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 and data-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-modal 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. Code is publicly available at https://github.com/1ranGuan/thinkomni
Figures & tables
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.
Figure 2: Performance comparison.
Figure 3: Guidance decoding methods. “Guid.” denotes the guiding model, and “Amat.” denotes the amateur model.
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.
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 α limits the ability of the model to optimally adapt to task-specific needs, highlighting the need for a more flexible approach.
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.
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
-
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 ‘ ∗ ’ are evaluated using our own evaluation scripts.
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 )
Table 2: Comparison with several training-free methods. All are built upon Qwen2.5-Omni-7B.
Figure 7: Ablation on guidance weight. Left: Performance varies with the constant guidance weight α . Each task’s optimal α range differs, with α=0 as the OLLM baseline. Right: ThinkOmni uses adaptive dynamic weights, and the dynamic αr shows a similar distribution shift, indicating that stepwise contrastive scaling can flexibly adapt to different task requirements.
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
Figure 9: An OmniBench case study. This case study visualizes the reasoning process of ThinkOmni ‑Qwen2.5-Omni‑7B, highlighting the stepwise contrastive scaling coefficient αr .
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 kit’s true start time in the audio. 2 2 2 For the reader’s understanding, the colored markings in Fig. 2 are visual aids added afterward and were not provided to the model.
Guidence Decoding Method
Model Size
Prefill Time ↓
Generate Time ↓
None (Baseline)
7B
0.138s
0.025s
VCD ( Leng et al., 2024 )
7B
0.262s ( 1.89× )
0.050s ( 2.00× )
ProxyTuning ( Liu et al., 2024 )
7B + 3B + 3B
0.406s ( 2.94× )
0.086s ( 3.44× )
ThinkOmni (Ours)
7B +7B
0.191s ( 1.38× )
0.072s ( 2.88× )
ThinkOmni (Ours)
7B +1.5B
0.191s ( 1.38× )
0.069s ( 2.76× )
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.
Improving joint audio-visual reasoning in Omni Large Language Models typically incurs substantial data construction and training costs. Our diagnostics reveal multi-hop reasoning difficulties despite correct answers to all corresponding single-hop questions and suggest partial decoupling in the local optimization of perception and reasoning objectives. This motivates post-training with different emphases on these capabilities. Text-only reasoning training yields gains across data sources, model scales, and families. With the best-performing text-only configuration, supervised fine-tuning followed by reinforcement learning (RL) raises Qwen2.5-Omni-7B's geometric mean of nine reasoning scores by 25.83% over the base model, outperforming the complete native audio-visual route with 56.6% fewer GPU-hours. Training on data synthesized entirely by a text-only LLM raises this geometric mean by 21.01% without audio-visual data in construction or training. However, text-only training degrades perception. We therefore propose a text-centric post-training paradigm: text-only training provides the main reasoning optimization, and reduced-data native audio-visual RL then refines perception. Refinement uses about 90% fewer input tokens than full-data audio-visual RL, restores perception above the base level, and retains 93.5% of the best-performing text-only pipeline's reasoning gain.
Recent multimodal large language models (MLLMs) have shown strong chain-of-thought (CoT) reasoning ability on vision-language tasks, but their direct deployment in real-world systems is often limited by latency and resource constraints. In practice, smaller MLLMs are preferred for online serving, yet their reasoning performance is bottlenecked by the lack of large-scale, high-quality multimodal CoT supervision. In this paper, we present OmniThoughtVis, a scalable data curation and distillation pipeline for transferring multimodal reasoning capabilities from high-capacity teacher models to smaller, deployment-oriented MLLMs. Starting from a diverse open-source seed pool, our pipeline generates structured CoT traces and performs joint annotation of reasoning difficulty, answer quality, and semantic task tags. To maintain data quality at scale, we combine rule-based filtering, difficulty-aware selection, and tag-based diversity sampling, resulting in a curated corpus of 1.8M samples that supports controllable subset construction for downstream training. We use OmniThoughtVis to distill Qwen3-VL models from 2B to 8B parameters and evaluate them on nine multimodal reasoning benchmarks. The resulting distilled models show consistent gains across model scales, including improvements of up to +16.8 points on MathVerse and +5.6 points on MMMU-Pro for the 4B model. Notably, the distilled 4B model matches or surpasses the undistilled 8B baseline on several tasks, highlighting the practical value of scalable reasoning distillation for deployment-oriented MLLMs.
Omni Large Language Models (Omni-LLMs) have demonstrated impressive capabilities in holistic multi-modal perception, yet they consistently falter in complex scenarios requiring synergistic omni-modal reasoning. Beyond understanding global multimodal context, effective reasoning also hinges on fine-grained cross-modal alignment, especially identifying shared referents across modalities, yet this aspect has been largely overlooked. To bridge this gap, we formalize the challenge as a cross-modal coreference problem, where a model must localize a referent in a source modality and re-identify it in a target modality. Building on this paradigm, we introduce CrossOmni, a dataset comprising nine tasks equipped with human-designed reasoning rationales to evaluate and enhance this capability. Experiments on 13 Omni-LLMs reveal systematic weaknesses in cross-modal coreference, which we attribute to the absence of coreference-aware thinking patterns. To address this, we enhance cross-modal alignment via two strategies: a training-free In-Context Learning method and a training-based SFT+GRPO framework designed to induce such thinking patterns. Both approaches yield substantial performance gains and generalize effectively to collaborative reasoning tasks. Overall, our findings highlight cross-modal coreference as a crucial missing piece for advancing robust omni-modal reasoning.