Omni-R1: Towards the Unified Generative Paradigm for Multimodal Reasoning

Abstract: Multimodal LLMs (MLLMs) are making significant progress in multimodal reasoning. Early approaches focus on pure text-based reasoning. More recent studies have incorporated multimodal information into the reasoning steps; however, they often follow a single task-specific reasoning pattern, which limits their generalizability across various multimodal tasks. In fact, there are numerous multimodal tasks requiring diverse reasoning skills, such as zooming in on a specific region or marking an object within an image. To address this, we propose unified generative multimodal reasoning, which unifies diverse multimodal reasoning skills by generating intermediate images during the reasoning process. We instantiate this paradigm with Omni-R1, a two-stage SFT+RL framework featuring perception alignment loss and perception reward, thereby enabling functional image generation. Additionally, we introduce Omni-R1-Zero, which eliminates the need for multimodal annotations by bootstrapping step-wise visualizations from text-only reasoning data. Empirical results show that Omni-R1 achieves unified generative reasoning across a wide range of multimodal tasks, and Omni-R1-Zero can match or even surpass Omni-R1 on average, suggesting a promising direction for generative multimodal reasoning.

• The paper introduces Omni-R1, a unified generative paradigm that interleaves textual and visual outputs to enable comprehensive multimodal reasoning.
• It employs a two-stage training approach combining supervised fine-tuning and RL with perception-calibrated rewards to optimize intermediate visual evidence.
• Empirical results demonstrate significant improvements across diverse tasks, including zero-shot settings using synthetic visual traces.

Unified Generative Paradigm for Multimodal Reasoning: An Expert Analysis of Omni-R1

Motivation and Paradigm Shift

Multimodal LLMs (MLLMs) have advanced beyond mere perception to increasingly sophisticated forms of multimodal reasoning. Early paradigms largely separated perception (vision) and reasoning (language), confining images to static context roles and limiting intermediate computations to the text domain. Progress towards interleaved-modal reasoning has allowed visual information to inform intermediate steps, yet prevailing approaches are bound to single-task patterns, with limited cross-task transfer or composition of visual reasoning skills. As illustrated by tasks that require spatial zoom-in, object marking, dynamic scene prediction, or embedding multiple forms of annotation, there is a clear need to support a broad array of visual reasoning operations within a unified framework.

Figure 1: An example illustrating the necessity of incorporating visual information into the intermediate reasoning steps for multimodal tasks.

Omni-R1 proposes a generative and unified paradigm: all intermediate reasoning—in both the textual and visual channels—is managed as an interleaved sequence, where the MLLM directly generates not only text but also images representing functional, intermediate visual evidence. This approach internalizes “thinking with images” and generalizes the reasoning protocol across diverse multimodal tasks.

Taxonomy of Multimodal Reasoning Tasks and Skills

The paper defines a comprehensive taxonomy (Omni-Bench) capturing the diversity of real-world multimodal tasks (“Uni-Tasks”) and the constituent multimodal reasoning skills (“Uni-Skills”). These range from grounding and zoom for VQA, to visual prediction for operational scenes, overlaid marking for structured figures, and auxiliary lines for geometry reasoning.

Figure 2: The illustration of various multimodal tasks and the corresponding diverse multimodal reasoning skills, with multimodal tasks shown in the top row and the required reasoning skills summarized in the bottom row.

Uni-Skills are explicitly grounded in atomic action protocols—e.g., $\textsf{ZOOM-in}$, $\textsf{BBOX}$, $\textsf{MARK}$, $\textsf{LINE}$, and $\textsf{PRED}$—allowing fine-grained control of visual reasoning in a formalized, executable manner.

Methodology: Frameworks of Omni-R1 and Omni-R1-Zero

Data and Pipeline

The foundational insight of this work is that a single MLLM can—if trained appropriately—internalize both the format and semantics of interleaved multimodal reasoning. This is non-trivial due to the unnatural and highly functional nature of intermediate visual artifacts, as well as the scarcity of annotated interleaved trajectories.

