OmniVerifier-C1: Decomposed Graded Verifier for Rule-Governed Generative World Modeling

Bowei Liu; Xinchen Zhang; Xuhuan Li; Tianlin Zhang; Kaian Jiang; Jingwei Liu; Congyang Zhao; Yifei Qian

cs.CLOct 15, 2026

OmniVerifier-C1: Decomposed Graded Verifier for Rule-Governed Generative World Modeling

Authors: Bowei Liu, Xinchen Zhang, Xuhuan Li, Tianlin Zhang, Kaian Jiang, Jingwei Liu, Congyang Zhao, Yifei Qian,

Organizations: Tsinghua University · Amap, Alibaba Group · Peking University · Microcyto

Abstract

Generative visual verifiers play a crucial role in advancing multimodal intelligence. To enhance the robustness of visual verifiers in complex multimodal scenarios involving compositional requirements, we move beyond conventional binary verifiers that produce only True/False discrete judgments and introduce a verifier that automatically decomposes complex prompts or environmental rules into fine-grained components and assigns pointwise scores to provide a continuous and more informative critique signal. Our experiments further demonstrate that these structured and interpretable outputs improve the verifier's generalization to out-of-domain multimodal critic tasks. To train this verifier, we introduce a Decoupled RL training strategy that jointly optimizes sub-question-level binary judgment and global alignment scoring, using precise local feedback to develop fine-grained verification capabilities that provide a foundation for reliable continuous scoring at the global level. Under a general reinforcement learning framework, we first empirically demonstrate that the proposed pointwise graded verifier significantly outperforms binary verifiers in traditional text-to-image scenarios. We then extend our investigation to rule-governed generative world modeling where generative models are leveraged for world simulation and world reasoning under explicit rules. From a theoretical standpoint, we further show that a pairwise verification paradigm yields more accurate critiques than the pointwise formulation when ground-truth images are available, with empirical results on diverse structured world-modeling tasks such as Maze and Sudoku further validating the effectiveness of our approach. Overall, our findings suggest a key insight: the transition from generative models to structured world-modeling agents critically hinges on the availability of accurate pairwise graded visual verifiers.

Figures & tables

Figure 1: Panels (a)–(d) highlight the advantages of the graded verifier over the binary verifier as a generative reward model, while panels (e)–(h) show that pairwise verification provides more reliable and accurate feedback signals than pointwise verification.

Figure 2: Overview of our training pipeline. Stage 1 trains the Decomposed Graded Verifier with decoupled reinforcement learning to produce fine-grained continuous rewards. Stage 2 introduces a pairwise judgment paradigm that provides cleaner and more reliable reward signals for generative model reinforcement learning.

Methods	Concept Existence			Object Relation		World Dynamics		Image Annotation			State Value Evaluation				STEM		Overall
Methods	Obj.	Attr.	Abs.P.	Spat.	N-Spat.	S-Phy	D-Phy	BBox	Point	Count	Maze	F.Lake	Robot.	GUI	Chart	LaTeX	Overall
Qwen3-VL-8B	0.710	0.690	0.562	0.642	0.716	0.604	0.593	0.831	0.741	0.489	0.420	0.693	0.671	0.787	0.540	0.773	0.654
+ RL(Binary)	0.732	0.727	0.579	0.713	0.754	0.600	0.596	0.843	0.770	0.566	0.434	0.674	0.645	0.782	0.526	0.795	0.671
+ SFT(Binary)	0.689	0.695	0.620	0.654	0.703	0.546	0.627	0.802	0.735	0.504	0.414	0.714	0.661	0.792	0.538	0.802	0.656
+SFT(Binary) + RL(Binary)	0.735	0.721	0.554	0.719	0.748	0.521	0.636	0.840	0.765	0.532	0.446	0.687	0.682	0.799	0.553	0.814	0.672
+SFT(Graded)	0.724	0.709	0.535	0.679	0.657	0.613	0.596	0.817	0.752	0.543	0.427	0.661	0.576	0.769	0.656	0.830	0.659

Table 1: Performance on ViVerBench of the Binary Verifier and Graded Verifier.

Figure 3: Qualitative visualization of traditional text-to-image generation.

Methods	GenEval							CLIPScore
Methods	Single Obj.	Two Obj.	Counting	Colors	Position	Attri. Bind.	Overall	CLIPScore
Stable-Diffusion-3.5-medium	0.99	0.78	0.69	0.79	0.25	0.45	0.66	32.8
+RL(UnifiedReward)	0.99	0.84	0.75	0.86	0.36	0.55	0.73	33.7
+RL(Discrete Binary Verifier)	0.99	0.80	0.72	0.81	0.29	0.47	0.68	33.1
+RL(Decomposed Graded Verifier)	1.00	0.90	0.79	0.89	0.41	0.63	0.77	34.4
FLUX.1-dev	0.99	0.85	0.61	0.84	0.22	0.58	0.68	33.2

Table 2: Performance on GenEval and CLIPScore under different reward models.

