reformulate Eq. (3) and introduce an additional contrastive logits term:
$$
\hat{P}=\mathrm{Softmax}\Big[\,\mathit{M}_{\mathit{O}}(x_{<t},\mathit{O})+\alpha^{r}_{t}\cdot\big(\mathit{M}_{\mathit{R}}(x_{<t})-\mathit{M}_{\mathit{O}}(x_{<t})\big)+\alpha^{p}_{t}\cdot\big(\mathit{M}_{\mathit{O}}(x_{<t},\mathit{O})-\mathit{M}_{\mathit{O}}(x_{<t})\big)\Big], \tag{5}
$$
where $\alpha^{r}_{t}$ acts as the original guidance weight, capturing enhanced reasoning capability, whereas the difference contributed by $\alpha^{p}_{t}$ serves as an aggressive visual contrastive term (Leng et al., 2024) (i.e., by directly removing non-textual inputs rather than adding noise), reflecting augmented perceptual capability. To improve decoding stability, we employ a normalization strategy to ensure $\alpha^{r}_{t}+\alpha^{p}_{t}=1$ , with the coefficients determined by the relative magnitudes of $D_{R}$ and $D_{P}$ . During the initial decoding steps, we constrain the magnitude of $\alpha_{r}$ to implement a warmup for the reasoning task. More implementation details are provided in Appendix B.
As shown in Fig. 4, the entire ThinkOmni procedure is training-free, requiring no additional finetuning or corpus statistics. Leveraging stepwise contrastive scaling, LRM-as-a-Guide can autonomously evaluate the relative contributions of perceptual and reasoning signals at each generation step, seamlessly balancing these complementary abilities without manual hyperparameter tuning.
## 4 Experiment
### 4.1 Experiment Setup
#### Models
To validate the effectiveness of ThinkOmni, we conduct experiments on three OLLMs: Qwen2.5-Omni-3B / 7B (Xu et al., 2025) and Omni-R1 (Zhong et al., 2025). We utilize the DeepSeek-R1-Distill series (Guo et al., 2025) and the Qwen3 series (Yang et al., 2025a), both in thinking mode, as our LRMs to guide decoding.
#### Benchmarks
To demonstrate the generalizability of ThinkOmni, we evaluate it on omni-modal scenarios using six benchmarks, comprising over $10,000$ test samples in total: MathVista (test-mini) (Lu et al., 2024), MathVision (Wang et al., 2024), MathVerse (test-mini) (Zhang et al., 2024), MMAU-v05.15.25 (test-mini) (Sakshi et al., 2025), Daily-Omni (Zhou et al., 2025), OmniBench (Li et al., 2024). More details are provided in Appendix B.
#### Evaluation
We first use template matching for multiple-choice questions to extract the option from the modelâs output. If the answer cannot be extracted directly, we use GPT-4o to extract it and then compare the extracted answer to the gold answer. For free-form questions, we first use GPT-4o to extract the answer from the modelâs output, then compare the extracted answer to the gold answer to determine if their meanings are consistent. This process is designed to account for various expressions in the answers.
### 4.2 Main Result
Table 1: Model performance on several omni-modal reasoning benchmarks. Here, DeepSeek refers to DeepSeek-R1-Distill-Qwen-7B (Guo et al., 2025), and Qwen3 denotes Qwen3-8B (Yang et al., 2025a). The numbers in parentheses indicate the performance changes compared to the base OLLMs Qwen2.5-Omni-3B / 7B (Xu et al., 2025) and Omni-R1 (Zhong et al., 2025). Models marked with â â â are evaluated using our own evaluation scripts.