Omni-R1 employs a two-stage protocol: a Perception-Aligned Supervised Fine-Tuning (PeSFT) phase with joint cross-entropy and perceptual alignment loss, and a Perception-Calibrated Relative Policy Optimization (PeRPO) phase leveraging RL with composite rewards (answer accuracy, trajectory format, and perceptual consistency). Notably, a differentiable perception loss leverages a fixed codebook to stabilize training on functional image generation, and RL is applied to optimize the model against structured rewards measuring correctness and the quality of visual generations.

Figure 3: Training pipeline for Omni-R1 and Omni-R1-Zero, highlighting the supervised and RL-based training stages with human or synthetic visual traces.

Omni-R1-Zero addresses the annotation bottleneck by synthesizing stepwise visualizations from text-only chain-of-thought (CoT) data. These synthetic visuals provide bootstrapped supervision so that the framework can induce interleaved multimodal reasoning even in the absence of ground-truth annotated traces.

Empirical Results and Analysis

Results on the composite Omni-Bench suite and established benchmarks (MME, MM-Vet, V$^*$-Bench, MMVP, POPE, BLINK) demonstrate that both Omni-R1 and Omni-R1-Zero outperform strong interleaved generation and fine-tuning baselines. The largest relative improvements are seen in tasks requiring active visual evidence composition (Vision-Op.), and in settings where annotation is limited.

Ablations show that RL with perception calibration (PeRPO) is essential: removing it results in significant regression, particularly on operational and diagrammatic tasks. The perception-calibrated reward term yields consistent improvements by promoting visually coherent intermediate generations.

t-SNE analyses of generated visuals show that supervised learning yields more compact, canonical distributional modes, while RL with synthetic trajectories (Omni-R1-Zero) supports broader yet effective exploration, as evidenced by tight clustering of correct samples.

Figure 4: The t-SNE visualization of generated images from Omni-R1 and Omni-R1-Zero, with markers indicating correctness and demonstrating distributional differences in intermediate generations.

Rich case studies highlight the model's capacity to realize flexible Uni-Skill composition, such as sequenced zoom-in grounding, functional attribute annotation, visual prediction of state transitions, and complex multi-step marking.

Figure 5: Case studies of Omni-R1's multimodal reasoning skills, including grounding, marking, and visual prediction maneuvers.

Critically, Omni-R1-Zero, trained without ground-truth visual traces, generates intermediate visual evidence that not only supports accurate answers but also provides interpretable, stepwise rationales.

Figure 6: Omni-R1-Zero’s generative reasoning process for a commonsense multimodal question, reflecting internal visual “thought.”

Practical and Theoretical Implications

These results validate the feasibility of unified generative multimodal reasoning under moderate or minimal multimodal supervision. The paradigm shift brings several implications:

• Unified Model Deployment: Task-agnostic models supporting callable visual actions reduce the need for task- or domain-specific architectures.
• Interpretability: Interleaved generations (text + functional visual evidence) make models’ intermediate reasoning transparent, critical for high-stakes applications.
• RL for Visual Reasoning: Perception-calibrated rewards coupled with transformer-based RL pipelines can optimize the quality of intermediate visualizations even where direct supervision is scarce or bootstrapped.
• Composable Reasoning: The atomic action protocol enables composition and transfer of reasoning primitives—key for generalist agents and scalable modular AI.

Future Directions

Theoretical advances will focus on scaling supervision signals in low-resource or zero-shot settings, automating the acquisition and evaluation of synthetic interleaved traces, and exploring more sophisticated perception-based reward architectures. Practically, this approach paves the way for more robust, interpretable, and generalist multimodal AI—critical for embodied agents, scientific understanding, and open-world decision-making systems.

Conclusion

Omni-R1 redefines multimodal reasoning as a unified, generative interleaving of textual and visual intermediate states, underpinned by a formal atomic action protocol, robust perception-aligned losses, and RL with structured rewards. The work demonstrates powerful improvements across diverse tasks under both supervised and synthetic-annotation regimes, establishing a foundation for future AI systems with general, interpretable, and compositional multimodal reasoning capabilities (2601.09536).