Figure 4: Qualitative visualization of rule-governed world-modeling tasks.

Methods	Maze		Sudoku		Aux. Line Construction		Sequence Prediction		Word Search		3D Projection
Methods	Normal	Hard	Normal	Hard	Normal	Hard	Normal	Hard	Normal	Hard	Normal	Hard
Qwen-Image-Edit-2509	17.31	8.73	22.21	5.62	16.79	7.32	9.18	4.33	6.28	2.46	3.86	1.71
+ SFT	56.77	18.96	63.62	29.99	57.62	26.84	46.53	27.96	37.36	17.29	42.94	23.37
+SFT+RL(Pointwise)	67.81	34.78	77.74	50.71	66.98	37.49	61.27	34.18	50.92	26.94	52.19	33.16
+SFT+RL(Pairwise)	85.31	51.92	89.97	62.99	76.19	46.12	73.48	63.64	62.18	34.27	69.28	47.51

Table 3: Performance comparison on game-based image-to-image tasks for world modeling.

Methods	RISEBench					UniREdit-Real					UniREdit-Game					Overall
Methods	T	C	S	L	Overall	AM	ST	PI	PR	Avg.	SI	SR	LHP	LPS	Avg.	Overall
Qwen-Image-Edit-2509	2.4	7.8	14.0	1.2	6.7	70.04	68.49	67.26	64.11	67.48	52.31	32.96	42.46	37.89	41.41	53.16
+ SFT	3.5	7.8	16.0	5.9	8.6	70.56	68.68	67.55	64.98	67.94	66.92	47.48	56.06	43.98	53.61	59.14
+ SFT + RL(Pointwise)	5.9	8.9	18.0	12.9	11.7	71.24	68.73	68.38	65.77	68.53	68.73	49.16	60.32	56.41	58.66	62.37
+ SFT + RL(Pairwise)	7.1	10.0	19.0	23.5	15.0	71.87	69.02	68.72	66.08	68.92	70.23	51.33	68.93	63.53	63.51	65.15

Table 4: Performance comparison on RISEBench and UniREditBench.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Training dynamics of pairwise and pointwise verification across six rule-governed tasks. Each panel shows validation scores and the score gap between the two methods.

Methods	Annotation Source
Methods	GPT-5.5	Gemini-3.1-Pro	Human
Qwen3-VL-8B	68.5	69.5	67.0
+ SFT(Graded)	76.0	76.5	74.5
+ SFT(Graded) + RL(Graded)	94.5	96.0	94.0

Appendix

Table 5: Sub-question-level verification accuracy (%) on 200 held-out samples.

Methods	GenEval							CLIPScore
Methods	Single Obj.	Two Obj.	Counting	Colors	Position	Attri. Bind.	Overall	CLIPScore
Qwen-Image-2512	0.99	0.90	0.77	0.86	0.38	0.52	0.74	34.1
+RL(Discrete Binary Verifier)	1.00	0.92	0.79	0.87	0.42	0.55	0.76	34.4
+RL(Decomposed Graded Verifier)	1.00	0.96	0.85	0.92	0.51	0.69	0.82	35.3

Appendix

Table 6: Performance of Qwen-Image-2512 on GenEval and CLIPScore under different reward models.

Methods	Concept Existence		Object Relation		Image Annotation			Overall (16 tasks)
Methods	Obj.	Attr.	Spat.	N-Spat.	BBox	Point	Count	Overall (16 tasks)
InternVL3-8B	0.513	0.578	0.667	0.619	0.750	0.619	0.604	0.594
+SFT(Graded)	0.526	0.585	0.652	0.632	0.757	0.611	0.618	0.606
+SFT(Graded)+RL(Graded)	0.678	0.684	0.707	0.724	0.821	0.665	0.649	0.652

Appendix

Table 7: Performance on ViVerBench with InternVL3-8B as the verifier backbone. We report seven selected sub-metrics with notable improvements after Decoupled RL. Overall denotes the mean score across all 16 task categories, including those not displayed.

Number of Sub-questions	Mean Alignment Error $\downarrow$
$\leq 10$	0.083
11–20	0.039
21–30	0.072

Appendix

Table 8: Effect of the number of sub-questions on global score estimation on 200 held-out test samples. Lower mean alignment error is better.

Multimodal Model Evaluation

Multimodal model evaluation measures how reliably AI systems understand and generate content across text, images, audio, video, and other data.