| Model | MathVista test-mini | MathVision test | MathVerse test-mini | MMAU (v05.15.25) test-mini | DailyOmni test | OmniBench test |
| --- | --- | --- | --- | --- | --- | --- |
| Close-Sourse Models | | | | | | |
| GPT-4o | 63.8 | 30.4 | 50.8 | 62.5 | 56.5 | - |
| Gemini-2.0-Flash | 73.1 | 41.3 | 59.3 | 70.5 | 67.8 | - |
| Open-Sourse Omni Models | | | | | | |
| Baichuan-Omni-1.5 | 63.6 | - | - | 66.2 | 50.0 | 42.9 |
| Ola | 68.4 | - | - | 70.3 | 50.71 | - |
| Open-Sourse RFT Omni Models | | | | | | |
| Omni-R1 â | 64.7 | 25.4 | 39.8 | 70.5 | 59.6 | 43.0 |
| HumanOmniV2 â | 68.8 | 25.4 | 37.3 | 75.3 | 58.5 | 41.9 |
| ThinkOmni -Qwen2.5-Omni-3B | | | | | | |
| Qwen2.5-Omni-3B | 56.0 | 18.2 | 32.0 | 69.4 | 56.6 | 37.5 |
| + DeepSeek | 56.1(+0.1) | 20.2(+2.0) | 33.5(+1.5) | 70.1(+0.7) | 57.1(+0.5) | 39.9(+2.4) |
| + Qwen3 | 58.1(+2.1) | 25.3(+7.1) | 38.8(+6.8) | 70.6(+1.2) | 57.3(+0.7) | 39.5(+2.0) |
| ThinkOmni -Qwen2.5-Omni-7B | | | | | | |
| Qwen2.5-Omni-7B | 66.8 | 25.0 | 40.2 | 71.5 | 57.9 | 42.1 |
| + DeepSeek | 68.8(+2.0) | 28.2(+3.2) | 42.0(+1.8) | 73.8(+2.3) | 59.8(+1.9) | 43.2(+1.1) |
| + Qwen3 | 70.2(+3.4) | 32.9(+7.9) | 45.1(+4.9) | 75.5(+4.0) | 59.5(+1.6) | 43.6(+1.5) |
| ThinkOmni -Omni-R1-7B | | | | | | |
| Omni-R1 | 64.7 | 25.4 | 39.8 | 70.5 | 59.6 | 43.0 |
| + DeepSeek | 66.1(+1.4) | 27.0(+1.6) | 43.1(+3.3) | 73.1(+2.6) | 60.3(+0.7) | 43.5(+0.5) |
| + Qwen3 | 71.3(+6.6) | 31.5(+6.1) | 45.2(+5.4) | 75.4(+4.9) | 59.8(+0.2) | 43.4(+0.4) |
To evaluate the generality and scalability of ThinkOmni, we benchmark the improvements of different LRM guides on several OLLMs with varying capability levels. Our main results are presented in Tab. 1. The experiment result shows that ThinkOmni brings extensive improvements across all OLLMs, LRMs, and benchmarks. For example, with the Qwen3 guide, ThinkOmni brings remarkable improvement to Qwen2.5-Omni-7B on MathVision by $7.9\$ , achieving the final score of $32.9\$ . Since LRMs do not have access to omni-modal data contents, our results demonstrate that ThinkOmni indeed lifts the complex reasoning of LRMs to the omni-modal scenario.
We compare our approach with methods trained using reinforcement learning finetuning (RFT) (i.e., Omni-R1 (Zhong et al., 2025) and HumanOmniV2 (Yang et al., 2025b)). Based on the same foundation model, Qwen2.5-Omni-7B, our DeepSeek-guided model achieves comparable performance, while our Qwen3-guided model consistently outperforms all the RFT-based methods. Moreover, our approach can be applied to models already undergoing RFT, further demonstrating broad performance improvements.
In addition, we observe differences in performance gains, which the following factors can explain: 1) the capabilities of LRM models, newer models like Qwen3, with stronger logical understanding and reasoning abilities, achieve greater improvements compared to DeepSeek under identical settings; 2) the training data of LRMs is biased towards scientific and mathematical content, leading to more pronounced gains on these tasks; 3) the tested tasks themselves differ in their demands for reasoning ability, with scientific and mathematical tasks typically requiring more reasoning than audio or general omni-modal tasks.
### 4.3 Compare with Training-free Methods
We use the original evaluation results of the OLLMs as a baseline. In addition, we compare our method with several other training-free methods: 1) Average Logits Fusion, which directly averages the output logits of the OLLM and LRM during inference. 2) Caption-then-Answer, where the OLLM generates a detailed caption for the omni-modal input, and the LRM answers the question based on this caption. 3) Visual Contrastive Decoding (VCD) (Leng et al., 2024), which mitigates hallucinations by contrasting predictions from original inputs with those from distorted ones. In this ablation, we implement the negative baseline by directly removing the multi-modal inputs. As shown in Tab. 2, ThinkOmni significantly outperforms the base OLLM. For Average Logits Fusion, although simple mixing of logits allows the model to generate outputs, it negatively impacts answer accuracy due to improper integration. The Caption-then-Answer experiment demonstrates that when the LRM alone is responsible for answering, even with multi-modal information provided by the OLLM, performance drops significantly because information transmission is one-way. The OLLM cannot respond to the LRMâs specific needs. VCD is designed to enhance attention to multi-modal information rather than reasoning ability, so its performance declines on MathVista, which requires stronger reasoning skills.
Table 2: Comparison with several training-free methods. All are built upon Qwen2.5-Omni-7B.
| Method | MathVista test-mini | MMAU (v05.15.25) test-mini | OmniBench test |
| --- | --- | --- | --- |
| Base Model | 66.8 | 71.5 | 42.1 |
| Average Logits Fusion | 55.0(-11.8) | 55.7(-15.8) | 36.1(-6.0) |
| Caption-then-Answer | 61.0(-5.8) | 59.7(-11.8) | 32.3(-9.8) |
| VCD | 66.5(-0.3) | 72.2(+0.7) | 43.1(+1.0) |
| ThinkOmni (Ours) | 68.8(+2.0) | 73.8(+2.3) | 43.2(+1.1) |
### 4.4 Ablation Study
*[Error downloading image: 2602.23306v1/x7.png]*
Figure 7: Ablation on guidance weight. Left: Performance varies with the constant guidance weight $\alpha$ . Each taskâs optimal $\alpha$ range differs, with $\alpha=0$ as the OLLM baseline. Right: ThinkOmni uses adaptive dynamic weights, and the dynamic $\alpha^{r}$ shows a similar distribution shift, indicating that stepwise contrastive scaling can flexibly adapt to different task requirements. *[Error downloading image: 2602.23306v1/x8.png]*
Figure 8: LRM-as-a-Guide performance scaling. We replace the Guide LRM on several benchmarks to study the impact of different LRM sizes and different LRM series on the performance of our method. The baseline refers to the performance of the Qwen2.5-Omni-7B
#### Ablation study on fixed $\alpha$ and adaptive $\alpha^{r}$
OLLM has limited capability in complex reasoning, while LRM cannot access multi-modal content. Over-reliance on either component leads to sub-optimal performance. As shown in Fig. 7 (a), adjusting the fixed guiding weight $\alpha$ markedly impacts results: when $\alpha=0$ , performance matches the original OLLM, and extreme $\alpha$ values reduce scores on both benchmarks. In contrast, our Stepwise Contrastive Scaling (Full ThinkOmni) consistently achieves superior results across both benchmarks. Furthermore, Fig. 7 (b) visualizes the distribution of the dynamic $\alpha^{r}$ , revealing distinct shifts across different tasks and underscoring the adaptive nature of our method in autonomously tuning parameters to meet specific task requirements.
#### Ablation study on different LRMs
Fig. 8 demonstrates the performance of Qwen2.5-Omni-7B under the guidance of LRM across different model sizes and series. When the LRM size is too small, performance on MathVista and OmniBench both degrades, as the language capabilities of LRM and the base OLLM are insufficiently matched. A sufficiently strong LRM positively impacts all benchmarks, with larger LRMs generally yielding better results than smaller ones. However, increasing the size to Qwen3-14B does not lead to further improvements on MMAU and OmniBench, suggesting that enhanced reasoning ability has a limited effect on these tasks, although the results still surpass the baseline.
### 4.5 Analysis
#### Qualitative analysis
Fig. 9 illustrates the generation process of ThinkOmni, where darker colors indicate greater LRM contributions to each token. These tokens, often logical connectives (e.g., âbutâ, âThereforeâ) and key terms (e.g., âtraditionalâ, âcommonâ), are evenly distributed, showing that LRM consistently guides reasoning rather than merely supplementing OLLM. This highlights LRMâs role in analyzing multi-modal clues and driving logical inference. In contrast, lighter tokens are mainly function words and specific terms reflecting multi-modal content, suggesting that OLLM focuses on retrieving information and constructing fluent responses under LRMâs guidance. For more cases, see Appendix D.
#### Failure case analysis
We identified several representative failure cases from the response. 1) ThinkOmni demonstrates correct multimodal perception, but conflicting information within the input leads to erroneous reasoning. As shown in Fig. 2 (a), the model correctly recognizes that the highest visible marking on the beaker is 400ml. However, due to the beaker being labeled as 600ml, the model incorrectly infers that only a portion of the beaker is visible in the image, resulting in the wrong final answer. 2) Insufficient perceptual ability and limited sensitivity to subtle differences in the input lead to incorrect answers. As illustrated in Fig. 2 (b), the model fails to accurately detect the actual onset of the drum kit in the audio and mistakenly identifies the beginning of the audio as the start of the drum kitâs performance, ultimately producing the wrong answer.
#### Efficiency analysis
Although our method introduces some additional computation, it remains efficient. We measured generation latency on an H800-80G GPU with KV cache in the generate stage, benchmarking VCD (Leng et al., 2024), ProxyTuning (Liu et al., 2024), and ThinkOmni using 100 random OmniBench (Li et al., 2024) samples. VCD performs two forward passes per step, while ProxyTuning and ThinkOmni require three. As shown in Tab. 3, our method (7B+7B setting) incurs $1.38\times$ in the prefill stage (first token generation) and $2.88\times$ in the generate stage (response generation utilizing KV cache). Importantly, during decoding, the guiding model in ThinkOmni processes only text, which helps to reduce latency. *[Error downloading image: 2602.23306v1/x9.png]*
Figure 9: An OmniBench case study. This case study visualizes the reasoning process of ThinkOmni âQwen2.5-Omniâ7B, highlighting the stepwise contrastive scaling coefficient $\alpha^{r}$ . *[Error downloading image: 2602.23306v1/x10.png]*
Figure 10: Failure case study. (a) ThinkOmni makes a reasoning error due to conflicting information between the visible beaker markings (400ml) and the label (600ml). (b) ThinkOmni fails because it cannot accurately detect the drum kitâs true start time in the audio. For the readerâs understanding, the colored markings in Fig. 2 are visual aids added afterward and were not provided to the model.