See all 770 papers

May 27, 2026

OmniVerifier-M1: Multimodal Meta-Verifier with Explicit Structured Recalibration

Visual outcomes are increasingly central to multimodal large language models, making reliable and fine-grained verification essential for scaling generalist foundation models. In this work, we investigate multimodal meta-verification, which leverages verifier-generated rationales rather than decision-only signals, and explore how to effectively incorporate meta-verification feedback into multimodal verifier training. We identify two key findings. First, symbolic verifier outputs (e.g., bounding boxes) outperform textual explanations as meta-verification rationales, enabling efficient rule-based reinforcement learning rewards while avoiding reliance on model-based rewards from auxiliary judge models. Second, decoupling reinforcement learning objectives for binary judgment and meta-verification substantially outperforms joint reward optimization, due to intrinsic differences in output structure and learning dynamics. Based on these insights, we train OmniVerifier-M1, a generalist visual verifier leveraging symbolic meta-verification and decoupled reinforcement learning. OmniVerifier-M1 provides robust verification and fine-grained error localization, and further enables M1-TTS, a verifier-driven agentic generation system achieving dynamic region-level self-correction. This approach paves the way for more reliable, interpretable, and fine-grained multimodal verification, supporting safer and more controllable foundation model deployment.

Xinchen Zhang, Bowei Liu, Jiale Liu +7

Tsinghua University · Pennsylvania State University · University of Southern California +2

Multimodal Large Language Models Reinforcement Learning with Verifiable Rewards Multimodal Foundation Models Multimodal Model Evaluation

Ask PDF

May 10, 2026

DeltaRubric: Generative Multimodal Reward Modeling via Joint Planning and Verification

Aligning Multimodal Large Language Models (MLLMs) requires reliable reward models, yet existing single-step evaluators can suffer from lazy judging, exploiting language priors over fine-grained visual verification. While rubric-based evaluation mitigates these biases in text-only settings, extending it to multimodal tasks is bottlenecked by the complexity of visual reasoning. The critical differences between responses often depend on instance-specific visual details. Robust evaluation requires dynamically synthesizing rubrics that isolate spatial and factual discrepancies. To address this, we introduce

\textbf{DeltaRubric}

, an approach that reformulates multimodal preference evaluation as a plan-and-execute process within a single MLLM. DeltaRubric operates in two steps: acting first as a

\textit{Disagreement Planner}

, the model generates a neutral, instance-specific verification checklist. Transitioning into a

\textit{Checklist Verifier}

, it executes these self-generated checks against the image and question to produce the final grounded judgment. We formulate DeltaRubric as a multi-role reinforcement learning problem, jointly optimizing planning and verification capabilities. Validated on Qwen3-VL 4B and 8B Instruct models, DeltaRubric achieves solid empirical gains. For instance, On VL-RewardBench, it improves base model overall accuracy by

\textbf{+22.6}

(4B) and

\textbf{+18.8}

(8B) points, largely outperforming standard no-rubric baselines. The results demonstrate that decomposing evaluation into structured, verifiable steps leads to more reliable and generalizable multimodal reward modeling.

Rui Liu, Dian Yu, Zhenwen Liang +6

1Tencent Hunyuan · University of Maryland, College Park · University of North Carolina, Chapel Hill

Multimodal Understanding Multimodal Reward Modeling Multimodal Large Language Models Rubric-Based Evaluation Multimodal Foundation Models Vision-Language Model-as-a-Judge+3

Ask PDF

Aug 11, 2026

VERDICT: Training-Free Step-Wise Verification of Multimodal Reasoning via Disagreement-Aware Consensus

Multimodal large language models often generate reasoning chains containing subtle errors that lead to incorrect answers. Current verification approaches have notable limitations. Existing approaches either require expensive labelled supervision with inconsistent cross-task performance or aggregate scores from multiple sources by simple aggregations, missing a key insight: when these scores disagree, that disagreement itself carries important information about whether a reasoning step is truly valid or not. We formalise this as a coupled scoring problem among disparate, frozen verifiers, interpretable as a coordination game with a unique closed-form equilibrium where agreement signals valid steps while disagreement reveals instability. Towards this end, we propose a training-free domain-agnostic step-wise verification approach we call VERDICT: VERification via Disagreement-Informed Coupled Thresholding. To our knowledge, VERDICT is the first training-free verifier that makes the structure of cross-modal disagreement explicit and actionable. It computes consensus scores through a closed-form solution, enabling both disagreement-aware filtering and stability-conscious ranking of reasoning steps. Evaluated across six benchmarks, \method consistently improves over the base model by up to +5.95%, and performs competitively with domain-specific critics that demand extensive supervision, demonstrating that cross-modal agreement provides robust verification signals without task-specific adaptation and Training-Free Verification

Rohit Sinha, Kunal Tilaganji, Tanuja Ganu +3

Indian Institute of Technology, Hyderabad · Microsoft Research

Multimodal Large Language Models Multimodal CoT Reasoning MLLM Reasoning MLLM Evaluation

Ask PDF

Abstract

Multimodal Model Evaluation

Explore similar work

OmniVerifier-M1: Multimodal Meta-Verifier with Explicit Structured Recalibration

DeltaRubric: Generative Multimodal Reward Modeling via Joint Planning and Verification

VERDICT: Training-Free Step-Wise Verification of Multimodal Reasoning via Disagreement-Aware Consensus