Table 3: Generation latency comparison. Prefill Time: first token generation latency; Generate Time: response generation with KV cache. Results are averaged over 100 samples from OmniBench.
| Guidence Decoding Method | Model Size | Prefill Time $\downarrow$ | Generate Time $\downarrow$ |
| --- | --- | --- | --- |
| None (Baseline) | 7B | 0.138s | 0.025s |
| VCD (Leng et al., 2024) | 7B | 0.262s ( $1.89\times$ ) | 0.050s ( $2.00\times$ ) |
| ProxyTuning (Liu et al., 2024) | 7B + 3B + 3B | 0.406s ( $2.94\times$ ) | 0.086s ( $3.44\times$ ) |
| ThinkOmni (Ours) | 7B +7B | 0.191s ( $1.38\times$ ) | 0.072s ( $2.88\times$ ) |
| ThinkOmni (Ours) | 7B +1.5B | 0.191s ( $1.38\times$ ) | 0.069s ( $2.76\times$ ) |
## 5 Related Work
#### Omni-modal Large Language Models
With the rapid advancement of large language models, expanding their capabilities to omni-modal domains has become a key focus. Omni-modal Large Language Models (OLLM) align and process information from multiple modalities, capturing richer semantics and context than single-modal systems. Proprietary models (Hurst et al., 2024; Deepmind, 2025) demonstrate impressive real-time multi-modal interaction, while open-source efforts such as Qwen2.5-Omni (Xu et al., 2025; Li et al., 2025b; Liu et al., 2025c; Yang et al., 2025b; Fu et al., 2025; Xie and Wu, 2024; Chen et al., 2025) are quickly closing the gap in modality alignment and deployment.
#### Large Reasoning Model
Recent advances in large-scale reasoning models, such as o1 (OpenAI, 2025) and DeepSeek-R1 (Guo et al., 2025), highlight the challenge of robust general reasoning. Early methods used supervised fine-tuning with chain-of-thought data (Xu et al., 2024; Linger et al., 2025), while recent work leverages reinforcement learning (Shao et al., 2024; Xiaomi et al., 2025; Team et al., 2025; Yu et al., 2025; Tan et al., 2025; Zhu et al., 2026) for autonomous reasoning. As OLLMs tackle complex cross-modal reasoning, bridging perception and reasoning remains a core challenge.
#### Decoding-time Algorithm
Decoding-time algorithms refine language model outputs at inference in a training-free manner. Contrastive Decoding (OâBrien and Lewis, 2023) improves long-form generation by avoiding degenerate outputs, and Visual Contrastive Decoding (Leng et al., 2024) reduces hallucination via visual input perturbation. ProxyTuning (Liu et al., 2024) combines expert outputs and injects knowledge from finetuned models. ProxyThinker (Xiao et al., 2026) extends ProxyTuning to multi-modal reasoning tasks. Beyond visual modalities, ThinkOmni adaptively integrates reasoning with omni-modal perception, achieving a flexible fusion of fast and slow thinking.
## 6 Conclusion
We present ThinkOmni, a training-free inference-time framework that achieves robust and generalizable reasoning enhancement by introducing an off-the-shelf LRM guide decoding with a stepwise adaptive scaling mechanism. Across six challenging multi-modal benchmarks, it delivers consistent gains, often matching or surpassing reinforcement-fine-tuned models, suggesting a general and extensible paradigm for omni-modal reasoning.
#### Limitation
ThinkOmni requires shared vocabularies between the OLLM and LRM for logit fusion and introduces extra inference overhead due to additional forward passes. Nevertheless, we believe our approach offers valuable insights for bridging the gap between multi-modal and textual LLMs and provides a sustainable direction for future LLM improvements.
## Aknowledegment
This work was done during the research internship of Yiran Guan, Sifan Tu, Linghao Zhu, and Dingkang Liang at Xiaomi Inc. MiLM Plus team.
This work was supported by the National Natural Science Foundation of China (NO. 62441615, 62225603, 62576147).
## Ethics statement
This work introduces a trainingâfree framework using only publicly available models (Qwen2.5-Omni (Xu et al., 2025), Omni-R1 (Zhong et al., 2025), DeepSeek-R1 (Guo et al., 2025), Qwen3 (Yang et al., 2025a)) and benchmarks (see Appendix B), without collecting new human, biometric, or sensitive data. Risks stem from inherited biases, possible hallucinated crossâmodal attributions, and overâtrust in generated reasoning chains; the method does not ensure factuality or safety in highâstakes domains. Before deployment, we advise bias auditing, human oversight, and external safety / factuality filters. Environmental impact is reduced relative to finetuning because no additional training is performed.
## Reproducibility statement
Reproducibility is supported by: 1) the explicit inference formulation (Eq. 3 and Eq. 5); 2) unified decoding hyperparameters (see Appendix B); 3) public benchmark splits and the twoâstage answer extraction pipeline follow Xiao et al. (2026); 4) hardware specification (80G VRAM GPU, KV cache enabled). The code will be coming up soon upon acceptance.
## LLM Usage
LLMs were used only for 1) minor code scaffolding / refactoring (boilerplate, log parsing), 2) linguistic polishing after technical content was finalized, and 3) answer string normalization in the evaluation pipeline (format extraction, not scoring). They were not used for ideas, model / method design, experiments, analyses, claims, or interpretations. The researchers authored, validated, and cross-checked all algorithms, equations, results, and conclusions. Every LLM-assisted output was manually reviewed to prevent hallucination. Thus, LLM involvement provides no creative contribution and does not affect the authenticity or reliability of the paper.
## References
- L. Chen, H. Hu, M. Zhang, Y. Chen, Z. Wang, Y. Li, P. Shyam, T. Zhou, H. Huang, M. Yang, et al. (2025) Omnixr: evaluating omni-modality language models on reasoning across modalities. Proc. of Intl. Conf. on Learning Representations. Cited by: §5.
- Y. Chuang, Y. Xie, H. Luo, Y. Kim, J. Glass, and P. He (2024) Dola: decoding by contrasting layers improves factuality in large language models. Proc. of Intl. Conf. on Learning Representations. Cited by: §3.2.
- G. Deepmind (2025) Gemini 2.5: Our most intelligent AI model â blog.google. Note: https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/#gemini-2-5-thinking Cited by: §5.
- C. Fu, H. Lin, X. Wang, Y. Zhang, Y. Shen, X. Liu, H. Cao, Z. Long, H. Gao, K. Li, et al. (2025) Vita-1.5: towards gpt-4o level real-time vision and speech interaction. arXiv preprint arXiv:2501.01957. Cited by: §1, §5.
- C. Gan, Y. Zhang, J. Wu, B. Gong, and J. B. Tenenbaum (2020) Look, listen, and act: towards audio-visual embodied navigation. In 2020 IEEE International Conference on Robotics and Automation, pp. 9701â9707. Cited by: §1.
- D. Guo, D. Yang, H. Zhang, J. Song, P. Wang, Q. Zhu, R. Xu, R. Zhang, S. Ma, X. Bi, et al. (2025) DeepSeek-r1 incentivizes reasoning in llms through reinforcement learning. Nature 645 (8081), pp. 633â638. Cited by: Appendix B, §1, §4.1, Table 1, §5, Ethics statement.
- A. Hurst, A. Lerer, A. P. Goucher, A. Perelman, A. Ramesh, A. Clark, A. Ostrow, A. Welihinda, A. Hayes, A. Radford, et al. (2024) Gpt-4o system card. arXiv preprint arXiv:2410.21276. Cited by: §5.
- S. Leng, H. Zhang, G. Chen, X. Li, S. Lu, C. Miao, and L. Bing (2024) Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In Proc. of IEEE Intl. Conf. on Computer Vision and Pattern Recognition, pp. 13872â13882. Cited by: §2.2, §3.2, §4.3, §4.5, Table 3, §5.
- G. Li, J. Liu, H. Dinkel, Y. Niu, J. Zhang, and J. Luan (2025a) Reinforcement learning outperforms supervised fine-tuning: a case study on audio question answering. arXiv preprint arXiv:2503.11197. Cited by: §1.
- X. L. Li, A. Holtzman, D. Fried, P. Liang, J. Eisner, T. Hashimoto, L. Zettlemoyer, and M. Lewis (2022) Contrastive decoding: open-ended text generation as optimization. arXiv preprint arXiv:2210.15097. Cited by: §1, §2.2.
- Y. Li, J. Liu, T. Zhang, S. Chen, T. Li, Z. Li, L. Liu, L. Ming, G. Dong, D. Pan, et al. (2025b) Baichuan-omni-1.5 technical report. arXiv preprint arXiv:2501.15368. Cited by: §1, §5.
- Y. Li, G. Zhang, Y. Ma, R. Yuan, K. Zhu, H. Guo, Y. Liang, J. Liu, Z. Wang, J. Yang, et al. (2024) Omnibench: towards the future of universal omni-language models. arXiv preprint arXiv:2409.15272. Cited by: 6th item, Appendix B, §4.1, §4.5.
- Z. Lin, Y. Gao, X. Zhao, Y. Yang, and J. Sang (2025) Mind with eyes: from language reasoning to multimodal reasoning. arXiv preprint arXiv:2503.18071. Cited by: §1.
- D. Linger, L. Zhu, Y. Liu, Y. Wang, Q. Xie, J. Wu, G. Zhang, Y. Zhu, and X. Bai (2025) Theorem-validated reverse chain-of-thought problem generation for geometric reasoning. Proc. of Conf. on Empirical Methods in Natural Language Processing, pp. 718â735. Cited by: §5.
- A. Liu, X. Han, Y. Wang, Y. Tsvetkov, Y. Choi, and N. A. Smith (2024) Tuning language models by proxy. First Conference on Language Modeling. Cited by: §1, §2.2, §4.5, Table 3, §5.
- Y. Liu, B. Peng, Z. Zhong, Z. Yue, F. Lu, B. Yu, and J. Jia (2025a) Seg-zero: reasoning-chain guided segmentation via cognitive reinforcement. arXiv preprint arXiv:2503.06520. Cited by: §1.
- Z. Liu, Z. Sun, Y. Zang, X. Dong, Y. Cao, H. Duan, D. Lin, and J. Wang (2025b) Visual-rft: visual reinforcement fine-tuning. Porc. of IEEE Intl. Conf. on Computer Vision. Cited by: §1.
- Z. Liu, Y. Dong, J. Wang, Z. Liu, W. Hu, J. Lu, and Y. Rao (2025c) Ola: pushing the frontiers of omni-modal language model with progressive modality alignment. arXiv e-prints, pp. arXivâ2502. Cited by: §1, §5.
- P. Lu, H. Bansal, T. Xia, J. Liu, C. Li, H. Hajishirzi, H. Cheng, K. Chang, M. Galley, and J. Gao (2024) Mathvista: evaluating mathematical reasoning of foundation models in visual contexts. Proc. of Intl. Conf. on Learning Representations. Cited by: 1st item, Appendix B, Figure 5, §4.1.
- R. Luo, T. Lin, H. Zhang, Y. Wu, X. Liu, M. Yang, Y. Li, L. Chen, J. Li, L. Zhang, et al. (2025) OpenOmni: advancing open-source omnimodal large language models with progressive multimodal alignment and real-time self-aware emotional speech synthesis. Proc. of Advances in Neural Information Processing Systems. Cited by: §1.
- S. OâBrien and M. Lewis (2023) Contrastive decoding improves reasoning in large language models. arXiv preprint arXiv:2309.09117. Cited by: §5.
- OpenAI (2025) OpenAI o1 System Card â openai.com. Note: https://cdn.openai.com/o1-system-card.pdf Cited by: §1, §5.
- A. Rouditchenko, S. Bhati, E. Araujo, S. Thomas, H. Kuehne, R. Feris, and J. Glass (2025) Omni-r1: do you really need audio to fine-tune your audio llm?. arXiv preprint arXiv:2505.09439. Cited by: §1.
- S. Sakshi, U. Tyagi, S. Kumar, A. Seth, R. Selvakumar, O. Nieto, R. Duraiswami, S. Ghosh, and D. Manocha (2025) Mmau: a massive multi-task audio understanding and reasoning benchmark. Proc. of Intl. Conf. on Learning Representations. Cited by: 4th item, Appendix B, Figure 5, §4.1.
- Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y. Li, Y. Wu, et al. (2024) Deepseekmath: pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300. Cited by: §1, §5.
- W. Tan, J. Li, J. Ju, Z. Luo, J. Luan, and R. Song (2025) Think silently, think fast: dynamic latent compression of llm reasoning chains. Proc. of Advances in Neural Information Processing Systems. Cited by: §5.
- K. Team, A. Du, B. Gao, B. Xing, C. Jiang, C. Chen, C. Li, C. Xiao, C. Du, C. Liao, et al. (2025) Kimi k1. 5: scaling reinforcement learning with llms. arXiv preprint arXiv:2501.12599. Cited by: §5.
- K. Wang, J. Pan, W. Shi, Z. Lu, H. Ren, A. Zhou, M. Zhan, and H. Li (2024) Measuring multimodal mathematical reasoning with math-vision dataset. Proc. of Advances in Neural Information Processing Systems 37, pp. 95095â95169. Cited by: 2nd item, Appendix B, §4.1.
- Y. Wang, Z. Wang, B. Xu, Y. Du, K. Lin, Z. Xiao, Z. Yue, J. Ju, L. Zhang, D. Yang, et al. (2025) Time-r1: post-training large vision language model for temporal video grounding. Proc. of Advances in Neural Information Processing Systems. Cited by: §1.
- Z. Xiao, J. Koo, S. Ouyang, J. Hernandez, Y. Meng, and V. Ordonez (2026) ProxyThinker: test-time guidance through small visual reasoners. Proc. of Intl. Conf. on Learning Representations. Cited by: §2.2, §5, Reproducibility statement.
- L. Xiaomi, B. Xia, B. Shen, D. Zhu, D. Zhang, G. Wang, H. Zhang, H. Liu, J. Xiao, J. Dong, et al. (2025) MiMo: unlocking the reasoning potential of language modelâfrom pretraining to posttraining. arXiv preprint arXiv:2505.07608. Cited by: §5.
- Z. Xie and C. Wu (2024) Mini-omni2: towards open-source gpt-4o with vision, speech and duplex capabilities. arXiv preprint arXiv:2410.11190. Cited by: §5.
- G. Xu, P. Jin, L. Hao, Y. Song, L. Sun, and L. Yuan (2024) Llava-o1: let vision language models reason step-by-step. arXiv preprint arXiv:2411.10440. Cited by: §1, §5.
- J. Xu, Z. Guo, J. He, H. Hu, T. He, S. Bai, K. Chen, J. Wang, Y. Fan, K. Dang, et al. (2025) Qwen2. 5-omni technical report. arXiv preprint arXiv:2503.20215. Cited by: §1, §1, §4.1, Table 1, §5, Ethics statement.
- A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, et al. (2025a) Qwen3 technical report. arXiv preprint arXiv:2505.09388. Cited by: Appendix B, §4.1, Table 1, Ethics statement.
- Q. Yang, S. Yao, W. Chen, S. Fu, D. Bai, J. Zhao, B. Sun, B. Yin, X. Wei, and J. Zhou (2025b) HumanOmniV2: from understanding to omni-modal reasoning with context. arXiv preprint arXiv:2506.21277. Cited by: §1, §4.2, §5.
- Y. Yang, X. He, H. Pan, X. Jiang, Y. Deng, X. Yang, H. Lu, D. Yin, F. Rao, M. Zhu, et al. (2025c) R1-onevision: advancing generalized multimodal reasoning through cross-modal formalization. arXiv preprint arXiv:2503.10615. Cited by: §1.
- Q. Yu, Z. Zhang, R. Zhu, Y. Yuan, X. Zuo, Y. Yue, T. Fan, G. Liu, L. Liu, X. Liu, et al. (2025) DAPO: an open-source llm reinforcement learning system at scale. Proc. of Advances in Neural Information Processing Systems. Cited by: §1, §5.
- R. Zhang, D. Jiang, Y. Zhang, H. Lin, Z. Guo, P. Qiu, A. Zhou, P. Lu, K. Chang, Y. Qiao, et al. (2024) Mathverse: does your multi-modal llm truly see the diagrams in visual math problems?. In Proc. of European Conference on Computer Vision, pp. 169â186. Cited by: 3rd item, Appendix B, §4.1.
- S. Zhang, S. Guo, Q. Fang, Y. Zhou, and Y. Feng (2025) Stream-omni: simultaneous multimodal interactions with large language-vision-speech model. arXiv preprint arXiv:2506.13642. Cited by: §1.
- J. Zhao, X. Wei, and L. Bo (2025) R1-omni: explainable omni-multimodal emotion recognition with reinforcement learning. arXiv preprint arXiv:2503.05379. Cited by: §1.
- H. Zhong, M. Zhu, Z. Du, Z. Huang, C. Zhao, M. Liu, W. Wang, H. Chen, and C. Shen (2025) Omni-r1: reinforcement learning for omnimodal reasoning via two-system collaboration. Proc. of Advances in Neural Information Processing Systems. Cited by: §1, §4.1, §4.2, Table 1, Ethics statement.
- Z. Zhou, R. Wang, and Z. Wu (2025) Daily-omni: towards audio-visual reasoning with temporal alignment across modalities. arXiv preprint arXiv:2505.17862. Cited by: 5th item, Appendix B, §4.1.
- L. Zhu, Y. Guan, D. Liang, J. Ju, Z. Luo, B. Qin, J. Luan, Y. Liu, and X. Bai (2026) Shuffle-r1: efficient rl framework for multimodal large language models via data-centric dynamic shuffle. Proc. of Intl. Conf. on Learning Representations. Cited by: §5.
## Appendix A A detailed explanation for ThinkOmni
At each decoding step $t$ , the Omni-modal Large Language Model (OLLM) generates two token probability distributions: conditioned on the full omni-modal input and the textual only input. In contrast, the Large Reasoning Model (LRM) outputs a single distribution based solely on the textual input.
We would like to decide how much of the next-token decision should be driven by perception ( $P_{O}^{(t)}$ ) and how much by abstract reasoning ( $P_{R}^{(t)}$ ). To gauge the disagreement between these signals, we compute two JensenâShannon divergences
$$
D_{R}^{(t)}\;=\;\mathrm{JS}\bigl(P_{R}^{(t)}\;\|\;P^{(t)}\bigr),\qquad D_{P}^{(t)}\;=\;\mathrm{JS}\bigl(P_{O}^{(t)}\;\|\;P^{(t)}\bigr).
$$
$D_{R}^{(t)}$ is large when the reasoner (LRM) wants a different token than the perceiver (OLLM text-only); $D_{P}^{(t)}$ is large when masking the multi-modal content changes the OLLMâs belief is crucial, i.e. when perception.
1. If the current step mainly requires perception (e.g. reading a label in the image or detecting a sound), masking the modalities hurts the OLLM, so $D_{P}^{(t)}$ is large while $D_{R}^{(t)}$ stays small. Consequently $\alpha_{t}^{r}\approx 0$ and the model trusts the perception logits.
1. If the step calls for reasoning (e.g. algebraic manipulation after the visual information is extracted), the text-only LRM disagrees with $P^{(t)}$ , so $D_{R}^{(t)}$ dominates and $\alpha_{t}^{r}\approx 1$ . The generation is therefore guided by the stronger logical signal from the LRM.
1. For mixed cases, $\alpha_{t}$ smoothly interpolates between the two extremes, allowing the decoder to blend perception and reasoning in real time.
## Appendix B Experiment Details
#### Hyperparameter Settings
During our experiments, we used the following inference parameters to ensure standard model outputs and to prevent performance degradation and endless repetitions, following (Yang et al., 2025a): a temperature of $0.6$ , a top_p value of $0.95$ , a repetition_penalty of $1.03$ , and a max_new_tokens of $4096$ . In addition, all LRMs append a <think> tag to the end of the prompt during inference to ensure the reasoning state is activated (Guo et al., 2025).
#### Dynamic Weight and Mixing.
Given the calculated divergences $D_{R}$ and $D_{P}$ , we determine the reasoning weight $\alpha^{r}_{t}$ by measuring the reasoning surplus, clamped to $[0,1]$ :
$$
\alpha^{r}_{t}=\text{clip}\left(D_{R}-D_{P},0,1.0\right). \tag{6}
$$
To ensure stability during the initial phase, we apply a linear warmup for the first $T_{\text{warm}}=5$ steps:
$$
\alpha^{r}_{t}\leftarrow\min\left(\alpha^{r}_{t},0.1\cdot t\right). \tag{7}
$$
Finally, the mixed logits are computed efficiently via the following implementation, which integrates the base, reasoning, and negative distributions:
$$
\hat{z}_{t}=(2-\alpha^{r}_{t})\cdot\mathit{M}_{\mathit{O}}(x_{<t},\mathit{O})+\alpha^{r}_{t}\cdot\mathit{M}_{\mathit{R}}(x_{<t})-\mathit{M}_{\mathit{O}}(x_{<t}). \tag{8}
$$
#### Benchmarks Details
The datasets used in our experiments include MathVista (Lu et al., 2024), MathVision (Wang et al., 2024), MathVerse (Zhang et al., 2024), MMAU (Sakshi et al., 2025), Daily-Omni (Zhou et al., 2025) and OmniBench (Li et al., 2024).
- MathVista (Lu et al., 2024): We evaluate on the test-mini split of MathVista (1,000 samples), a unified benchmark for mathematical reasoning in visual contexts, which includes three newly introduced datasets (IQTest, FunctionQA, and PaperQA), as well as 9 MathQA and 19 VQA datasets from previous work.
- MathVision (Wang et al., 2024): We evaluate on the MathVision dataset (3,040 samples), a curated collection of high-quality mathematical problems with visual contexts. Covering 16 mathematical disciplines and five difficulty levels, MATHVision offers a comprehensive and diverse benchmark for assessing the mathematical reasoning abilities of LMMs.
- MathVerse (Zhang et al., 2024): We evaluate on the test-mini split of MathVerse (3,940 samples), a visual math benchmark with 2,612 multi-subject problems and diagrams. Each issue is annotated into six multi-modal versions, totaling 15K samples, to assess MLLMsâ understanding of visual information in math reasoning.
- MMAU (Sakshi et al., 2025): We evaluate on the test-mini split of MMAU (1,000 samples), which consists of 10K audio clips with natural language questions and answers across speech, sounds, and music. Covering 12 retrieval and 15 reasoning types, MMAU challenges models with expert-level, domain-specific audio understanding and reasoning.
- Daily-Omni (Zhou et al., 2025): We evaluate on Daily-Omni (1,197 samples), which contains 684 videos from 11 YouTube categories and questions requiring integration of audio, visual, and textual information. The benchmark covers 30-second and 60-second videos to assess multi-modal reasoning abilities.
- OmniBench (Li et al., 2024): We evaluate on OmniBench (1,142 samples), which covers a wide range of reasoning and cognitive skills, from perception to complex reasoning. Tasks include object recognition, temporal and spatial reasoning, symbolic and quantitative processing, and various audio types, including speech, sound events, and music.
## Appendix C Future Work
ThinkOmni represents an attempt to introduce omni-modal capabilities based on textual reasoning abilities. We plan to explore additional modalities, such as 3D point clouds, protein structures, and reasoning applications in image/video generation scenarios. Moreover, we are also curious about âwhat truly works during the reasoning processâ, which is of great significance for understanding why reasoning abilities in the textual domain can generalize to a wider range of modalities.
## Appendix D More Cases
*[Error downloading image: 2602.23306v1/x11.png]*
Figure 11: Omni-modal Reasoning. *[Error downloading image: 2602.23306v1/x12.png]*
Figure 12: Audio Reasoning. *[Error downloading image: 2602.23306v1/x13.png]*
Figure 13: Visual Reasoning.